OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning

# OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning
Source: [https://arxiv.org/html/2609.16057](https://arxiv.org/html/2609.16057)
1\]Beihang University 2\]The Chinese University of Hong Kong 3\]National University of Singapore\\contribution\[\*\]Corresponding authors: Xu Xu \(\), Jinxiu Liu \(\)\\checkdata\[Resources\][Project Page](https://omniharness.github.io/)[Code](https://github.com/OmniHarness/OmniHarness)

###### Abstract

Unified multimodal large language models \(MLLMs\) and multi\-agent systems have advanced visual generation\. However, three limitations remain\. \(1\) Existing methods often distill task\-specific experience with limited generalizability\. \(2\) Reflection is often deferred until task completion\. \(3\) Knowledge is often acquired only in response to downstream task demands\. To address these limitations, we introduceOmniHarness, a framework for generalizable visual generation viasymbolic policy learning\. OmniHarness abstracts verified executions into symbolic policies for visual generation task families, capturing shared procedures and applicability conditions while removing instance\-specific inputs\. The harness instantiates, adapts, and composes these policies for new tasks\. Intermediate verification guides refinement and failure recovery during execution\. Through self\-directed inquiry, OmniHarness autonomously generates and executes practice tasks near its capability limits before downstream objectives are specified\. Execution feedback continually refines the policies while model parameters remain fixed\. Experiments across six benchmarks, three MLLM backbones, and three visual agent frameworks demonstrate strong performance and continual capability expansion\. On ComfyBench’s Creative tasks, OmniHarness achieves a 95\.0% resolve rate, exceeding the strongest baseline by 27\.5 percentage points\. Frozen policy snapshots improve existing visual agent systems through plug\-and\-play reuse\.

## 1Introduction

Visual intelligence is advancing along two complementary paths for effective scaling\. The first develops end\-to\-end unified multimodal large language models \(MLLMs\) for visual perception\[[47](https://arxiv.org/html/2609.16057#bib.bib7),[10](https://arxiv.org/html/2609.16057#bib.bib8),[13](https://arxiv.org/html/2609.16057#bib.bib9),[36](https://arxiv.org/html/2609.16057#bib.bib19)\], multimodal reasoning\[[54](https://arxiv.org/html/2609.16057#bib.bib10),[74](https://arxiv.org/html/2609.16057#bib.bib11),[20](https://arxiv.org/html/2609.16057#bib.bib12)\], and image generation\[[19](https://arxiv.org/html/2609.16057#bib.bib15),[30](https://arxiv.org/html/2609.16057#bib.bib16),[6](https://arxiv.org/html/2609.16057#bib.bib17),[71](https://arxiv.org/html/2609.16057#bib.bib18)\]\. However, as shown in Figure[1](https://arxiv.org/html/2609.16057#S1.F1)\(a\), this approach relies on large\-scale training data\. Standalone models offer limited support for explicit verification and self\-correction and can struggle with complex reasoning tasks\. These limitations motivate the second path, which explores MLLM\-based multi\-agent systems \(MAS\)\[[49](https://arxiv.org/html/2609.16057#bib.bib20),[69](https://arxiv.org/html/2609.16057#bib.bib21),[76](https://arxiv.org/html/2609.16057#bib.bib22),[73](https://arxiv.org/html/2609.16057#bib.bib26)\]\.

Self\-Evolving Visual MAS\.Kahneman’s dual\-process theory distinguishes fast intuition from deliberate reasoning\[[24](https://arxiv.org/html/2609.16057#bib.bib13)\], while neuroscientific evidence suggests that language may express rather than underlie reasoning\[[14](https://arxiv.org/html/2609.16057#bib.bib14)\]\. This perspective motivates visual MAS that combine direct generation, collaborative reasoning, and memory to learn from experience\[[9](https://arxiv.org/html/2609.16057#bib.bib68),[21](https://arxiv.org/html/2609.16057#bib.bib29),[67](https://arxiv.org/html/2609.16057#bib.bib30)\]\. However, as illustrated in Figure[1](https://arxiv.org/html/2609.16057#S1.F1)\(b\), many systems still rely on manually defined ComfyUI workflows\[[32](https://arxiv.org/html/2609.16057#bib.bib28),[65](https://arxiv.org/html/2609.16057#bib.bib3),[22](https://arxiv.org/html/2609.16057#bib.bib4)\]or fixed communication topologies\[[35](https://arxiv.org/html/2609.16057#bib.bib23),[34](https://arxiv.org/html/2609.16057#bib.bib24),[63](https://arxiv.org/html/2609.16057#bib.bib25)\]\. Recent methods automate prompt or topology optimization, yet adapting coordination based on collaboration experience remains difficult\[[62](https://arxiv.org/html/2609.16057#bib.bib31),[38](https://arxiv.org/html/2609.16057#bib.bib32),[75](https://arxiv.org/html/2609.16057#bib.bib33)\]\. Retaining successful executions does not necessarily yield reusable skills or reliable cross\-task transfer\[[18](https://arxiv.org/html/2609.16057#bib.bib5),[37](https://arxiv.org/html/2609.16057#bib.bib6)\]\. When confined to individual cases, such experience preserves specific solutions without revealing the principles shared across a task family, much likegiving a fish without teaching how to fish\.

Harness Design for Self\-Evolving Agents\.A harness coordinates tools, workflows, and memory, shaping agent behavior alongside data and models\[[11](https://arxiv.org/html/2609.16057#bib.bib75),[68](https://arxiv.org/html/2609.16057#bib.bib74)\]\. Enabling self\-evolution through harness design raises three questions: ❶Can task\-specific experience reveal generalizable patterns?Miss the forest for the trees\.Existing methods distill execution experience but often remain focused on individual solutions, overlooking patterns shared across a task family and limiting transfer to new tasks\. ❷Can post\-task reflection alone ensure reliable execution?Hindsight offers lessons, but errors do not wait\.Many existing methods reflect only after task completion, allowing intermediate errors to propagate without timely verification or recovery\. ❸Can reactive learning prepare agents for future tasks?Necessity is a late teacher\.Existing methods often acquire knowledge only in response to downstream task demands, leaving capability gaps unaddressed until they hinder execution\. These challenges motivate a central question:How can we build a visual generation system that generalizes beyond individual cases, reflects as it acts, and learns through self\-directed exploration?

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/2.jpg)Figure 1:Comparison of visual generation paradigms\. \(a\) End\-to\-end unified MLLMs enable fast multimodal generation but lack deliberate reasoning and self\-correction\. \(b\) Existing multi\-agent systems improve collaborative reasoning but lack unified coordination and persistent knowledge accumulation\. \(c\) OmniHarness integrates self\-directed inquiry and feedback\-guided execution to learn reusable symbolic policies for generalizable visual generation\.To address this central question, we introduceOmniHarness, a framework for generalizable visual generation viasymbolic policy learning\. As shown in Figure[1](https://arxiv.org/html/2609.16057#S1.F1)\(c\), OmniHarness distills verified executions into symbolic policies for families of visual generation tasks, capturing shared procedures and applicability conditions while removing instance\-specific inputs\. The harness instantiates, adapts, and composes these policies for new tasks\. During execution, it verifies intermediate outputs and repairs failed steps\. Motivated by Chinese philosopher Wang Yangming’s interpretation ofthe investigation of things and the extension of knowledge\[[52](https://arxiv.org/html/2609.16057#bib.bib76)\], we incorporate self\-directed inquiry\. Before downstream objectives are specified, OmniHarness autonomously generates and executes practice tasks to probe its capability limits and gather experience for policy learning\. Execution feedback continually refines these policies while model parameters remain fixed\. Frozen policy snapshots support plug\-and\-play reuse across visual agent frameworks\.

Our contributions are summarized as follows:

- •Symbolic Policy Learning\.We distill verified executions into symbolic policies for visual generation task families, capturing shared procedures and applicability conditions while removing instance\-specific inputs\.
- •Feedback\-Guided Execution\.We design a harness that instantiates, adapts, and composes symbolic policies for new tasks\. Verification of intermediate outputs guides workflow refinement and failure recovery during execution, limiting error propagation\.
- •Self\-Directed Inquiry\.We introduce self\-directed inquiry to learn reusable symbolic policies before downstream objectives are specified\. OmniHarness autonomously generates and executes practice tasks, probing its capability limits and using execution feedback to guide policy learning\.
- •Experimental Evaluation\.Experiments across six benchmarks demonstrate strong performance, continual capability expansion, and transfer across visual agent frameworks\. On ComfyBench’s Creative tasks, OmniHarness achieves a 95\.0% resolve rate, exceeding the state\-of\-the\-art baseline by 27\.5 percentage points\.

## 2Related Work

End\-to\-End Unified MLLMs\.Unified MLLMs integrate visual understanding and generation within a single architecture\[[50](https://arxiv.org/html/2609.16057#bib.bib34),[15](https://arxiv.org/html/2609.16057#bib.bib35),[61](https://arxiv.org/html/2609.16057#bib.bib36)\]\. Recent work advances reasoning through cross\-modal chain\-of\-thought\[[33](https://arxiv.org/html/2609.16057#bib.bib37),[5](https://arxiv.org/html/2609.16057#bib.bib38)\], multi\-representation mutual reinforcement\[[46](https://arxiv.org/html/2609.16057#bib.bib39)\], and shared\-context visual tokenization\[[42](https://arxiv.org/html/2609.16057#bib.bib40)\]\. However, standard inference offers limited support for explicit verification, failure recovery, and persistent workflow reuse\.

Agentic Systems\.ComfyBench evaluates autonomous workflow construction in ComfyUI\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\], while related systems combine planning and feedback to construct workflows for assigned tasks\[[22](https://arxiv.org/html/2609.16057#bib.bib4),[18](https://arxiv.org/html/2609.16057#bib.bib5)\]\. Recent methods evolve execution checks and recovery for embodied agents\[[11](https://arxiv.org/html/2609.16057#bib.bib75)\], synthesize task\-specific harnesses\[[68](https://arxiv.org/html/2609.16057#bib.bib74)\], or learn reusable symbolic concepts from incoming tasks\[[37](https://arxiv.org/html/2609.16057#bib.bib6)\]\. OmniHarness learns symbolic policies from verified executions, capturing principles shared across visual generation task families\. The harness adapts and composes these policies for new tasks, using intermediate verification to guide refinement and recovery during execution\. Self\-directed inquiry autonomously generates and executes practice tasks to probe capability limits before downstream objectives are specified\. Execution feedback continually refines the policies while model parameters remain fixed\. Frozen policy snapshots support plug\-and\-play reuse across visual agent frameworks\.

## 3OmniHarness: Symbolic Policy Learning

### 3\.1Self\-Directed Inquiry

Motivated by Wang Yangming’s interpretation ofthe investigation of things and the extension of knowledge\[[52](https://arxiv.org/html/2609.16057#bib.bib76)\], OmniHarness uses self\-directed inquiry to learn symbolic policies before downstream objectives are specified\. Figure[2](https://arxiv.org/html/2609.16057#S3.F2)shows the architecture\. Appendix Sections[A\.1](https://arxiv.org/html/2609.16057#A1.SS1)and[A\.2](https://arxiv.org/html/2609.16057#A1.SS2)detail the formulation, learning procedure, and inquiry configuration\.

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/3.jpg)Figure 2:Architecture of OmniHarness\. Self\-directed inquiry and feedback\-guided execution drive symbolic policy learning for generalizable visual generation\. The policy library evolves during OmniHarness execution, while frozen snapshots support plug\-and\-play reuse across visual agents\.Candidate Task Generation\.At iterationtt, the proposer generates𝒯t=Propose⁡\(𝒢,ct,𝒮t\)\\mathcal\{T\}\_\{t\}=\\operatorname\{Propose\}\(\\mathcal\{G\},c\_\{t\},\\mathcal\{S\}\_\{t\}\)using capability space𝒢\\mathcal\{G\}, scene contextctc\_\{t\}, and policy library state𝒮t=\(ℒt,ℱt\)\\mathcal\{S\}\_\{t\}=\(\\mathcal\{L\}\_\{t\},\\mathcal\{F\}\_\{t\}\)\. The contextctc\_\{t\}summarizes capability coverage, workflow reliability, and available source images\. The workflow libraryℒt\\mathcal\{L\}\_\{t\}stores symbolic policies as reusable workflow templates for visual generation task families, while the failure libraryℱt\\mathcal\{F\}\_\{t\}stores failure evidence and corrective strategies\. Each candidateτ=\(qτ,mτ,𝒢τ,xτ\)\\tau=\(q\_\{\\tau\},m\_\{\\tau\},\\mathcal\{G\}\_\{\\tau\},x\_\{\\tau\}\)specifies a descriptionqτq\_\{\\tau\}, modalitymτ∈\{T2I,I2I\}m\_\{\\tau\}\\in\\\{\\mathrm\{T2I\},\\mathrm\{I2I\}\\\}, required capabilities𝒢τ⊆𝒢\\mathcal\{G\}\_\{\\tau\}\\subseteq\\mathcal\{G\}, and source imagexτ∈𝒳∪\{⊥\}x\_\{\\tau\}\\in\\mathcal\{X\}\\cup\\\{\\bot\\\}, wherexτ=⊥x\_\{\\tau\}=\\botfor T2I\.

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/comfybench.jpg)Figure 3:Qualitative comparison on challenging tasks from the Complex and Creative subsets\.Generation encourages diverse capability combinations, filters near\-duplicates, and avoids known failure patterns\.

Exploration within Reach\.Candidates are scored by capability novelty𝒩⁡\(τ\)\\mathcal\{N\}\(\\tau\)and competence frontier score𝒞⁡\(τ\)\\mathcal\{C\}\(\\tau\), with time indices omitted\. Setzτ=zT2Iz\_\{\\tau\}=z\_\{\\mathrm\{T2I\}\}for T2I andzτ=xτz\_\{\\tau\}=x\_\{\\tau\}for I2I\. Given attempt countsnt​\(z,g\)n\_\{t\}\(z,g\)for contextzzand capabilitygg,𝒩⁡\(τ\)=1\|𝒢τ\|​∑g∈𝒢τ1nt​\(zτ,g\)\+1\\mathcal\{N\}\(\\tau\)=\\frac\{1\}\{\|\\mathcal\{G\}\_\{\\tau\}\|\}\\sum\_\{g\\in\\mathcal\{G\}\_\{\\tau\}\}\\frac\{1\}\{\\sqrt\{n\_\{t\}\(z\_\{\\tau\},g\)\+1\}\}favors underexplored context–capability pairs\. Capability support uses applicable, non\-suspended workflows𝒲t​\(τ,g\)⊆ℒt\\mathcal\{W\}\_\{t\}\(\\tau,g\)\\subseteq\\mathcal\{L\}\_\{t\}\. Workflow reliability is the lower endpoint of a 95% Wilson confidence interval based on usage and success counts\. The highest reliability definesrt​\(τ,g\)r\_\{t\}\(\\tau,g\), with a small priorε\\varepsilonwhen𝒲t​\(τ,g\)\\mathcal\{W\}\_\{t\}\(\\tau,g\)is empty\. Estimated task competence follows the weakest required capability,r¯t​\(τ\)=ming∈𝒢τ⁡rt​\(τ,g\)\\bar\{r\}\_\{t\}\(\\tau\)=\\min\_\{g\\in\\mathcal\{G\}\_\{\\tau\}\}r\_\{t\}\(\\tau,g\)\. See Appendix Section[A\.1\.2](https://arxiv.org/html/2609.16057#A1.SS1.SSS2)for reliability details\. Following the Goldilocks principle\[[1](https://arxiv.org/html/2609.16057#bib.bib1),[26](https://arxiv.org/html/2609.16057#bib.bib2)\], the learnability heuristic𝒞⁡\(τ\)=4​r¯t​\(τ\)​\(1−r¯t​\(τ\)\)\\mathcal\{C\}\(\\tau\)=4\\bar\{r\}\_\{t\}\(\\tau\)\\left\(1\-\\bar\{r\}\_\{t\}\(\\tau\)\\right\)peaks atr¯t​\(τ\)=0\.5\\bar\{r\}\_\{t\}\(\\tau\)=0\.5and downweights tasks with very low or high estimated competence\. OmniHarness selectsτt=arg​maxτ∈𝒯t⁡𝒩​\(τ\)​𝒞​\(τ\)\\tau\_\{t\}=\\operatorname\*\{arg\\,max\}\_\{\\tau\\in\\mathcal\{T\}\_\{t\}\}\\mathcal\{N\}\(\\tau\)\\mathcal\{C\}\(\\tau\), favoring novel tasks near the competence frontier as𝒮t\\mathcal\{S\}\_\{t\}evolves\.

### 3\.2Feedback\-Guided Execution

For a practice or downstream taskτt\\tau\_\{t\}, OmniHarness executes symbolic policies through\(πt,ot\)=Run⁡\(τt,ct,𝒦,𝒮t\)\(\\pi\_\{t\},o\_\{t\}\)=\\allowbreak\\operatorname\{Run\}\(\\tau\_\{t\},\\allowbreak c\_\{t\},\\allowbreak\\mathcal\{K\},\\allowbreak\\mathcal\{S\}\_\{t\}\), whereπt\\pi\_\{t\}is the task’s executable workflow andoto\_\{t\}records status, verifier feedback, and evidence\. Agents share plans, programs, verification results, and corrections\. The planner constructs the ordered planρt=\(at,1,…,at,Jt\)\\rho\_\{t\}=\(a\_\{t,1\},\\ldots,a\_\{t,J\_\{t\}\}\)viaρt=Plan⁡\(τt,ct,𝒦,𝒮t\)\\rho\_\{t\}=\\operatorname\{Plan\}\(\\tau\_\{t\},c\_\{t\},\\mathcal\{K\},\\mathcal\{S\}\_\{t\}\)\. It instantiates, adapts, and composes policies from the workflow libraryℒt\\mathcal\{L\}\_\{t\}, guided by failure patterns and remedies inℱt\\mathcal\{F\}\_\{t\}\. A plan verifier checks step order, dependencies, and task alignment\. The writer generates a Python\-like Code\-as\-Policy programπ~t=Write⁡\(ρt,𝒦\)\\widetilde\{\\pi\}\_\{t\}=\\operatorname\{Write\}\(\\rho\_\{t\},\\mathcal\{K\}\), which a reversible interpreter compiles intoπt=Compile⁡\(π~t\)\\pi\_\{t\}=\\operatorname\{Compile\}\(\\widetilde\{\\pi\}\_\{t\}\)\. Function calls represent ComfyUI nodes, and data flow defines their connections\. Workflows and components are reused when their preconditions hold\. Verification checks executability, the intended effect of eachat,ja\_\{t,j\}, and whether the output satisfiesqτtq\_\{\\tau\_\{t\}\}and the constraints inτt\\tau\_\{t\}\. On failure, the diagnoser identifies the affected step and retrieves a correction fromℱt\\mathcal\{F\}\_\{t\}\. The harness repairs that component while preserving verified steps\. A subagent supplies a reusable subworkflow when needed\. Verification repeats until success or the retry budget is exhausted\.

Table 1:Quantitative comparison on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\.
### 3\.3Symbolic Policy Learning

After each task, OmniHarness updates its context and policy library through\(ct\+1,𝒮t\+1\)=Update⁡\(ct,𝒮t,τt,πt,ot\)\(c\_\{t\+1\},\\mathcal\{S\}\_\{t\+1\}\)=\\allowbreak\\operatorname\{Update\}\(c\_\{t\},\\allowbreak\\mathcal\{S\}\_\{t\},\\allowbreak\\tau\_\{t\},\\allowbreak\\pi\_\{t\},\\allowbreak o\_\{t\}\), where𝒮t=\(ℒt,ℱt\)\\mathcal\{S\}\_\{t\}=\(\\mathcal\{L\}\_\{t\},\\mathcal\{F\}\_\{t\}\)\. On verified success,πt\\pi\_\{t\}is distilled into a symbolic policy inℒt\+1\\mathcal\{L\}\_\{t\+1\}for its visual generation task family\. This abstraction captures shared procedures and applicability conditions while removing instance\-specific inputs\. Equivalent workflows are merged, and usage, success, and reliability statistics are updated to guide condition\-aware retrieval and composition\. Failures are recorded inℱt\+1\\mathcal\{F\}\_\{t\+1\}withτt\\tau\_\{t\},πt\\pi\_\{t\}, execution evidence, and verifier feedback\. Their analysis identifies root causes, workflow antipatterns, remedies, and applicable scope\. A curator periodically merges redundant workflows, consolidates corrective strategies, updates reliability tiers, and may construct missing workflows\. It refreshesct\+1c\_\{t\+1\}from the libraries and source image pool𝒳\\mathcal\{X\}to guide future task proposals\. Updates apply to both practice and downstream tasks, continually refining the policy library\. The policy library learned through self\-directed inquiry is exported as a frozen snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}for plug\-and\-play reuse by external visual agents\.

## 4Experiments

### 4\.1Autonomous Workflow Construction

We evaluate autonomous workflow construction on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\], where each agent must construct an executable ComfyUI workflow that satisfies the task requirements\. Table[1](https://arxiv.org/html/2609.16057#S3.T1)shows that both OmniHarness variants achieve a 100\.0% Pass rate across all subsets\. GPT\-4o \+ OmniHarness and Codex GPT\-4o \+ OmniHarness achieve Total Resolve rates of 89\.5% and 92\.5%, respectively\. The latter exceeds SymbOmni by 6\.5 percentage points overall, with the largest gain on Creative tasks, where it achieves 95\.0% Resolve compared with SymbOmni’s 67\.5%\. On Complex tasks, OmniHarness matches SymbOmni at 83\.3% Resolve, while ComfyMind achieves 85\.0%\. Figure[3](https://arxiv.org/html/2609.16057#S3.F3)provides qualitative examples of multi\-step editing, reference\-style transfer, restoration, and content preservation\.

### 4\.2Text\-to\-Image Generation

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/Text-to-Image_Generation.jpg)Figure 4:Qualitative comparison on representative text\-to\-image tasks\.Table 2:Quantitative comparison on GenEval\[[17](https://arxiv.org/html/2609.16057#bib.bib45)\]\.Table 3:Quantitative comparison on GenEval2\[[25](https://arxiv.org/html/2609.16057#bib.bib53)\]\.Table 4:Quantitative comparison on WISE\[[41](https://arxiv.org/html/2609.16057#bib.bib57)\]\.We evaluate text\-to\-image generation on GenEval\[[17](https://arxiv.org/html/2609.16057#bib.bib45)\], GenEval2\[[25](https://arxiv.org/html/2609.16057#bib.bib53)\], and WISE\[[41](https://arxiv.org/html/2609.16057#bib.bib57)\]\. GenEval measures six object\-centric compositional skills, GenEval2 tests fine\-grained attributes, counting, and spatial and transitive verb relations, while WISE assesses knowledge\-informed synthesis across cultural, spatiotemporal, and scientific domains\. For these evaluations, self\-directed inquiry uses a general T2I generative capability space without access to evaluation tasks from these benchmarks\. As shown in Tables[2](https://arxiv.org/html/2609.16057#S4.T2)–[4](https://arxiv.org/html/2609.16057#S4.T4), OmniHarness achieves the highest GenEval overall score of 0\.997, reaching 1\.00 in five categories and 0\.98 in attribute binding\. On GenEval2, it leads in Attribute, Count, Position, and Verb with scores of 94\.0, 94\.0, 76\.9, and 89\.0, exceeding the best competing scores by 2\.6, 19\.2, 6\.7, and 2\.3 points, respectively\. Its Object score of 95\.0 matches SymbOmni but remains below Qwen\-Image and Gemini 2\.5 Flash Image\. On WISE, it achieves the highest overall WiScore of 0\.86, exceeding both SymbOmni and GPT\-Image\-1 by 0\.06\. It leads in Time, Biology, Physics, and Chemistry and remains within 0\.03 of the best Cultural and Space scores\. Figure[4](https://arxiv.org/html/2609.16057#S4.F4)further illustrates adherence to object counts, attribute combinations, spatial and action relations, and world\-knowledge constraints\.

### 4\.3Image Editing

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/reason-edit.jpg)Figure 5:Qualitative results on \(a\) Understanding Scenarios, \(b\) Reasoning Scenarios\.Table 5:Quantitative comparison on Reason\-Edit\[[23](https://arxiv.org/html/2609.16057#bib.bib63)\]\.We evaluate instruction\-based image editing on Reason\-Edit\[[23](https://arxiv.org/html/2609.16057#bib.bib63)\], which includes explicit target cues in Understanding Scenarios and indirect target descriptions in Reasoning Scenarios\. During self\-directed inquiry, OmniHarness may access raw source images used by ComfyBench, but downstream task instructions, target outputs, reference workflows, benchmark annotations, and evaluation labels are withheld to prevent task\-level leakage\. As shown in Table[5](https://arxiv.org/html/2609.16057#S4.T5), OmniHarness leads on all four metrics in Understanding Scenarios, with 23\.89 dB PSNR, 0\.86 SSIM, 0\.05 LPIPS, and 24\.55 CLIP Score\. In Reasoning Scenarios, it achieves the highest SSIM of 0\.80 and CLIP Score of 21\.32, while its LPIPS of 0\.05 matches the best baselines at two\-decimal precision\. Figure[5](https://arxiv.org/html/2609.16057#S4.F5)provides complementary qualitative evidence\.

### 4\.4Ablation Study

Table 6:Quantitative ablation results on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\.Table 7:Ablation of task selection in self\-directed inquiry on the Creative subset\.Table 8:Ablation of feedback\-guided execution on the Complex subset\.Table[6](https://arxiv.org/html/2609.16057#S4.T6)shows that removing self\-directed inquiry lowers Total Resolve from 92\.5% to 87\.0% and Creative Resolve from 95\.0% to 72\.5%\. Creative tasks test skill application beyond curriculum examples\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\], and the larger decline supports prior policy acquisition for new generation requirements\. Disabling online policy updates lowers Total Resolve to 88\.5%, supporting continual refinement through execution feedback\. Table[7](https://arxiv.org/html/2609.16057#S4.T7)shows that removing capability novelty or the competence frontier score lowers Creative Resolve to 87\.5% and 85\.0%, respectively\. Their combination outperforms either alone, supporting the complementary roles of exploration and estimated learnability\. Complex tasks require combining multiple workflows\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\], testing composition and coordination across dependent steps\. In Table[8](https://arxiv.org/html/2609.16057#S4.T8), removing planning, intermediate verification, or localized recovery lowers Complex Resolve from 83\.3% to 55\.0%, 68\.3%, and 76\.7%, respectively\. Removing intermediate verification also reduces Complex Pass from 100\.0% to 75\.0%\. These declines support planning for dependency coordination and intermediate feedback for workflow refinement and recovery, consistent with limiting error propagation\. Additional ablation results appear in Appendix Section[A\.4](https://arxiv.org/html/2609.16057#A1.SS4)\.

## 5Conclusion

Visual agents need to generalize across tasks, correct errors during execution, and learn before new task demands arise\. We introduce OmniHarness, a framework for generalizable visual generation via symbolic policy learning\. It abstracts verified executions into reusable symbolic policies for visual generation task families\. The harness adapts and composes these policies for new tasks, with intermediate verification guiding refinement and localized recovery\. Self\-directed inquiry acquires policies before downstream objectives are specified, while execution feedback continually refines them without model fine\-tuning\. Experiments across six benchmarks demonstrate effectiveness\. On ComfyBench’s Creative tasks, OmniHarness achieves a 95\.0% Resolve rate, exceeding the strongest baseline by 27\.5 percentage points\. Frozen policy snapshots improve external agents through plug\-and\-play reuse\. Policies learned through image\-only inquiry also transfer to unseen video generation tasks, supporting reuse across tasks, frameworks, and modalities\.

## References

- \[1\]\(2013\)Active learning of inverse models with intrinsically motivated goal exploration in robots\.Robotics and Autonomous Systems61\(1\),pp\. 49–73\.External Links:ISSN 0921\-8890,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.robot.2012.05.008),[Link](https://www.sciencedirect.com/science/article/pii/S0921889012000644)Cited by:[§A\.1\.2](https://arxiv.org/html/2609.16057#A1.SS1.SSS2.p7.1),[§3\.1](https://arxiv.org/html/2609.16057#S3.SS1.p4.1)\.
- \[2\]T\. Brooks, A\. Holynski, and A\. A\. Efros\(2023\)InstructPix2Pix: learning to follow image editing instructions\.In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 18392–18402\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52729.2023.01764)Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.14.1),[Table 5](https://arxiv.org/html/2609.16057#S4.T5.5.3.1)\.
- \[3\]T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell, S\. Agarwal, A\. Herbert\-Voss, G\. Krueger, T\. Henighan, R\. Child, A\. Ramesh, D\. Ziegler, J\. Wu, C\. Winter, C\. Hesse, M\. Chen, E\. Sigler, M\. Litwin, S\. Gray, B\. Chess, J\. Clark, C\. Berner, S\. McCandlish, A\. Radford, I\. Sutskever, and D\. Amodei\(2020\)Language models are few\-shot learners\.InAdvances in Neural Information Processing Systems,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\.F\. Balcan, and H\. Lin \(Eds\.\),Vol\.33,pp\. 1877–1901\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf)Cited by:[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.4.1)\.
- \[4\]J\. Chen, J\. YU, C\. GE, L\. Yao, E\. Xie, Z\. Wang, J\. Kwok, P\. Luo, H\. Lu, and Z\. Li\(2024\)PixArt\-α\\alpha: fast training of diffusion transformer for photorealistic text\-to\-image synthesis\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 57611–57640\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/fe989bb038b5dcc44181255dd6913e43-Paper-Conference.pdf)Cited by:[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.9.1)\.
- \[5\]L\. L\. Chen, H\. Ma, Z\. Fan, Z\. Huang, A\. Sinha, X\. Dai, J\. Wang, Z\. He, J\. Yang, C\. Li, J\. Sun, C\. Wang, S\. Yeung\-Levy, and F\. Juefei\-Xu\(2026\)UniT: unified multimodal chain\-of\-thought test\-time scaling\.External Links:2602\.12279,[Link](https://arxiv.org/abs/2602.12279)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1)\.
- \[6\]S\. Chen, Z\. Xing, T\. Ye, X\. Geng, Y\. Lin, J\. Lai, X\. He, F\. Zhai, J\. Gao, and L\. Zhu\(2026\)GenEvolve: self\-evolving image generation agents via tool\-orchestrated visual experience distillation\.External Links:2605\.21605,[Link](https://arxiv.org/abs/2605.21605)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[7\]X\. Chen, Z\. Wu, X\. Liu, Z\. Pan, W\. Liu, Z\. Xie, X\. Yu, and C\. Ruan\(2025\)Janus\-pro: unified multimodal understanding and generation with data and model scaling\.External Links:2501\.17811,[Link](https://arxiv.org/abs/2501.17811)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.15.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.16.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.17.1)\.
- \[8\]G\. Comanici, E\. Bieber, M\. Schaekermann, I\. Pasupat, N\. Sachdeva, I\. Dhillon, M\. Blistein, O\. Ram, D\. Zhang, E\. Rosen,et al\.\(2025\)Gemini 2\.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities\.arXiv preprint arXiv:2507\.06261\.Cited by:[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.11.1)\.
- \[9\]Y\. Dang, C\. Qian, X\. Luo, J\. Fan, Z\. Xie, R\. Shi, W\. Chen, C\. Yang, X\. Che, Y\. Tian, X\. Xiong, L\. Han, Z\. Liu, and M\. Sun\(2025\)Multi\-agent collaboration via evolving orchestration\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38, Main Conference,pp\. 165025–165059\.External Links:[Document](https://dx.doi.org/10.52202/085713-5502),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/f1320d2e2842169c6fc89dcbd80e94d0-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[10\]C\. Deng, D\. Zhu, K\. Li, C\. Gou, F\. Li, Z\. Wang, S\. Zhong, W\. Yu, X\. Nie, Z\. Song,et al\.\(2025\)Emerging properties in unified multimodal pretraining\.arXiv preprint arXiv:2505\.14683\.Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.7.1),[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.8.1),[§1](https://arxiv.org/html/2609.16057#S1.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.17.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.9.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.24.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.25.1)\.
- \[11\]X\. Ding, L\. Mi, M\. Huang, Z\. Wang, C\. Zhang, Z\. Hao, F\. Chen, X\. Li, Y\. Zheng, Y\. Guo, W\. Wang, K\. Li, H\. Wu, Y\. Liu, and T\. Cao\(2026\)Zettaζ\\zeta: an efficient closed\-loop embodied harness for self\-evolving physical intelligence\.External Links:2608\.16590,[Link](https://arxiv.org/abs/2608.16590)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p3.1),[§2](https://arxiv.org/html/2609.16057#S2.p2.1)\.
- \[12\]P\. Esser, S\. Kulal, A\. Blattmann, R\. Entezari, J\. Müller, H\. Saini, Y\. Levi, D\. Lorenz, A\. Sauer, F\. Boesel, D\. Podell, T\. Dockhorn, Z\. English, and R\. Rombach\(2024\)Scaling rectified flow transformers for high\-resolution image synthesis\.InForty\-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21\-27, 2024,R\. Salakhutdinov, Z\. Kolter, K\. A\. Heller, A\. Weller, N\. Oliver, J\. Scarlett, and F\. Berkenkamp \(Eds\.\),Proceedings of Machine Learning Research, Vol\.235,pp\. 12606–12633\.External Links:[Link](https://proceedings.mlr.press/v235/esser24a.html)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.7.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.5.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.6.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.6.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.7.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.8.1)\.
- \[13\]R\. Fang, C\. Duan, K\. Wang, L\. Huang, H\. Li, H\. Tian, S\. Yan, W\. Yu, X\. Zeng, J\. Dai, X\. Liu, and H\. Li\(2025\)GoT: unleashing reasoning capability of mllm for visual generation and editing\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38, Main Conference,pp\. 67680–67708\.External Links:[Document](https://dx.doi.org/10.52202/085713-2270),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/61960fdfda4d4e95fa1c1f6e64bfe8bc-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.16.1)\.
- \[14\]E\. Fedorenko, S\. T\. Piantadosi, and E\. A\. F\. Gibson\(2024\)Language is primarily a tool for communication rather than thought\.Nature630\(8017\),pp\. 575–586\.External Links:[Document](https://dx.doi.org/10.1038/s41586-024-07522-w),[Link](https://doi.org/10.1038/s41586-024-07522-w),ISSN 1476\-4687Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[15\]Y\. Ge, S\. Zhao, J\. Zhu, Y\. Ge, K\. Yi, L\. Song, C\. Li, X\. Ding, and Y\. Shan\(2025\)SEED\-x: multimodal models with unified multi\-granularity comprehension and generation\.External Links:2404\.14396,[Link](https://arxiv.org/abs/2404.14396)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.11.1)\.
- \[16\]Z\. Geng, B\. Yang, T\. Hang, C\. Li, S\. Gu, T\. Zhang, J\. Bao, Z\. Zhang, H\. Li, H\. Hu, D\. Chen, and B\. Guo\(2024\)InstructDiffusion: a generalist modeling interface for vision tasks\.In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 12709–12720\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52733.2024.01208)Cited by:[Table 5](https://arxiv.org/html/2609.16057#S4.T5.5.5.1)\.
- \[17\]D\. Ghosh, H\. Hajishirzi, and L\. Schmidt\(2023\)GenEval: an object\-focused framework for evaluating text\-to\-image alignment\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 52132–52152\.External Links:[Document](https://dx.doi.org/10.52202/075280-2270),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/a3bf71c7c63f0c3bcb7ff67c67b1e7b1-Paper-Datasets_and_Benchmarks.pdf)Cited by:[Figure 13](https://arxiv.org/html/2609.16057#A1.F13),[Figure 13](https://arxiv.org/html/2609.16057#A1.F13.4),[§A\.5](https://arxiv.org/html/2609.16057#A1.SS5.p1.1),[§4\.2](https://arxiv.org/html/2609.16057#S4.SS2.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.4)\.
- \[18\]L\. Guo, X\. Xu, L\. Wang, J\. Lin, J\. Zhou, Z\. Zhang, B\. Su, and Y\. Chen\(2025\)ComfyMind: toward general\-purpose generation via tree\-based planning and reactive feedback\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38, Main Conference,pp\. 45128–45164\.External Links:[Document](https://dx.doi.org/10.52202/085713-1503),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/40168e00bf87869c5d153e934d8a3602-Paper-Conference.pdf)Cited by:[§A\.3](https://arxiv.org/html/2609.16057#A1.SS3.p1.1),[§A\.6](https://arxiv.org/html/2609.16057#A1.SS6.p1.1),[Table 13](https://arxiv.org/html/2609.16057#A1.T13.5.4.1),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.5.4.1.1),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.5.7.1.1),[Table 17](https://arxiv.org/html/2609.16057#A1.T17.5.4.1.1),[§1](https://arxiv.org/html/2609.16057#S1.p2.1),[§2](https://arxiv.org/html/2609.16057#S2.p2.1),[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.14.1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.21.1.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.29.1.1)\.
- \[19\]S\. Han, H\. Ji, S\. Xin, J\. Shi, S\. Qiu, X\. Ye, P\. Xia, J\. Liu, Z\. Chen, Y\. Zhou, L\. Li, L\. Wang, and H\. Yao\(2026\)Paper2Figure: a multi\-agent collaborative system for figure generation towards academic research paper\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 29157–29166\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[20\]Z\. Hao, H\. Wang, J\. Luo, J\. Zhang, Y\. Zhou, Q\. Lin, C\. Wang, H\. Dong, and J\. Chen\(2026\)ReCreate: reasoning and creating domain agents driven by experience\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 31018–31046\.External Links:[Link](https://aclanthology.org/2026.acl-long.1432/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1432),ISBN 979\-8\-89176\-390\-6Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[21\]J\. Huang, L\. Chen, Z\. Xu, Y\. Li, X\. Xu, S\. Chen, J\. Liu, M\. Kong, and Q\. Zhu\(2026\)META: meta evolution of tool trajectory adaptation for long\-video understanding\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 9837–9846\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[22\]O\. Huang, Y\. Ma, Z\. Zhao, M\. Wu, J\. Ji, R\. Zhang, Z\. Hu, X\. Sun, and R\. Ji\(2025\)ComfyGPT: a self\-optimizing multi\-agent system for comprehensive comfyui workflow generation\.External Links:2503\.17671,[Link](https://arxiv.org/abs/2503.17671)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1),[§2](https://arxiv.org/html/2609.16057#S2.p2.1)\.
- \[23\]Y\. Huang, L\. Xie, X\. Wang, Z\. Yuan, X\. Cun, Y\. Ge, J\. Zhou, C\. Dong, R\. Huang, R\. Zhang, and Y\. Shan\(2024\)SmartEdit: exploring complex instruction\-based image editing with multimodal large language models\.In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 8362–8371\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52733.2024.00799)Cited by:[Figure 15](https://arxiv.org/html/2609.16057#A1.F15),[Figure 15](https://arxiv.org/html/2609.16057#A1.F15.4),[§A\.5](https://arxiv.org/html/2609.16057#A1.SS5.p1.1),[§4\.3](https://arxiv.org/html/2609.16057#S4.SS3.p1.1),[Table 5](https://arxiv.org/html/2609.16057#S4.T5),[Table 5](https://arxiv.org/html/2609.16057#S4.T5.4),[Table 5](https://arxiv.org/html/2609.16057#S4.T5.5.6.1),[Table 5](https://arxiv.org/html/2609.16057#S4.T5.5.7.1)\.
- \[24\]D\. Kahneman\(2011\)Thinking, fast and slow\.Farrar, Straus and Giroux\.External Links:ISBN 9780374275631,[Link](https://us.macmillan.com/books/9780374275631/thinkingfastandslow/)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[25\]A\. Kamath, K\. Chang, R\. Krishna, L\. Zettlemoyer, Y\. Hu, and M\. Ghazvininejad\(2025\)GenEval 2: addressing benchmark drift in text\-to\-image evaluation\.External Links:2512\.16853,[Link](https://arxiv.org/abs/2512.16853)Cited by:[Figure 14](https://arxiv.org/html/2609.16057#A1.F14),[Figure 14](https://arxiv.org/html/2609.16057#A1.F14.4),[§A\.5](https://arxiv.org/html/2609.16057#A1.SS5.p1.1),[§4\.2](https://arxiv.org/html/2609.16057#S4.SS2.p1.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.4)\.
- \[26\]C\. Kidd, S\. T\. Piantadosi, and R\. N\. Aslin\(2012\)The goldilocks effect: human infants allocate attention to visual sequences that are neither too simple nor too complex\.PLOS ONE7\(5\),pp\. 1–8\.External Links:[Document](https://dx.doi.org/10.1371/journal.pone.0036399),[Link](https://doi.org/10.1371/journal.pone.0036399)Cited by:[§A\.1\.2](https://arxiv.org/html/2609.16057#A1.SS1.SSS2.p7.1),[§3\.1](https://arxiv.org/html/2609.16057#S3.SS1.p4.1)\.
- \[27\]S\. Kou, J\. Jin, Z\. Liu, C\. Liu, Y\. Ma, J\. Jia, Q\. Chen, P\. Jiang, and Z\. Deng\(2025\)Orthus: autoregressive interleaved image\-text generation with modality\-specific heads\.External Links:2412\.00127,[Link](https://arxiv.org/abs/2412.00127)Cited by:[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.21.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.22.1)\.
- \[28\]B\. F\. Labs, S\. Batifol, A\. Blattmann, F\. Boesel, S\. Consul, C\. Diagne, T\. Dockhorn, J\. English, Z\. English, P\. Esser, S\. Kulal, K\. Lacey, Y\. Levi, C\. Li, D\. Lorenz, J\. Müller, D\. Podell, R\. Rombach, H\. Saini, A\. Sauer, and L\. Smith\(2025\)FLUX\.1 kontext: flow matching for in\-context image generation and editing in latent space\.External Links:2506\.15742,[Link](https://arxiv.org/abs/2506.15742)Cited by:[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.8.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.11.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.12.1)\.
- \[29\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.InAdvances in Neural Information Processing Systems,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\.F\. Balcan, and H\. Lin \(Eds\.\),Vol\.33,pp\. 9459–9474\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205f780e1bc26945df7481e5-Paper.pdf)Cited by:[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.7.1)\.
- \[30\]C\. Li, Q\. Wu, J\. Pan, K\. Hui, J\. Hu, Y\. Jiang, B\. Sheng, X\. Liu, W\. Gong, and Z\. Liu\(2026\)CoDrawAgents: a multi\-agent dialogue framework for compositional image generation\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\) Findings,pp\. 9802–9812\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[31\]D\. Li, A\. Kamko, E\. Akhgari, A\. Sabet, L\. Xu, and S\. Doshi\(2024\)Playground v2\.5: three insights towards enhancing aesthetic quality in text\-to\-image generation\.External Links:2402\.17245,[Link](https://arxiv.org/abs/2402.17245)Cited by:[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.10.1)\.
- \[32\]Z\. Li, D\. Liu, F\. Liu, Y\. Zhou, X\. Wu, J\. Chen, J\. Xie, X\. Wu, and L\. Sun\(2026\)COMFYCLAW: self\-evolving skill harnesses for image generation workflows\.External Links:2607\.01709,[Link](https://arxiv.org/abs/2607.01709)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[33\]Q\. Liang, Y\. Wu, K\. Li, J\. Wei, S\. He, J\. Guo, and N\. Xie\(2025\)MM\-r1: unleashing the power of unified multimodal large language models for personalized image generation\.External Links:2508\.11433,[Link](https://arxiv.org/abs/2508.11433)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1)\.
- \[34\]J\. Lin, C\. Jiang, X\. Lin, R\. Zhang, X\. Zhu, J\. Liu, C\. Tang, Y\. Du, S\. Gao, J\. Ning,et al\.\(2026\)EvoGraph\-r1: self\-evolving multimodal knowledge hypergraphs for agentic retrieval\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 756–765\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[35\]Y\. Lin, L\. Wang, K\. Lin, Z\. Lin, K\. Gong, W\. Li, B\. Lin, Z\. Li, S\. Zhang, Y\. Peng, W\. Dai, X\. Ding, C\. Wang, and Q\. Lu\(2026\)JarvisEvo: towards a self\-evolving photo editing agent with synergistic editor\-evaluator optimization\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 27291–27302\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[36\]H\. Liu, W\. Yan, M\. Zaharia, and P\. Abbeel\(2025\)World model on million\-length video and language with blockwise ringattention\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 45953–45977\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/71859ac75d53879d9bbd2f4b77b59929-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.10.1)\.
- \[37\]J\. Liu, J\. Li, T\. Kuang, X\. Liu, K\. Mei, Y\. Wen, and W\. Liu\(2026\)SymbOmni: evolving agentic omni models via symbolic concept learning\.InProceedings of the European Conference on Computer Vision \(ECCV\),Cited by:[§A\.3](https://arxiv.org/html/2609.16057#A1.SS3.p1.1),[§A\.6](https://arxiv.org/html/2609.16057#A1.SS6.p1.1),[Table 13](https://arxiv.org/html/2609.16057#A1.T13.5.5.1),[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.16.1.1),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.5.5.1.1),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.5.8.1.1),[Table 17](https://arxiv.org/html/2609.16057#A1.T17.5.5.1.1),[§1](https://arxiv.org/html/2609.16057#S1.p2.1),[§2](https://arxiv.org/html/2609.16057#S2.p2.1),[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.16.1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.22.1.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.13.1.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.30.1.1)\.
- \[38\]M\. Liu, X\. Ma, Y\. Xie, Q\. Chen, X\. Liu, and Y\. Ma\(2026\)ROGA: scaling generalist agents for office productivity tasks via tool generation\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 18183–18199\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/1ed4723f12853cbd02aecb8160f5e0c9-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[39\]S\. Liu, Y\. Han, P\. Xing, F\. Yin, R\. Wang, W\. Cheng, J\. Liao, Y\. Wang, H\. Fu, C\. Han, G\. Li, Y\. Peng, Q\. Sun, J\. Wu, Y\. Cai, Z\. Ge, R\. Ming, L\. Xia, X\. Zeng, Y\. Zhu, B\. Jiao, X\. Zhang, G\. Yu, and D\. Jiang\(2025\)Step1X\-edit: a practical framework for general image editing\.arXiv preprint arXiv:2504\.17761\.Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.9.1)\.
- \[40\]Y\. Ma, X\. Liu, X\. Chen, W\. Liu, C\. Wu, Z\. Wu, Z\. Pan, Z\. Xie, H\. Zhang, X\. Yu, L\. Zhao, Y\. Wang, J\. Liu, and C\. Ruan\(2025\)JanusFlow: harmonizing autoregression and rectified flow for unified multimodal understanding and generation\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 7739–7751\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52734.2025.00725)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.14.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.15.1)\.
- \[41\]Y\. Niu, M\. Ning, M\. Zheng, B\. Lin, P\. Jin, J\. Liao, K\. Ning, B\. Zhu, and L\. Yuan\(2025\)WISE: A world knowledge\-informed semantic evaluation for text\-to\-image generation\.CoRRabs/2503\.07265\.External Links:[Link](https://doi.org/10.48550/arXiv.2503.07265),[Document](https://dx.doi.org/10.48550/ARXIV.2503.07265),2503\.07265Cited by:[Figure 16](https://arxiv.org/html/2609.16057#A1.F16),[Figure 16](https://arxiv.org/html/2609.16057#A1.F16.4),[§A\.5](https://arxiv.org/html/2609.16057#A1.SS5.p1.1),[§4\.2](https://arxiv.org/html/2609.16057#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.4)\.
- \[42\]W\. Peng, L\. Meng, Y\. Cai, X\. Zhuang, Y\. Yang, R\. Fang, C\. Wu, J\. Lin, Z\. Wu, and S\. Bai\(2026\)Unified multimodal autoregressive modeling with shared context\-visual tokenizer is key to unification\.External Links:2606\.18249,[Link](https://arxiv.org/abs/2606.18249)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1)\.
- \[43\]D\. Podell, Z\. English, K\. Lacey, A\. Blattmann, T\. Dockhorn, J\. Müller, J\. Penna, and R\. Rombach\(2024\)SDXL: improving latent diffusion models for high\-resolution image synthesis\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 1862–1874\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/081b08068e4733ae3e7ad019fe8d172f-Paper-Conference.pdf)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.5.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.4.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.5.1)\.
- \[44\]A\. Ramesh, P\. Dhariwal, A\. Nichol, C\. Chu, and M\. Chen\(2022\)Hierarchical text\-conditional image generation with clip latents\.External Links:2204\.06125,[Link](https://arxiv.org/abs/2204.06125)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.6.1)\.
- \[45\]R\. Rombach, A\. Blattmann, D\. Lorenz, P\. Esser, and B\. Ommer\(2022\)High\-resolution image synthesis with latent diffusion models\.In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 10674–10685\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52688.2022.01042)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.3.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.4.1),[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.3.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.3.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.4.1)\.
- \[46\]Z\. Su, H\. Wei, K\. Cen, Y\. Wang, G\. Chen, C\. Yuan, and X\. Chu\(2026\)Generation enhances understanding in unified multimodal models via multi\-representation generation\.External Links:2601\.21406,[Link](https://arxiv.org/abs/2601.21406)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1)\.
- \[47\]P\. Sun, Y\. Jiang, S\. Chen, S\. Zhang, B\. Peng, P\. Luo, and Z\. Yuan\(2024\)Autoregressive model beats diffusion: llama for scalable image generation\.arXiv preprint arXiv:2406\.06525\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.9.1)\.
- \[48\]Q\. Sun, Y\. Cui, X\. Zhang, F\. Zhang, Q\. Yu, Y\. Wang, Y\. Rao, J\. Liu, T\. Huang, and X\. Wang\(2024\)Generative multimodal models are in\-context learners\.In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 14398–14409\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52733.2024.01365)Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.10.1)\.
- \[49\]K\. Wang, G\. Zhang, M\. Ye, X\. Deng, D\. Wang, X\. Hu, J\. Guo, Y\. Liu, and Y\. Guo\(2026\)MAS2\{\}^\{2\}: self\-generative, self\-configuring, self\-rectifying multi\-agent systems\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 113586–113613\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/b88318174aad2cc174a4e05ab6bfad80-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[50\]X\. Wang, X\. Zhang, Z\. Luo, Q\. Sun, Y\. Cui, J\. Wang, F\. Zhang, Y\. Wang, Z\. Li, Q\. Yu, Y\. Zhao, Y\. Ao, X\. Min, T\. Li, B\. Wu, B\. Zhao, B\. Zhang, L\. Wang, G\. Liu, Z\. He, X\. Yang, J\. Liu, Y\. Lin, T\. Huang, and Z\. Wang\(2024\)Emu3: next\-token prediction is all you need\.External Links:2409\.18869,[Link](https://arxiv.org/abs/2409.18869)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.12.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.23.1)\.
- \[51\]X\. Wang, J\. Wei, D\. Schuurmans, Q\. V\. Le, E\. H\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou\(2023\)Self\-Consistency Improves Chain of Thought Reasoning in Language Models\.InInternational Conference on Learning Representations,External Links:[Link](https://mlanthology.org/iclr/2023/wang2023iclr-selfconsistency/)Cited by:[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.6.1)\.
- \[52\]Y\. Wang\(1916\)The philosophy of wang yang\-ming\.Open Court Publishing Company\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p4.1),[§3\.1](https://arxiv.org/html/2609.16057#S3.SS1.p1.1)\.
- \[53\]J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, b\. ichter, F\. Xia, E\. Chi, Q\. V\. Le, and D\. Zhou\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InAdvances in Neural Information Processing Systems,S\. Koyejo, S\. Mohamed, A\. Agarwal, D\. Belgrave, K\. Cho, and A\. Oh \(Eds\.\),Vol\.35,pp\. 24824–24837\.External Links:[Document](https://dx.doi.org/10.52202/068431-1800),[Link](https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf)Cited by:[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.5.1)\.
- \[54\]Y\. Wen, Y\. Ye, Y\. Zhang, D\. Yang, and H\. Zhu\(2026\)Real\-time reasoning agents in evolving environments\.InInternational Conference on Learning Representations,External Links:[Link](https://iclr.cc/virtual/2026/poster/10007516)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[55\]C\. Wu, J\. Li, J\. Zhou, J\. Lin, K\. Gao, K\. Yan, S\. Yin, S\. Bai, X\. Xu, Y\. Chen, Y\. Chen, Z\. Tang, Z\. Zhang, Z\. Wang, A\. Yang, B\. Yu, C\. Cheng, D\. Liu, D\. Li, H\. Zhang, H\. Meng, H\. Wei, J\. Ni, K\. Chen, K\. Cao, L\. Peng, L\. Qu, M\. Wu, P\. Wang, S\. Yu, T\. Wen, W\. Feng, X\. Xu, Y\. Wang, Y\. Zhang, Y\. Zhu, Y\. Wu, Y\. Cai, and Z\. Liu\(2025\)Qwen\-image technical report\.External Links:2508\.02324,[Link](https://arxiv.org/abs/2508.02324)Cited by:[Table 3](https://arxiv.org/html/2609.16057#S4.T3.5.10.1)\.
- \[56\]C\. Wu, X\. Chen, Z\. Wu, Y\. Ma, X\. Liu, Z\. Pan, W\. Liu, Z\. Xie, X\. Yu, C\. Ruan, and P\. Luo\(2025\)Janus: decoupling visual encoding for unified multimodal understanding and generation\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 12966–12977\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52734.2025.01210)Cited by:[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.13.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.14.1)\.
- \[57\]Y\. Wu, Z\. Zhang, J\. Chen, H\. Tang, D\. Li, Y\. Fang, L\. Zhu, E\. Xie, H\. Yin, L\. Yi, S\. Han, and Y\. Lu\(2025\)VILA\-u: a unified foundation model integrating visual understanding and generation\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 93620–93638\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/e9e140df6de01afb672cb859d203c307-Paper-Conference.pdf)Cited by:[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.20.1)\.
- \[58\]Y\. Wu, Z\. Li, X\. Hu, X\. Ye, X\. Zeng, G\. Yu, W\. Zhu, B\. Schiele, M\. Yang, and X\. Yang\(2025\)KRIS\-bench: benchmarking next\-level intelligent image editing models\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38, Main Conference,pp\.\.External Links:[Document](https://dx.doi.org/10.52202/085713-5242),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/e619b285582fb12f4c3de3a507b8b99c-Paper-Datasets_and_Benchmarks_Track.pdf)Cited by:[Figure 17](https://arxiv.org/html/2609.16057#A1.F17),[Figure 17](https://arxiv.org/html/2609.16057#A1.F17.4),[§A\.5\.6](https://arxiv.org/html/2609.16057#A1.SS5.SSS6.p1.1),[§A\.5](https://arxiv.org/html/2609.16057#A1.SS5.p1.1),[Table 15](https://arxiv.org/html/2609.16057#A1.T15),[Table 15](https://arxiv.org/html/2609.16057#A1.T15.4)\.
- \[59\]S\. Xiao, Y\. Wang, J\. Zhou, H\. Yuan, X\. Xing, R\. Yan, C\. Li, S\. Wang, T\. Huang, and Z\. Liu\(2025\)OmniGen: unified image generation\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 13294–13304\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52734.2025.01241)Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.13.1)\.
- \[60\]J\. Xie, W\. Mao, Z\. Bai, D\. J\. Zhang, W\. Wang, K\. Q\. Lin, Y\. Gu, Z\. Chen, Z\. Yang, and M\. Z\. Shou\(2025\)Show\-o: one single transformer to unify multimodal understanding and generation\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 28240–28264\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/45f0d179ef7e10eb7366550cd4e574ae-Paper-Conference.pdf)Cited by:[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.18.1),[Table 4](https://arxiv.org/html/2609.16057#S4.T4.5.19.1)\.
- \[61\]J\. Xie, Z\. Yang, and M\. Z\. Shou\(2025\)Show\-o2: improved native unified multimodal models\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38, Main Conference,pp\. 47490–47518\.External Links:[Document](https://dx.doi.org/10.52202/085713-1584),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/43e8fd8b9581faa71a6a61602bc28435-Paper-Conference.pdf)Cited by:[§2](https://arxiv.org/html/2609.16057#S2.p1.1)\.
- \[62\]H\. Xin, X\. Li, Z\. Liu, Y\. Yan, S\. Wang, C\. Yang, Y\. Gu, G\. Yu, and M\. Sun\(2026\)MetaMem: evolving meta\-memory for knowledge utilization through self\-reflective symbolic optimization\.InFindings of the Association for Computational Linguistics: ACL 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 5473–5492\.External Links:[Link](https://aclanthology.org/2026.findings-acl.270/),[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.270),ISBN 979\-8\-89176\-395\-1Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[63\]Z\. Xiong, Y\. Lin, W\. Xie, P\. He, Z\. Liu, J\. Tang, H\. Lakkaraju, and Z\. Xiang\(2026\)How memory management impacts LLM agents: an empirical study of experience\-following behavior\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 623–645\.External Links:[Link](https://aclanthology.org/2026.acl-long.27/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.27),ISBN 979\-8\-89176\-390\-6Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[64\]Y\. Xu, J\. Kong, J\. Wang, X\. Pan, B\. Lin, and Q\. Liu\(2025\)InsightEdit: towards better instruction following for image editing\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 2694–2703\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52734.2025.00257)Cited by:[Table 5](https://arxiv.org/html/2609.16057#S4.T5.5.8.1)\.
- \[65\]X\. Xue, Z\. Lu, D\. Huang, Z\. Wang, W\. Ouyang, and L\. Bai\(2025\)ComfyBench: benchmarking llm\-based agents in comfyui for autonomously designing collaborative ai systems\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 24614–24624\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52734.2025.02292)Cited by:[Figure 10](https://arxiv.org/html/2609.16057#A1.F10),[Figure 10](https://arxiv.org/html/2609.16057#A1.F10.4),[Figure 11](https://arxiv.org/html/2609.16057#A1.F11),[Figure 11](https://arxiv.org/html/2609.16057#A1.F11.4),[Figure 12](https://arxiv.org/html/2609.16057#A1.F12),[Figure 12](https://arxiv.org/html/2609.16057#A1.F12.4),[Figure 18](https://arxiv.org/html/2609.16057#A1.F18),[Figure 18](https://arxiv.org/html/2609.16057#A1.F18.4),[Figure 19](https://arxiv.org/html/2609.16057#A1.F19),[Figure 19](https://arxiv.org/html/2609.16057#A1.F19.4),[Figure 21](https://arxiv.org/html/2609.16057#A1.F21),[Figure 21](https://arxiv.org/html/2609.16057#A1.F21.4),[Figure 7](https://arxiv.org/html/2609.16057#A1.F7),[Figure 7](https://arxiv.org/html/2609.16057#A1.F7.4),[Figure 8](https://arxiv.org/html/2609.16057#A1.F8),[Figure 8](https://arxiv.org/html/2609.16057#A1.F8.7),[Figure 9](https://arxiv.org/html/2609.16057#A1.F9),[Figure 9](https://arxiv.org/html/2609.16057#A1.F9.4),[§A\.3](https://arxiv.org/html/2609.16057#A1.SS3.p1.1),[§A\.4\.3](https://arxiv.org/html/2609.16057#A1.SS4.SSS3.p1.1),[§A\.5](https://arxiv.org/html/2609.16057#A1.SS5.p1.1),[§A\.6](https://arxiv.org/html/2609.16057#A1.SS6.p1.1),[§A\.8](https://arxiv.org/html/2609.16057#A1.SS8.p1.1),[Table 13](https://arxiv.org/html/2609.16057#A1.T13),[Table 13](https://arxiv.org/html/2609.16057#A1.T13.4),[Table 13](https://arxiv.org/html/2609.16057#A1.T13.5.3.1),[Table 14](https://arxiv.org/html/2609.16057#A1.T14),[Table 14](https://arxiv.org/html/2609.16057#A1.T14.4),[Table 16](https://arxiv.org/html/2609.16057#A1.T16),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.4),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.5.3.1.1),[Table 16](https://arxiv.org/html/2609.16057#A1.T16.5.6.1.1),[Table 17](https://arxiv.org/html/2609.16057#A1.T17),[Table 17](https://arxiv.org/html/2609.16057#A1.T17.4),[Table 17](https://arxiv.org/html/2609.16057#A1.T17.5.3.1.1),[§1](https://arxiv.org/html/2609.16057#S1.p2.1),[§2](https://arxiv.org/html/2609.16057#S2.p2.1),[Table 1](https://arxiv.org/html/2609.16057#S3.T1),[Table 1](https://arxiv.org/html/2609.16057#S3.T1.4),[Table 1](https://arxiv.org/html/2609.16057#S3.T1.5.12.1.1),[§4\.1](https://arxiv.org/html/2609.16057#S4.SS1.p1.1),[§4\.4](https://arxiv.org/html/2609.16057#S4.SS4.p1.1),[Table 2](https://arxiv.org/html/2609.16057#S4.T2.5.20.1.1),[Table 6](https://arxiv.org/html/2609.16057#S4.T6),[Table 6](https://arxiv.org/html/2609.16057#S4.T6.4)\.
- \[66\]Q\. Yu, W\. Chow, Z\. Yue, K\. Pan, Y\. Wu, X\. Wan, J\. Li, S\. Tang, H\. Zhang, and Y\. Zhuang\(2025\)AnyEdit: mastering unified high\-quality image editing for any idea\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 26125–26135\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52734.2025.02433)Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.11.1)\.
- \[67\]Z\. Zeng, H\. Hua, and J\. Luo\(2026\)MIRA: multimodal iterative reasoning agent for image editing\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\) Findings,pp\. 9563–9573\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[68\]G\. Zhang, L\. Lu, F\. Xie, K\. Zhu, J\. Wang, Z\. Xie, Z\. Yu, Z\. Liu, Z\. Sun, Q\. Li, Y\. Liao, H\. Chang, X\. Hu, Q\. Ren, W\. Zhou, C\. Hu, Y\. Deng, and S\. Yan\(2026\)JIT\-agent: scaling harness intelligence via just\-in\-time harness evolution\.External Links:2608\.25593,[Link](https://arxiv.org/abs/2608.25593)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p3.1),[§2](https://arxiv.org/html/2609.16057#S2.p2.1)\.
- \[69\]G\. Zhang, J\. Wang, J\. Chen, W\. Zhou, K\. Wang, and S\. YAN\(2026\)AgenTracer: who is inducing failure in the llm agentic systems?\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 11377–11399\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/134ed7a477770f227f12450ef0cbb8f4-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[70\]K\. Zhang, L\. Mo, W\. Chen, H\. Sun, and Y\. Su\(2023\)MagicBrush: a manually annotated dataset for instruction\-guided image editing\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 31428–31449\.External Links:[Document](https://dx.doi.org/10.52202/075280-1365),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/64008fa30cba9b4d1ab1bd3bd3d57d61-Paper-Datasets_and_Benchmarks.pdf)Cited by:[Table 15](https://arxiv.org/html/2609.16057#A1.T15.5.12.1),[Table 5](https://arxiv.org/html/2609.16057#S4.T5.5.4.1)\.
- \[71\]S\. Zhang, Z\. Li, Y\. Zhang, J\. Fu, L\. Song, J\. Bian, J\. Zhang, Y\. Yang, and R\. Wang\(2026\)PixelCraft: a multi\-agent system for high\-fidelity visual reasoning on structured images\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 77608–77632\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/7d90c28e7820709792d969211815a2b3-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[72\]Y\. Zhang, H\. Zhang, F\. Gao, X\. Li, Z\. Liu, C\. Zhu, J\. Qiu, Y\. Yan, J\. Liu, W\. Tang, Z\. Fang, Y\. Nie, C\. Wei, Y\. Wang, W\. Ding, and C\. Yu\(2026\)Harness vla: steering frozen vlas into reliable manipulation primitives via memory\-guided agents\.External Links:2607\.08448,[Link](https://arxiv.org/abs/2607.08448)Cited by:[§A\.3](https://arxiv.org/html/2609.16057#A1.SS3.p1.1)\.
- \[73\]Y\. Zhang, S\. Huang, and L\. Wang\(2026\)IntentEdit: multi\-agent reasoning for intent\-driven complex image editing\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\) Findings,pp\. 8776–8785\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[74\]W\. Zhao, Y\. Zhang, Y\. Wang, Y\. Deng, Y\. Zhao, X\. Zhi, Y\. Huang, H\. He, W\. Che, B\. Qin, and T\. Liu\(2026\)On safety risks in experience\-driven self\-evolving agents\.InFindings of the Association for Computational Linguistics: ACL 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 42145–42169\.External Links:[Link](https://aclanthology.org/2026.findings-acl.2091/),[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.2091),ISBN 979\-8\-89176\-395\-1Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.
- \[75\]C\. Zhou, T\. Xu, J\. Lin, and D\. Ge\(2026\)StepORLM: a self\-evolving framework with generative process supervision for operations research language models\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 6914–6940\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/0bcfb525c8f8f07ae10a93d0b2a40e00-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p2.1)\.
- \[76\]Z\. Zhu, Y\. Sun, M\. Zhu, W\. Ma, S\. Zhang, Z\. Wang, Z\. Yang, D\. Zhang, K\. Lan, Z\. Zhao, D\. Liu, S\. Xiang, L\. Chen, and K\. Yu\(2026\)IEA: amateur\-friendly conversational image editing agent via three stages of multitask alignment\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\) Findings,pp\. 8672–8683\.Cited by:[§1](https://arxiv.org/html/2609.16057#S1.p1.1)\.

## Appendix AAppendix

### Appendix Contents

### A\.1Problem Setup and Symbolic Policy Learning

#### A\.1\.1Problem Formulation and Policy Learning Loop

This section details the operations and score computations in the main paper\. We distinguish the ComfyUI knowledge graph𝒦\\mathcal\{K\}, the policy library state𝒮t\\mathcal\{S\}\_\{t\}, and the frozen policy snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}\.

A visual workflow agent𝒜\\mathcal\{A\}receives an external task instructionℓ\\ell, scene contextcc, and ComfyUI knowledge graph𝒦\\mathcal\{K\}, and produces an executable workflow

π=𝒜⁡\(ℓ,c,𝒦\)\.\\pi=\\mathcal\{A\}\(\\ell,c,\\mathcal\{K\}\)\.\(1\)The instructionℓ\\ellcorresponds to the descriptionqτq\_\{\\tau\}in the structured task representation below\. OmniHarness introduces self\-directed inquiry to generate and execute practice tasks before downstream objectives are specified\.

Given a generative capability space𝒢\\mathcal\{G\}, a ComfyUI knowledge graph𝒦\\mathcal\{K\}, a source image pool𝒳\\mathcal\{X\}, andNNinquiry iterations, OmniHarness maintains scene contextctc\_\{t\}and policy library state𝒮t=\(ℒt,ℱt\)\\mathcal\{S\}\_\{t\}=\(\\mathcal\{L\}\_\{t\},\\mathcal\{F\}\_\{t\}\)\. The workflow libraryℒt\\mathcal\{L\}\_\{t\}stores symbolic policies as reusable workflow templates for visual generation task families, together with reliability evidence\. The failure libraryℱt\\mathcal\{F\}\_\{t\}stores failure evidence and corrective strategies\. The contextctc\_\{t\}summarizes capability coverage, workflow reliability, and available source images\. Initialization sets𝒮1=\(∅,∅\)\\mathcal\{S\}\_\{1\}=\(\\emptyset,\\emptyset\)andc1=InitContext⁡\(𝒳\)c\_\{1\}=\\operatorname\{InitContext\}\(\\mathcal\{X\}\)\.

At inquiry iterationtt, the proposer generates candidate tasks

𝒯t=Propose⁡\(𝒢,ct,𝒮t\)\.\\mathcal\{T\}\_\{t\}=\\operatorname\{Propose\}\(\\mathcal\{G\},c\_\{t\},\\mathcal\{S\}\_\{t\}\)\.\(2\)The proposer considers capability coverage, available source images, reusable policies, and observed failures\.

OmniHarness selects the next task according to

τt=arg​maxτ∈𝒯t⁡𝒩​\(τ\)​𝒞​\(τ\)\.\\tau\_\{t\}=\\operatorname\*\{arg\\,max\}\_\{\\tau\\in\\mathcal\{T\}\_\{t\}\}\\mathcal\{N\}\(\\tau\)\\mathcal\{C\}\(\\tau\)\.\(3\)Capability novelty𝒩⁡\(τ\)\\mathcal\{N\}\(\\tau\)favors underexplored context–capability pairs, while the competence frontier score𝒞⁡\(τ\)\\mathcal\{C\}\(\\tau\)favors tasks near the estimated competence frontier\. Their product balances novelty with estimated learnability\. Time indices are omitted from both scores\.

The harness performs feedback\-guided execution through

\(πt,ot\)=Run⁡\(τt,ct,𝒦,𝒮t\)\.\(\\pi\_\{t\},o\_\{t\}\)=\\operatorname\{Run\}\(\\tau\_\{t\},c\_\{t\},\\mathcal\{K\},\\mathcal\{S\}\_\{t\}\)\.\(4\)Here,πt\\pi\_\{t\}is the executable workflow forτt\\tau\_\{t\}, andoto\_\{t\}records execution status, verifier feedback, and evidence, including intermediate outputs and recovery results\. The operationRun\\operatorname\{Run\}includes planning, Code\-as\-Policy synthesis, execution, intermediate verification, and localized recovery\. It uses𝒦\\mathcal\{K\}to construct workflows, instantiates, adapts, and composes policies fromℒt\\mathcal\{L\}\_\{t\}, and consults failure patterns and remedies inℱt\\mathcal\{F\}\_\{t\}\.

After execution, OmniHarness updates its context and policy library

\(ct\+1,𝒮t\+1\)=Update⁡\(ct,𝒮t,τt,πt,ot\)\.\(c\_\{t\+1\},\\mathcal\{S\}\_\{t\+1\}\)=\\operatorname\{Update\}\(c\_\{t\},\\mathcal\{S\}\_\{t\},\\tau\_\{t\},\\pi\_\{t\},o\_\{t\}\)\.\(5\)Verified executions are abstracted into symbolic policies inℒt\+1\\mathcal\{L\}\_\{t\+1\}, capturing shared procedures and applicability conditions for visual generation task families while removing instance\-specific inputs\. Failure evidence and corrective strategies are recorded inℱt\+1\\mathcal\{F\}\_\{t\+1\}\. Updates apply to both practice and downstream tasks while model parameters remain fixed\. The exported snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}remains frozen for plug\-and\-play reuse by external visual agents\.

#### A\.1\.2Exploration within Reach

Each candidate task is represented asτ=\(qτ,mτ,𝒢τ,xτ\)\\tau=\(q\_\{\\tau\},m\_\{\\tau\},\\mathcal\{G\}\_\{\\tau\},x\_\{\\tau\}\), whereqτq\_\{\\tau\}is the task description,mτ∈\{T2I,I2I\}m\_\{\\tau\}\\in\\\{\\mathrm\{T2I\},\\mathrm\{I2I\}\\\}is the generation modality, and𝒢τ⊆𝒢\\mathcal\{G\}\_\{\\tau\}\\subseteq\\mathcal\{G\}is the required capability set\. The source image isxτ∈𝒳∪\{⊥\}x\_\{\\tau\}\\in\\mathcal\{X\}\\cup\\\{\\bot\\\}, withxτ=⊥x\_\{\\tau\}=\\botfor T2I\. Candidate generation encourages diverse capability combinations, filters near\-duplicates, and avoids known failure patterns inℱt\\mathcal\{F\}\_\{t\}\.

Capability novelty\.Letzτz\_\{\\tau\}denote the task’s context key\. For T2I,zτ=zT2Iz\_\{\\tau\}=z\_\{\\mathrm\{T2I\}\}, wherezT2Iz\_\{\\mathrm\{T2I\}\}is a fixed context token\. For I2I,zτ=xτz\_\{\\tau\}=x\_\{\\tau\}\. Letnt​\(z,g\)n\_\{t\}\(z,g\)count practice attempts involving contextzzand capabilityggbefore iterationtt\. Capability novelty is

𝒩⁡\(τ\)=1\|𝒢τ\|​∑g∈𝒢τ1nt​\(zτ,g\)\+1\.\\mathcal\{N\}\(\\tau\)=\\frac\{1\}\{\|\\mathcal\{G\}\_\{\\tau\}\|\}\\sum\_\{g\\in\\mathcal\{G\}\_\{\\tau\}\}\\frac\{1\}\{\\sqrt\{n\_\{t\}\(z\_\{\\tau\},g\)\+1\}\}\.\(6\)Underexplored pairs receive higher scores\. Averaging over𝒢τ\\mathcal\{G\}\_\{\\tau\}prevents a larger capability set from increasing novelty solely through its size\.

Conservative workflow reliability\.Novelty alone may favor tasks with insufficient support from the workflow library\. OmniHarness therefore estimates reliability before computing the competence frontier score\. For each required capabilityg∈𝒢τg\\in\\mathcal\{G\}\_\{\\tau\}, let𝒲t​\(τ,g\)⊆ℒt\\mathcal\{W\}\_\{t\}\(\\tau,g\)\\subseteq\\mathcal\{L\}\_\{t\}contain applicable, non\-suspended workflows that supportgg\. Each workflowwwrecords a usage countuw≥1u\_\{w\}\\geq 1and verified success countsws\_\{w\}, with empirical success ratepw=sw/uwp\_\{w\}=s\_\{w\}/u\_\{w\}\.

Reliability is the lower endpoint of a 95% Wilson confidence interval

r^​\(w\)=pw\+κ22​uw−κ​pw​\(1−pw\)uw\+κ24​uw21\+κ2uw,κ=1\.96\.\\widehat\{r\}\(w\)=\\frac\{p\_\{w\}\+\\frac\{\\kappa^\{2\}\}\{2u\_\{w\}\}\-\\kappa\\sqrt\{\\frac\{p\_\{w\}\(1\-p\_\{w\}\)\}\{u\_\{w\}\}\+\\frac\{\\kappa^\{2\}\}\{4u\_\{w\}^\{2\}\}\}\}\{1\+\\frac\{\\kappa^\{2\}\}\{u\_\{w\}\}\},\\qquad\\kappa=1\.96\.\(7\)This discounts success rates supported by limited execution evidence\.

Capability support is estimated from the most reliable applicable workflow

rt​\(τ,g\)=\{maxw∈𝒲t​\(τ,g\)⁡r^​\(w\),𝒲t​\(τ,g\)≠∅,ε,𝒲t​\(τ,g\)=∅,r\_\{t\}\(\\tau,g\)=\\begin\{cases\}\\displaystyle\\max\_\{w\\in\\mathcal\{W\}\_\{t\}\(\\tau,g\)\}\\widehat\{r\}\(w\),&\\mathcal\{W\}\_\{t\}\(\\tau,g\)\\neq\\emptyset,\\\\\[6\.0pt\] \\varepsilon,&\\mathcal\{W\}\_\{t\}\(\\tau,g\)=\\emptyset,\\end\{cases\}\(8\)whereε=0\.05\\varepsilon=0\.05provides a small prior when no applicable workflow exists\.

Estimated task competence follows the weakest required capability

r¯t​\(τ\)=ming∈𝒢τ⁡rt​\(τ,g\)\.\\bar\{r\}\_\{t\}\(\\tau\)=\\min\_\{g\\in\\mathcal\{G\}\_\{\\tau\}\}r\_\{t\}\(\\tau,g\)\.\(9\)
Competence frontier score\.Following the Goldilocks principle\[[1](https://arxiv.org/html/2609.16057#bib.bib1),[26](https://arxiv.org/html/2609.16057#bib.bib2)\], OmniHarness uses the learnability heuristic

𝒞⁡\(τ\)=4​r¯t​\(τ\)​\(1−r¯t​\(τ\)\)\.\\mathcal\{C\}\(\\tau\)=4\\bar\{r\}\_\{t\}\(\\tau\)\\left\(1\-\\bar\{r\}\_\{t\}\(\\tau\)\\right\)\.\(10\)It peaks atr¯t​\(τ\)=0\.5\\bar\{r\}\_\{t\}\(\\tau\)=0\.5and downweights tasks with very low or high estimated competence\. Combined with capability novelty in \([3](https://arxiv.org/html/2609.16057#A1.E3)\), it favors novel tasks near the competence frontier\. Task selection adapts as attempts accumulate and𝒮t\\mathcal\{S\}\_\{t\}changes\.

#### A\.1\.3Feedback\-Guided Execution

Agents share the task and context\(τt,ct,𝒦,𝒮t\)\(\\tau\_\{t\},c\_\{t\},\\mathcal\{K\},\\mathcal\{S\}\_\{t\}\), together with plans, programs, verification results, and corrections\.

The planner decomposesτt\\tau\_\{t\}into an ordered planρt=\(at,1,…,at,Jt\)\\rho\_\{t\}=\(a\_\{t,1\},\\ldots,a\_\{t,J\_\{t\}\}\), whereJtJ\_\{t\}is the number of steps andat,ja\_\{t,j\}is thejj\-th generation step

ρt=Plan⁡\(τt,ct,𝒦,𝒮t\)\.\\rho\_\{t\}=\\operatorname\{Plan\}\(\\tau\_\{t\},c\_\{t\},\\mathcal\{K\},\\mathcal\{S\}\_\{t\}\)\.\(11\)It uses𝒦\\mathcal\{K\}to construct the plan and instantiates, adapts, and composes policies fromℒt\\mathcal\{L\}\_\{t\}, guided by failure patterns and remedies inℱt\\mathcal\{F\}\_\{t\}\. The plan specifies step dependencies and verification criteria\. A plan verifier checks step order, dependencies, and task alignment\.

The workflow writer converts the verified plan into a Python\-like Code\-as\-Policy program

π~t=Write⁡\(ρt,𝒦\)\.\\widetilde\{\\pi\}\_\{t\}=\\operatorname\{Write\}\(\\rho\_\{t\},\\mathcal\{K\}\)\.\(12\)Function calls represent ComfyUI nodes, arguments specify node parameters, and data flow defines their connections\. Workflows and components are reused when their preconditions hold\.

A reversible interpreter compiles the program into an executable workflow

πt=Compile⁡\(π~t\)\.\\pi\_\{t\}=\\operatorname\{Compile\}\(\\widetilde\{\\pi\}\_\{t\}\)\.\(13\)The resultingπt\\pi\_\{t\}is the execution instance forτt\\tau\_\{t\}, preserving the program’s node types, parameters, dependencies, and execution order\. The executor runsπt\\pi\_\{t\}and records intermediate outputs\.

Verification covers four levels\. The plan verifier checks the plan before synthesis\. The workflow validator checks graph validity and executability in the current ComfyUI environment\. Intermediate verification checks each output against the intended effect ofat,ja\_\{t,j\}\. The goal verifier checks the final output againstqτtq\_\{\\tau\_\{t\}\}and the constraints inτt\\tau\_\{t\}\. Verifier feedback and evidence are recorded inoto\_\{t\}\.

On failure, the diagnoser identifies the affected step, analyzes its cause, and retrieves a correction fromℱt\\mathcal\{F\}\_\{t\}\. The harness repairs that component while preserving verified steps\. If the failure persists, a subagent solves the affected step and returns a reusable subworkflow\. The repaired component is integrated intoπt\\pi\_\{t\}and verified again\. Verification and recovery repeat until success or the retry budget is exhausted\.

#### A\.1\.4Symbolic Policy Learning and Export

Whenoto\_\{t\}confirms success,πt\\pi\_\{t\}is abstracted into a symbolic policy represented by a workflow entrywwinℒt\+1\\mathcal\{L\}\_\{t\+1\}\. The entry stores a reusable workflow template, modality, capability categories, description, preconditions, expected effects, and dependencies\. The template retains the shared ComfyUI graph structure while removing instance\-specific inputs, includingqτtq\_\{\\tau\_\{t\}\}andxτtx\_\{\\tau\_\{t\}\}\. Its procedures and applicability conditions support instantiation, adaptation, and composition for new tasks within a visual generation task family\.

Before insertion, OmniHarness compares the entry with existing workflows by function and executable structure\. Equivalent entries are updated rather than duplicated\. Each invocation ofwwincrementsuwu\_\{w\}, whilesws\_\{w\}increases only after verified success\.

A new entry starts in theProvisionaltier\. It is promoted toValidatedwhenuw≥3u\_\{w\}\\geq 3andpw≥0\.5p\_\{w\}\\geq 0\.5, and assigned toSuspendedwhenuw≥10u\_\{w\}\\geq 10andpw≤0\.2p\_\{w\}\\leq 0\.2\. Validated workflows are prioritized during retrieval, while suspended workflows are excluded from normal reuse\. The empirical ratepwp\_\{w\}controls tier updates, whereasr^​\(w\)\\widehat\{r\}\(w\)provides the conservative reliability estimate for task selection\.

Failures are recorded inℱt\+1\\mathcal\{F\}\_\{t\+1\}withτt\\tau\_\{t\},πt\\pi\_\{t\}, input conditions, generation constraints, intermediate outputs, execution feedback, and verifier results fromoto\_\{t\}\. Their analysis yields corrective strategies specifying root causes, workflow antipatterns, remedies, and applicable scope\. The scope may include modalities, capability categories, workflow stages, nodes, or models\.

A curator periodically consolidatesℒt\\mathcal\{L\}\_\{t\}andℱt\\mathcal\{F\}\_\{t\}\. It merges redundant workflows, refines their preconditions and expected effects, consolidates corrective strategies, and updates statistics and reliability tiers\. When repeated failures reveal a missing capability, it may trigger workflow construction and verification\. The updated libraries and source image pool𝒳\\mathcal\{X\}refreshct\+1c\_\{t\+1\}to guide future task proposals\.

AfterNNinquiry iterations, OmniHarness exports a frozen copy of the learned policy library

𝒦inquiry=Snapshot⁡\(𝒮N\+1\)\.\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}=\\operatorname\{Snapshot\}\(\\mathcal\{S\}\_\{N\+1\}\)\.\(14\)The snapshot contains both the workflow library and the failure library\. It supports plug\-and\-play reuse by external visual agents and remains fixed during evaluation\.

For downstream OmniHarness execution, let𝒟=\(ℓj\)j=1M\\mathcal\{D\}=\(\\ell\_\{j\}\)\_\{j=1\}^\{M\}denote a stream ofMMexternal task instructions\. Execution starts from the final inquiry state

\(c,𝒮\)=Execute⁡\(𝒟,cN\+1,𝒦,𝒮N\+1\)\.\(c,\\mathcal\{S\}\)=\\operatorname\{Execute\}\\left\(\\mathcal\{D\},c\_\{N\+1\},\\mathcal\{K\},\\mathcal\{S\}\_\{N\+1\}\\right\)\.\(15\)Each instruction is represented as a downstream taskτt\\tau\_\{t\}and processed by the same feedback\-guided execution and policy update operations\. The internal policy library𝒮\\mathcal\{S\}continues to learn from downstream feedback while model parameters remain fixed\. The exported snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}receives no downstream updates\.

#### A\.1\.5Overall Symbolic Policy Learning Procedure

Algorithm[1](https://arxiv.org/html/2609.16057#alg1)summarizes the complete OmniHarness procedure\. Self\-directed inquiry generates practice tasks before downstream objectives are specified\. Feedback\-guided execution provides verified experience for learning symbolic policies for visual generation task families\. These policies are adapted and refined during downstream execution, while a frozen snapshot of the inquiry state supports plug\-and\-play reuse\.

Algorithm 1Symbolic Policy Learning in OmniHarness1:Generative capability space

𝒢\\mathcal\{G\}, ComfyUI knowledge graph

𝒦\\mathcal\{K\}, source image pool

𝒳\\mathcal\{X\},

2:inquiry iterations

NN, and downstream task stream

𝒟=\(ℓj\)j=1M\\mathcal\{D\}=\(\\ell\_\{j\}\)\_\{j=1\}^\{M\}
3:Evolving policy library state

𝒮=\(ℒ,ℱ\)\\mathcal\{S\}=\(\\mathcal\{L\},\\mathcal\{F\}\)and frozen policy snapshot

𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}
4:

𝒮1←\(∅,∅\),c1←InitContext⁡\(𝒳\)\\mathcal\{S\}\_\{1\}\\leftarrow\(\\emptyset,\\emptyset\),\\quad c\_\{1\}\\leftarrow\\operatorname\{InitContext\}\(\\mathcal\{X\}\)⊳\\trianglerightinitialize the context and policy library

5:for

t=1,…,Nt=1,\\ldots,Ndo

6:

𝒯t←Propose⁡\(𝒢,ct,𝒮t\)\\mathcal\{T\}\_\{t\}\\leftarrow\\operatorname\{Propose\}\(\\mathcal\{G\},c\_\{t\},\\mathcal\{S\}\_\{t\}\)⊳\\trianglerightStep 1: propose and select a practice task

7:

τt←arg​maxτ∈𝒯t⁡𝒩​\(τ\)​𝒞​\(τ\)\\tau\_\{t\}\\leftarrow\\operatorname\*\{arg\\,max\}\_\{\\tau\\in\\mathcal\{T\}\_\{t\}\}\\mathcal\{N\}\(\\tau\)\\mathcal\{C\}\(\\tau\)
8:

\(πt,ot\)←Run⁡\(τt,ct,𝒦,𝒮t\)\(\\pi\_\{t\},o\_\{t\}\)\\leftarrow\\operatorname\{Run\}\(\\tau\_\{t\},c\_\{t\},\\mathcal\{K\},\\mathcal\{S\}\_\{t\}\)⊳\\trianglerightStep 2: plan, execute, verify, and recover

9:

\(ct\+1,𝒮t\+1\)←Update⁡\(ct,𝒮t,τt,πt,ot\)\(c\_\{t\+1\},\\mathcal\{S\}\_\{t\+1\}\)\\leftarrow\\operatorname\{Update\}\(c\_\{t\},\\mathcal\{S\}\_\{t\},\\tau\_\{t\},\\pi\_\{t\},o\_\{t\}\)⊳\\trianglerightStep 3: update the policy library

10:endfor

11:

\(c,𝒮\)←Execute⁡\(𝒟,cN\+1,𝒦,𝒮N\+1\)\(c,\\mathcal\{S\}\)\\leftarrow\\operatorname\{Execute\}\(\\mathcal\{D\},c\_\{N\+1\},\\mathcal\{K\},\\mathcal\{S\}\_\{N\+1\}\)⊳\\trianglerightStep 4a: adapt and refine policies downstream

12:

𝒦inquiry←Snapshot⁡\(𝒮N\+1\)\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}\\leftarrow\\operatorname\{Snapshot\}\(\\mathcal\{S\}\_\{N\+1\}\)⊳\\trianglerightStep 4b: export the frozen inquiry snapshot

13:return

𝒮,𝒦inquiry\\mathcal\{S\},\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}

### A\.2Self\-Directed Inquiry Configuration

Unless otherwise specified by an ablation, experiments start from the policy library state𝒮N\+1\\mathcal\{S\}\_\{N\+1\}learned throughN=50N=50iterations of self\-directed inquiry\. At each iterationtt, the proposer generates ten candidate practice tasks, with\|𝒯t\|=10\\lvert\\mathcal\{T\}\_\{t\}\\rvert=10, conditioned on the generative capability space𝒢\\mathcal\{G\}, scene contextctc\_\{t\}, and policy library state𝒮t=\(ℒt,ℱt\)\\mathcal\{S\}\_\{t\}=\(\\mathcal\{L\}\_\{t\},\\mathcal\{F\}\_\{t\}\)\. The curator consolidates the workflow libraryℒt\\mathcal\{L\}\_\{t\}and failure libraryℱt\\mathcal\{F\}\_\{t\}every five inquiry iterations\. For I2I practice, the source image pool𝒳\\mathcal\{X\}contains all original ComfyBench images and excludes video inputs\. These raw images are available during inquiry, while downstream task instructions, target outputs, reference workflows, benchmark annotations, and evaluation labels are withheld\.

The full image\-only capability space𝒢\\mathcal\{G\}contains six T2I and six I2I capabilities, listed in Table[9](https://arxiv.org/html/2609.16057#A1.T9)\. For GenEval, GenEval2, and WISE, inquiry uses only the T2I capabilities without access to benchmark evaluation tasks\.

Table 9:Generative capability space𝒢\\mathcal\{G\}for self\-directed inquiry\.The policy library𝒮t=\(ℒt,ℱt\)\\mathcal\{S\}\_\{t\}=\(\\mathcal\{L\}\_\{t\},\\mathcal\{F\}\_\{t\}\)combines reusable workflow templates with failure evidence and corrective strategies\. Verified executions are abstracted into symbolic policies inℒt\\mathcal\{L\}\_\{t\}, capturing shared procedures and applicability conditions for visual generation task families while removing instance\-specific inputs\. The harness instantiates, adapts, and composes these policies for downstream tasks\. Execution feedback continually refines the policy library while model parameters remain fixed\. External visual agents reuse the frozen snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}, which contains both libraries\. Table[10](https://arxiv.org/html/2609.16057#A1.T10)lists the workflow library schema\.

Table 10:Schema of the workflow libraryℒt\\mathcal\{L\}\_\{t\}\.The failure libraryℱt\\mathcal\{F\}\_\{t\}stores failure evidence and corrective strategies used during feedback\-guided execution\. Table[11](https://arxiv.org/html/2609.16057#A1.T11)lists its schema\.

Table 11:Schema of the failure libraryℱt\\mathcal\{F\}\_\{t\}\.The scene contextctc\_\{t\}summarizes capability coverage, workflow reliability, and available source images\. It is updated after each task and refreshed during periodic library consolidation\. Table[12](https://arxiv.org/html/2609.16057#A1.T12)lists its fields\.

Table 12:Fields of the current scene contextctc\_\{t\}\.
### A\.3Environment and Evaluation Protocols

All experiments were conducted within ComfyUI\. We evaluated two reasoning configurations, GPT\-4o and Codex GPT\-4o\. In the first configuration, GPT\-4o directly served as the reasoning engine of OmniHarness\. Following the planner\-instantiation design of Harness VLA\[[72](https://arxiv.org/html/2609.16057#bib.bib27)\], the second configuration used Codex as the agentic planner and GPT\-4o as its underlying inference model\. Table[1](https://arxiv.org/html/2609.16057#S3.T1)reports bothGPT\-4o \+ OmniHarnessandCodex GPT\-4o \+ OmniHarness\. Unless otherwise specified, all subsequent OmniHarness results use Codex GPT\-4o\. For all reasoning calls, we settemperatureto 0 andtop\_pto 1\. For fair comparison across agentic systems, we fixed the maximum number of retries at 4\[[37](https://arxiv.org/html/2609.16057#bib.bib6)\]\. Within each evaluation setting, workflows constructed by OmniHarness and the compared agents invoked the same underlying generative models\[[65](https://arxiv.org/html/2609.16057#bib.bib3),[18](https://arxiv.org/html/2609.16057#bib.bib5),[37](https://arxiv.org/html/2609.16057#bib.bib6)\]\. Unified generative models and published baselines followed their respective benchmark protocols\.

### A\.4Fine\-Grained Ablation Study

Table[6](https://arxiv.org/html/2609.16057#S4.T6)in the main paper evaluates self\-directed inquiry and online policy updates\.w/o Self\-Directed Inquiryremoves inquiry while retaining downstream updates to the workflow and failure libraries, assessing the benefit of learning before downstream tasks arrive\.w/o Online Policy Updatesretains the policy library learned through inquiry but disables updates to both libraries during downstream execution\. Policy instantiation, adaptation, composition, and feedback\-guided execution remain active\. This variant assesses continual policy refinement through execution feedback\.OmniHarnessenables both mechanisms\.

We conduct six fine\-grained ablation and diagnostic studies\. First, we examine capability novelty𝒩⁡\(τ\)\\mathcal\{N\}\(\\tau\), the competence frontier score𝒞⁡\(τ\)\\mathcal\{C\}\(\\tau\), and the number of inquiry iterations\. Second, we ablate planning, intermediate verification, and localized recovery in feedback\-guided execution\. Third, we analyze symbolic policy learning through workflow library growth, reliability, reuse, composition, capability distribution, and modality expansion\. Fourth, we measure reasoning efficiency through agent\-side token consumption and reasoning time across task difficulty levels\. Fifth, we evaluate different reasoning backbones and planner configurations\. Finally, we replace ComfyBench source images with independently generated images from the same capability space to assess whether the benefits of inquiry depend on the original source pool\.

#### A\.4\.1Fine\-Grained Ablation of Self\-Directed Inquiry

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/4.jpg)Figure 6:Visualization of the workflow library after 50 iterations of self\-directed inquiry\. \(a\) Word cloud summarizing the capabilities represented by stored policies\. \(b\) Distribution of workflow entries across capability categories\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/play_round_ablation_resolve_rates.png)Figure 7:Resolve rates with different numbers of self\-directed inquiry iterations on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\.After selectingτt\\tau\_\{t\}, OmniHarness constructs or adapts a workflow through feedback\-guided execution, including planning, Code\-as\-Policy synthesis, execution, intermediate verification, and localized recovery\. Whenoto\_\{t\}confirms success, OmniHarness abstracts the verified workflowπt\\pi\_\{t\}into a symbolic policy represented by a reusable workflow entryww\. Distillation removes instance\-specific inputs, including the task descriptionqτtq\_\{\\tau\_\{t\}\}and, for I2I tasks, the source imagexτtx\_\{\\tau\_\{t\}\}\. Shared procedures and applicability conditions support reuse across a visual generation task family\. As summarized in Table[10](https://arxiv.org/html/2609.16057#A1.T10), each entry records its modalitymτtm\_\{\\tau\_\{t\}\}, capability categories𝒢τt\\mathcal\{G\}\_\{\\tau\_\{t\}\}, reusable ComfyUI workflow template, name, description, preconditions, expected effects, dependencies, and reliability statistics\.

Before insertion, the distilled entry is compared with workflows inℒt\\mathcal\{L\}\_\{t\}by intended function and executable structure\. If an equivalent workflow exists, the verified execution updates that entry, incrementing its usage countuwu\_\{w\}and success countsws\_\{w\}\. The empirical success ratepw=sw/uwp\_\{w\}=s\_\{w\}/u\_\{w\}, conservative reliability estimater^​\(w\)\\widehat\{r\}\(w\), and reliability tier are then updated\. Otherwise,wwis inserted intoℒt\+1\\mathcal\{L\}\_\{t\+1\}as a newProvisionalworkflow entry\.

AfterN=50N=50inquiry iterations, the workflow libraryℒN\+1\\mathcal\{L\}\_\{N\+1\}contains 43 symbolic policies stored as workflow entries, as visualized in Figure[6](https://arxiv.org/html/2609.16057#A1.F6)\. The curriculum does not enforce a uniform category distribution\. Candidate generation considers the modality and capability coverage inctc\_\{t\}, while task selection maximizes𝒩⁡\(τ\)​𝒞​\(τ\)\\mathcal\{N\}\(\\tau\)\\mathcal\{C\}\(\\tau\)\. Capability novelty favors underexplored context–capability combinations\. The competence frontier score favors intermediate estimated competence, serving as a learnability heuristic\. A category may therefore be revisited when it contains novel task variants near the current competence frontier\.

Localized Image Editing forms the largest category, accounting for 16\.3% of the library\. Reference\-Guided Generation and Style Transfer & Repainting each account for 14\.0%\. Policies in these categories support localized control, content preservation, reference conditioning, and multi\-stage workflow composition\. Photorealistic Generation also accounts for 14\.0%, covering diverse subjects and scenes that require high perceptual fidelity\.

As reported in Table[6](https://arxiv.org/html/2609.16057#S4.T6), removing self\-directed inquiry reduces Creative Total Resolve from 95\.0% to 72\.5%\. The corresponding T2I and I2I Resolve declines are 33\.3 and 20\.0 percentage points, respectively\. These results support acquiring reusable symbolic policies before downstream Creative tasks arrive\. Table[7](https://arxiv.org/html/2609.16057#S4.T7)isolates the two task selection criteria\. All variants retain a 100\.0% Pass rate, with differences appearing in Resolve, which measures whether outputs satisfy all task requirements\. Removing capability novelty lowers Creative Total Resolve to 87\.5%\. T2I Resolve decreases from 83\.3% to 66\.7%, while I2I Resolve decreases from 93\.3% to 86\.7%\. These declines support𝒩⁡\(τ\)\\mathcal\{N\}\(\\tau\)in directing exploration toward underexplored context–capability pairs\. Removing the competence frontier score lowers Creative Total Resolve to 85\.0%\. T2I Resolve remains at 83\.3%, but I2I Resolve decreases to 80\.0%\. This larger I2I decline supports using𝒞⁡\(τ\)\\mathcal\{C\}\(\\tau\)to guide practice with source images according to estimated capability support\. The combined objective𝒩⁡\(τ\)​𝒞​\(τ\)\\mathcal\{N\}\(\\tau\)\\mathcal\{C\}\(\\tau\)achieves the highest Creative Total Resolve, matches the best T2I result, and improves I2I Resolve among the tested variants\. This supports the complementary roles of novelty and estimated learnability in task selection\.

As shown in Figure[7](https://arxiv.org/html/2609.16057#A1.F7), increasing self\-directed inquiry iterations improves task resolution, with the largest gains on Creative tasks\. Without inquiry, OmniHarness achieves Resolve rates of 96\.0%, 81\.7%, and 72\.5% on Vanilla, Complex, and Creative tasks, respectively, yielding 87\.0% overall Resolve\. Ten iterations increase Creative Resolve to 80\.0%, while 20 and 50 iterations further improve it to 87\.5% and 95\.0%\. The corresponding overall Resolve rates are 88\.5%, 91\.0%, and 92\.5%\. Vanilla reaches 97\.0%, while Complex improves modestly to 83\.3% after 50 iterations\. These results suggest that policy acquisition before downstream tasks is particularly beneficial for creative workflow discovery and composition\. Increasing inquiry from 50 to 100 iterations raises Creative Resolve to 97\.5% and overall Resolve to 93\.5%, adding only 1\.0 percentage point overall\. This diminishing performance gain supports using 50 iterations as a practical balance between task resolution and additional practice cost\.

#### A\.4\.2Fine\-Grained Ablation of Feedback\-Guided Execution

As shown in Table[8](https://arxiv.org/html/2609.16057#S4.T8), the full OmniHarness achieves a 100\.0% Pass rate and an 83\.3% Resolve rate on the Complex subset\. Removing planning causes the largest Resolve decline, reducing Complex Total Resolve from 83\.3% to 55\.0%\. T2I and I2I Resolve fall to 40\.0% and 56\.5%, respectively\. These results support the role of the ordered planρt\\rho\_\{t\}in decomposing multi\-step objectives, maintaining dependencies, and guiding workflow composition and synthesis\. Removing intermediate verification causes the largest Pass decline, reducing Complex Total Pass from 100\.0% to 75\.0%, while Resolve falls to 68\.3%\. Without intermediate checks, incorrect outputs fromat,ja\_\{t,j\}can propagate to later steps\. Removing localized recovery produces smaller but consistent declines, with Pass and Resolve decreasing to 93\.3% and 76\.7%, respectively\. This supports targeted repair that preserves verified steps when adapting symbolic policies to downstream tasks\.

#### A\.4\.3Symbolic Policy Learning Analysis

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/5.jpg)Figure 8:Symbolic policy learning during downstream execution on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. \(a\) Growth in total and validated workflow entries in the workflow library, starting after 50 inquiry iterations and continuing through the Vanilla, Complex, and Creative subsets\. \(b\) Changes in the proportions ofValidated,Provisional, andSuspendedworkflows before and after ComfyBench execution\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/6.jpg)Figure 9:Evolution and reuse of symbolic policies during downstream execution on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. \(a\) Expansion of the image\-only workflow library acquired after 50 inquiry iterations to include T2I, I2I, T2V, I2V, and V2V workflows after ComfyBench execution\. \(b\) Distribution of direct reuse, compositional reuse, and no\-prior\-match cases among resolved Vanilla, Complex, and Creative tasks, together with the overall distribution\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/7.png)Figure 10:Capability distribution shift in the image workflow library during downstream execution on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. The bars compare workflow proportions across capability categories after 50 inquiry iterations and after ComfyBench execution\. Percentages are normalized over T2I and I2I workflows only\. The comparison shows how the capabilities represented by stored symbolic policies change during downstream execution\.To examine symbolic policy learning during downstream execution, we keep library updates active while OmniHarness processes ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. Execution starts from the policy library state acquired after 50 inquiry iterations\. We track the workflow library’s size, reliability, modality coverage, reuse patterns, and capability distribution during downstream execution\.

As shown in Figure[8](https://arxiv.org/html/2609.16057#A1.F8), the workflow library grows from 43 entries after inquiry to 48, 62, and 81 entries after the Vanilla, Complex, and Creative subsets, respectively\. The number of Validated workflows follows a different trend\. Vanilla adds only five workflows but increases Validated workflows from 14 to 33\. This is consistent with repeated executions adding validation evidence to existing policies through updates touwu\_\{w\}andsws\_\{w\}\. Complex and Creative add 14 and 19 workflows, while Validated workflows increase by only eight and four\. Library expansion therefore outpaces the accumulation of sufficient validation evidence for newly acquired policies\. After ComfyBench execution, the Validated proportion increases from 32\.6% to 55\.6%, while the Provisional proportion decreases from 67\.4% to 42\.0%\. Two workflows are assigned the Suspended tier\. These updates expand the workflow library and adjust retrieval priorities according to accumulated execution evidence\.

Figure[9](https://arxiv.org/html/2609.16057#A1.F9)shows how symbolic policies are expanded and reused\. The initial library contains 16 T2I and 27 I2I workflows\. After ComfyBench execution, it contains 22 T2I, 36 I2I, 9 T2V, 7 I2V, and 7 V2V workflows\. The 23 newly acquired video workflows account for 28\.4% of the final library\. These additions extend the live workflow library from image generation to video generation during downstream execution\. Among the 185 resolved tasks, 110 use direct reuse and 60 use compositional reuse, yielding a combined reuse rate of 91\.9%\. Reuse patterns also vary with task difficulty\. Direct reuse accounts for 90\.7% of resolved Vanilla tasks, while compositional reuse accounts for 78\.0% of resolved Complex tasks\. For Creative tasks, the no\-prior\-match rate rises to 26\.3%\. These patterns suggest that simpler tasks mainly invoke existing policies, while harder tasks more often require composition or workflow construction when no applicable policy is available\.

Figure[10](https://arxiv.org/html/2609.16057#A1.F10)presents the capability distribution shift within the image workflow library, which grows from 43 to 58 workflows\. Localized Image Editing increases from 7 to 10 workflows, and Image Restoration & Refinement increases from 5 to 8\. Position\-Constrained Generation and Poster & Graphic Design each increase from 1 to 3\. These additions expand the library’s representation of localized editing, restoration, spatial constraints, and layout\-aware generation\. Category proportions can decrease even when their workflow counts remain stable or increase\. Reference\-Guided Generation and Style Transfer & Repainting each grow from 6 to 7 workflows, although their shares decrease from 14\.0% to 12\.1%\. Photorealistic Generation remains at 6 workflows, while its share decreases from 14\.0% to 10\.3%\. The distribution therefore reflects both library growth and changes in relative capability coverage\.

#### A\.4\.4Reasoning Efficiency Analysis

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/8.png)Figure 11:Reasoning efficiency of OmniHarness on the 200 ComfyBench tasks\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\], reported over consecutive groups of 20 tasks\. The bars show the average token consumption per task, while the line shows the average reasoning time per task\. Tasks 1–100, 101–160, and 161–200 correspond to the Vanilla, Complex, and Creative subsets, respectively\. Reasoning time includes agent\-side planning, workflow synthesis, verification, and recovery, but excludes visual generation model execution\.As shown in Figure[11](https://arxiv.org/html/2609.16057#A1.F11), OmniHarness reasoning cost varies with task difficulty and modality\. Across all 200 ComfyBench tasks, it consumes an average of 12\.03K tokens and 52\.9 s of agent\-side reasoning time per task\. For Tasks 1–100 in the Vanilla subset, token consumption ranges from 6\.74K to 9\.08K, while reasoning time ranges from 30\.8 s to 45\.9 s\. The lowest cost occurs in Tasks 1–20, which contain simple T2I requests\. Cost increases for Tasks 21–60, where T2V and I2V tasks require new modality\-specific workflow construction\. It decreases for Tasks 61–100, which mainly contain I2I tasks suitable for reusing symbolic policies learned through self\-directed inquiry\.

Tasks 101–160 in the Complex subset require 14\.36K–16\.28K tokens and 57\.9–66\.8 s per task\. Their dependent operations require an ordered plan, workflow composition, and intermediate verification\. Tasks 121–140 have the highest cost within this subset and include many video conversion and interpolation workflows\. The decrease for Tasks 141–160 coincides with their larger share of I2I tasks, for which symbolic policies learned through self\-directed inquiry are available\.

Tasks 161–200 in the Creative subset incur the highest reasoning cost\. Average token consumption reaches 16\.47K and 18\.63K in the final two groups, while reasoning time increases to 69\.8 s and 83\.5 s\. These tasks involve longer instructions, stricter content constraints, complex text and layout requirements, and identity\-preserving transformations\. The final group is also dominated by video generation and V2V editing, requiring workflow composition, intermediate verification, and localized recovery\. Token consumption and reasoning time follow similar trends across these task groups\.

#### A\.4\.5Backbone Ablation

Table 13:Backbone ablation of OmniHarness on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. The first three rows report representative existing agents, followed by OmniHarness instantiated with different reasoning backbones and planner configurations\. Pass and Resolve are reported as percentages, with higher values indicating better performance\.AgentVanillaComplexCreativeTotalPass↑\\uparrowRes\.↑\\uparrowPass↑\\uparrowRes\.↑\\uparrowPass↑\\uparrowRes\.↑\\uparrowPass↑\\uparrowRes\.↑\\uparrowGPT\-4o \+ ComfyAgent\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]67\.046\.048\.321\.740\.015\.056\.032\.5GPT\-4o \+ ComfyMind\[[18](https://arxiv.org/html/2609.16057#bib.bib5)\]100\.092\.0100\.085\.0100\.057\.5100\.083\.0Gemini\-2\.5\-Flash \+ SymbOmni\[[37](https://arxiv.org/html/2609.16057#bib.bib6)\]100\.095\.0100\.083\.3100\.067\.5100\.086\.0Gemini\-2\.5\-Flash \+ OmniHarness100\.096\.0100\.081\.7100\.087\.5100\.090\.0GPT\-4o \+ OmniHarness100\.095\.0100\.076\.7100\.095\.0100\.089\.5GPT\-5\.5 \+ OmniHarness100\.097\.0100\.081\.7100\.097\.5100\.092\.5Codex GPT\-4o \+ OmniHarness100\.097\.0100\.083\.3100\.095\.0100\.092\.5Codex GPT\-5\.5 \+ OmniHarness100\.098\.0100\.088\.3100\.097\.5100\.095\.0As shown in Table[13](https://arxiv.org/html/2609.16057#A1.T13), same\-backbone comparisons evaluate agent frameworks while holding the underlying reasoning model fixed\. With GPT\-4o, OmniHarness achieves a 100\.0% Total Pass rate and an 89\.5% Total Resolve rate, compared with 56\.0% and 32\.5% for ComfyAgent, and 100\.0% and 83\.0% for ComfyMind\. Its largest advantage over ComfyMind appears on Creative tasks, where Resolve increases from 57\.5% to 95\.0%, although ComfyMind performs better on Complex tasks\. With Gemini\-2\.5\-Flash, OmniHarness improves Total Resolve over SymbOmni from 86\.0% to 90\.0% and Creative Resolve from 67\.5% to 87\.5%, despite a small decrease on Complex tasks from 83\.3% to 81\.7%\. These comparisons show performance differences across frameworks using the same reasoning model\. Across OmniHarness configurations, all variants achieve a 100\.0% Pass rate, while Resolve varies with the model and planner configuration\. Replacing GPT\-4o with GPT\-5\.5 increases Total Resolve from 89\.5% to 92\.5%\. Codex GPT\-4o also reaches 92\.5%, outperforming GPT\-5\.5 on Complex tasks but trailing it on Creative tasks\. Codex GPT\-5\.5 achieves the highest overall result of 95\.0% among the tested configurations\. These results support using OmniHarness across reasoning models and planner configurations\. Its symbolic policies capture shared procedures for visual generation task families, while feedback\-guided execution adapts them to new tasks and supplies feedback for continual policy learning\.

#### A\.4\.6Source Image Pool Isolation Study

Table 14:Source image pool isolation study on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\.𝒳ComfyBench\\mathcal\{X\}\_\{\\mathrm\{ComfyBench\}\}denotes the original ComfyBench source image pool, while𝒳Gen\\mathcal\{X\}\_\{\\mathrm\{Gen\}\}replaces it with independently generated images from the same generative capability space\.To assess whether the benefits of inquiry depend on ComfyBench source images, we replace the original pool with 33 independently generated images from the same generative capability space\. As shown in Table[14](https://arxiv.org/html/2609.16057#A1.T14), OmniHarness with𝒳Gen\\mathcal\{X\}\_\{\\mathrm\{Gen\}\}achieves a Total Resolve rate of 91\.5%, compared with 92\.5% using𝒳ComfyBench\\mathcal\{X\}\_\{\\mathrm\{ComfyBench\}\}\. Vanilla and Complex results remain unchanged, while Creative Resolve decreases\. These results suggest that inquiry can acquire useful symbolic policies from independently generated images, with limited dependence on the original ComfyBench source pool in this evaluation\.

### A\.5Comprehensive Benchmark Evaluation

We evaluate OmniHarness on six complementary benchmarks covering diverse visual generation task families\. ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]assesses autonomous ComfyUI workflow construction and execution across Vanilla, Complex, and Creative generation and editing tasks\. GenEval\[[17](https://arxiv.org/html/2609.16057#bib.bib45)\]measures compositional text\-to\-image fidelity through object generation, object co\-occurrence, counting, color, relative position, and attribute binding\. GenEval2\[[25](https://arxiv.org/html/2609.16057#bib.bib53)\]provides a finer\-grained assessment of object generation, attribute rendering, counting, spatial relations, and transitive verb relations\. Reason\-Edit\[[23](https://arxiv.org/html/2609.16057#bib.bib63)\]evaluates explicit instruction understanding and commonsense\-driven image editing, emphasizing target localization and preservation of unrelated content\. WISE\[[41](https://arxiv.org/html/2609.16057#bib.bib57)\]examines knowledge\-guided synthesis across cultural, temporal, spatial, biological, physical, and chemical domains\. KRIS\-Bench\[[58](https://arxiv.org/html/2609.16057#bib.bib69)\]evaluates factual, conceptual, and procedural knowledge in generation and editing\. Together, these benchmarks assess performance across workflow construction, compositional generation, reasoning\-guided editing, and knowledge\-grounded visual synthesis\.

#### A\.5\.1Evaluation on ComfyBench

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/10.jpg)Figure 12:Qualitative comparison on representative I2I tasks from ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. The examples illustrate target localization, instruction following, and preservation of unrelated scene content\.Figures[12](https://arxiv.org/html/2609.16057#A1.F12)and[3](https://arxiv.org/html/2609.16057#S3.F3)present qualitative comparisons on representative ComfyBench tasks\. Figure[12](https://arxiv.org/html/2609.16057#A1.F12)compares localized I2I editing using the same input image\. In these examples, OmniHarness removes the plate, fork, or glass as instructed and replaces the fork with a spoon while preserving unrelated objects and the scene layout\. Some compared outputs retain the target object, modify non\-target content, or change the overall appearance\. Figure[3](https://arxiv.org/html/2609.16057#S3.F3)illustrates multi\-stage and appearance\-sensitive tasks\. OmniHarness reconstructs a scribble as a realistic red flower and transfers its visual style to a new flower\-field image\. It also restores and colorizes an old photograph while preserving identity and clothing details\. In facial appearance editing, it changes the requested attributes while retaining facial structure, pose, and clothing\.

As shown in Table[1](https://arxiv.org/html/2609.16057#S3.T1), both OmniHarness variants achieve a 100\.0% Pass rate across all task categories\. Codex GPT\-4o \+ OmniHarness obtains a Total Resolve rate of 92\.5%, exceeding SymbOmni and ComfyMind by 6\.5 and 9\.5 percentage points, respectively\. Its largest advantage appears on Creative tasks, where Resolve reaches 95\.0%, compared with 67\.5% for SymbOmni and 57\.5% for ComfyMind\. The ablation results in Tables[6](https://arxiv.org/html/2609.16057#S4.T6)and[7](https://arxiv.org/html/2609.16057#S4.T7)further support the contribution of self\-directed inquiry and its task selection criteria to Creative performance\. Before downstream objectives are specified, OmniHarness selects practice tasks using capability novelty𝒩⁡\(τ\)\\mathcal\{N\}\(\\tau\)and the competence frontier score𝒞⁡\(τ\)\\mathcal\{C\}\(\\tau\)\. Their product favors underexplored context–capability pairs near the estimated competence frontier\. Verified executions are abstracted into symbolic policies that capture shared procedures and applicability conditions for visual generation task families\. The harness instantiates, adapts, and composes these policies for downstream requirements\.

#### A\.5\.2Evaluation on GenEval

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/geneval.jpg)Figure 13:Qualitative comparison on representative GenEval\[[17](https://arxiv.org/html/2609.16057#bib.bib45)\]tasks covering single\-object generation, two\-object co\-occurrence, counting, color, spatial relations, and attribute binding\.As shown in Table[2](https://arxiv.org/html/2609.16057#S4.T2)and Figure[13](https://arxiv.org/html/2609.16057#A1.F13), OmniHarness achieves an overall GenEval score of 0\.997, exceeding SymbOmni, ComfyMind, and GPT\-Image\-1 by 0\.017, 0\.097, and 0\.157, respectively\. It reaches 1\.00 on single\-object generation, two\-object co\-occurrence, counting, color, and relative position, and 0\.98 on attribute binding\. Compared with SymbOmni, Position improves from 0\.97 to 1\.00 and Attribute Binding from 0\.95 to 0\.98\. GPT\-Image\-1 obtains 0\.75 and 0\.61 on these categories\. The qualitative examples illustrate these compositional requirements\. OmniHarness generates a clear parking meter, preserves bottle–refrigerator co\-occurrence, produces exactly two bears, renders the requested red zebra, places the bird below the skateboard, and binds the purple and brown attributes to the elephant and sports ball\. Some compared outputs violate counts, spatial relations, or attribute assignments\. These results indicate high compositional fidelity across the evaluated dimensions\. OmniHarness uses feedback\-guided execution to adapt and compose symbolic policies for task requirements, with intermediate verification guiding workflow refinement\.

#### A\.5\.3Evaluation on GenEval2

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/geneval2.jpg)Figure 14:Qualitative comparison on representative GenEval2\[[25](https://arxiv.org/html/2609.16057#bib.bib53)\]tasks covering object generation, attribute rendering, counting, spatial relations, and transitive verb relations\.As shown in Table[3](https://arxiv.org/html/2609.16057#S4.T3)and Figure[14](https://arxiv.org/html/2609.16057#A1.F14), OmniHarness achieves the best results on four of the five GenEval2 skills\. It obtains 94\.0 on Attribute and Count, 76\.9 on Position, and 89\.0 on Verb, exceeding the strongest competing results by 2\.6, 19\.2, 6\.7, and 2\.3 points, respectively\. Compared with SymbOmni, the gains are 10\.4 points on Attribute, 19\.2 on Count, 8\.1 on Position, and 24\.5 on Verb\. Its Object score of 95\.0 matches SymbOmni but remains below Qwen\-Image and Gemini 2\.5 Flash Image\. The strongest advantages therefore appear on tasks with interacting compositional constraints\. In the presented examples, OmniHarness generates six cars with one kangaroo, a glass lion with one bear and four kangaroos, five checkered umbrellas with a black candle, a lion jumping over a sparkling dog, and five penguins to the left of a striped violin\. Some competing outputs violate counts, assign attributes to the wrong objects, or miss spatial and action relations\. These results support strong performance on counting, attributes, spatial relations, and transitive verb relations, while isolated object generation remains less competitive\.

#### A\.5\.4Evaluation on Reason\-Edit

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/reason_edit_gpt_score.png)Figure 15:GPT\-score comparison on Reason\-Edit\[[23](https://arxiv.org/html/2609.16057#bib.bib63)\]across seven reasoning\-intensive image\-editing categories\. Higher scores indicate better performance\.As shown in Table[5](https://arxiv.org/html/2609.16057#S4.T5), Figure[15](https://arxiv.org/html/2609.16057#A1.F15), and Figure[5](https://arxiv.org/html/2609.16057#S4.F5), OmniHarness performs well on instruction understanding and reasoning\-guided editing\. In the Understanding Scenarios, it leads all four metrics, with PSNR of 23\.894 dB, SSIM of 0\.856, LPIPS of 0\.053, and CLIP Score of 24\.554\. Compared with the strongest baseline for each metric, PSNR improves by 0\.298 dB, SSIM by 0\.105, and CLIP Score by 0\.820, while LPIPS decreases by 0\.014\. In the Reasoning Scenarios, OmniHarness achieves the highest SSIM of 0\.796 and CLIP Score of 21\.318\. Its LPIPS is 0\.052, compared with the best result of 0\.049, and its PSNR remains below SmartEdit\-13B and InsightEdit\. The GPT\-score comparison provides another assessment\. OmniHarness ranks first on Left\-Right, Mirror, Color, Multiple\-Objects, and Reasoning, reaches 10\.00 on Add\-supp, and ranks second to Nano Banana on Relative\-Size\. The Understanding examples illustrate target identification through spatial position, color, mirror relations, and relative size\. OmniHarness replaces or adds the specified object while preserving other instances and scene structure\. In the Reasoning examples, it identifies implicit targets from commonsense descriptions, including a dog bowl, a fruit\-cutting knife, vitamin\-rich food, an animal lying on grass, a game ball, and a warning sign\. It then performs the requested removal or replacement with limited changes to unrelated regions\. These results support the effectiveness of OmniHarness on both explicit and reasoning\-intensive editing instructions\.

#### A\.5\.5Evaluation on WISE

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/WISE.jpg)Figure 16:Qualitative comparison on representative WISE\[[41](https://arxiv.org/html/2609.16057#bib.bib57)\]tasks spanning cultural, temporal, spatial, biological, physical, and chemical knowledge\.As shown in Table[4](https://arxiv.org/html/2609.16057#S4.T4)and Figure[16](https://arxiv.org/html/2609.16057#A1.F16), OmniHarness achieves the highest overall WiScore of 0\.86, exceeding SymbOmni and GPT\-Image\-1 by 0\.06 and ComfyMind by 0\.10\. It ranks first in Time, Biology, Physics, and Chemistry, with scores of 0\.85, 0\.84, 0\.82, and 0\.86, respectively\. Compared with the strongest competing result in each category, the gains are 0\.14, 0\.01, 0\.03, and 0\.08, respectively\. Its Cultural and Space scores reach 0\.88 and 0\.86, remaining 0\.02 and 0\.03 below the corresponding best results\. The qualitative examples illustrate the range of knowledge requirements\. OmniHarness generates Russian nesting dolls as the requested cultural craft, depicts the Sydney Opera House at night, and preserves the depth relation between a nearby coin and a distant cityscape\. It also depicts the formation of migrating geese, condensation on a cold can in humid air, and the precipitate formed by mixing copper sulfate and sodium hydroxide\. Some compared outputs capture the main objects but miss the implied temporal condition, spatial relation, physical state, or scientific consequence\. These results support the effectiveness of OmniHarness across the evaluated cultural, spatiotemporal, and scientific generation tasks\.

#### A\.5\.6Evaluation on KRIS\-Bench

Table 15:Quantitative comparison on KRIS\-Bench\[[58](https://arxiv.org/html/2609.16057#bib.bib69)\]\. Scores evaluate factual, conceptual, and procedural knowledge, together with overall performance\. Higher values indicate better performance\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/krisbench.jpg)Figure 17:Qualitative results of OmniHarness on representative KRIS\-Bench tasks\[[58](https://arxiv.org/html/2609.16057#bib.bib69)\], covering factual, conceptual, and procedural knowledge through knowledge\-guided generation, editing, visual reasoning, and multi\-image composition\.As shown in Table[15](https://arxiv.org/html/2609.16057#A1.T15)and Figure[17](https://arxiv.org/html/2609.16057#A1.F17), OmniHarness achieves an overall score of 77\.33 on KRIS\-Bench\[[58](https://arxiv.org/html/2609.16057#bib.bib69)\], exceeding SymbOmni by 5\.15 points and BAGEL\-Think, the strongest open\-source baseline overall, by 23\.97 points\. It trails GPT\-Image\-1 by 2\.76 points\. OmniHarness achieves the best Conceptual score of 81\.89, exceeding GPT\-Image\-1 and SymbOmni by 0\.52 and 9\.61 points, respectively\. Its Factual and Procedural scores reach 74\.81 and 73\.22, improving over SymbOmni by 1\.48 and 2\.93 points\. The qualitative examples illustrate diverse knowledge requirements\. OmniHarness completes numerical patterns, applies scientific, geographic, and biological knowledge, reasons about physical processes, performs landmark and color transformations, composes reference images, annotates geometric and medical content, controls object counts, completes missing structures, and follows poster\-design constraints\. These tasks require knowledge to be translated into visual operations under specific constraints\. The harness instantiates, adapts, and composes symbolic policies to meet these requirements\. Intermediate verification guides workflow refinement and localized recovery during execution\.

### A\.6Detailed Plug\-and\-Play Analysis

Table 16:Quantitative plug\-and\-play evaluation on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. The first three rows report the original host agents, while the last three augment them with the same frozen policy snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}learned through self\-directed inquiry, without model fine\-tuning\.To evaluate policy portability, we keep the snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}learned through self\-directed inquiry frozen and map its entries to each host agent’s native knowledge interface\. The snapshot contains reusable workflow templates, failure evidence, and corrective strategies from the workflow and failure libraries\. Host model weights and original control flow remain unchanged\. Host agents retrieve, instantiate, and adapt the exported policies through their existing execution procedures\. For ComfyAgent\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\], each workflow template is converted into its native Python\-like representation and inserted into theReferencememory with its name, description, preconditions, expected effects, and dependencies\. The RetrieveAgent selects relevant entries, while the CombineAgent and AdaptAgent compose or modify the retrieved workflows\. The RefineAgent checks the adapted workflow before it enters theWorkspace\. Failure evidence and corrective strategies provide correction references for the PlanAgent and RefineAgent\. For ComfyMind\[[18](https://arxiv.org/html/2609.16057#bib.bib5)\], each compatible workflow entry is wrapped as a Semantic Workflow Interface module\. Its description, preconditions, and expected effects define the semantic function interface, while its ComfyUI graph template provides the canonical JSON implementation\. The Planning Agent selects relevant modules, and the Execution Agent supplies task\-specific parameters before execution\. Corrective strategies are appended to the local\-feedback context as correction guidance\. For SymbOmni\[[37](https://arxiv.org/html/2609.16057#bib.bib6)\], each workflow entry is converted into a Symbolic Concept\. Its description and preconditions define the concept semantics, its reusable procedure defines the Symbolic Workflow Instruction, its dependencies and parameters define the parameter configuration, and its reliability statistics define the concept score\. Failure library entries become negative or refinement concepts\. The resulting concepts are inserted into the Symbolic Concept Box and accessed through the original retrieval and transduction process\. The exported snapshot receives no updates during evaluation\.

As shown in Table[16](https://arxiv.org/html/2609.16057#A1.T16), the same𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}improves task resolution in all three host agents without model fine\-tuning\. For ComfyAgent, Total Pass increases from 56\.0% to 79\.0%, while Total Resolve increases from 32\.5% to 57\.0%\. Creative Resolve rises from 15\.0% to 42\.5%, indicating gains in workflow executability and task completion\. ComfyMind and SymbOmni retain their 100\.0% Pass rates, with improvements appearing in Resolve\. ComfyMind increases Total Resolve from 83\.0% to 88\.0% and Creative Resolve from 57\.5% to 75\.0%\. SymbOmni increases Total Resolve from 86\.0% to 89\.0% and Creative Resolve from 67\.5% to 77\.5%\. The larger Creative gains suggest that the exported symbolic policies are particularly useful for these tasks\. Improvements through code references, semantic workflow modules, and symbolic concepts support plug\-and\-play reuse across the three evaluated visual agent frameworks\.

### A\.7Cross\-Modal Generalization Evaluation

Table 17:Quantitative cross\-modal generalization results on the video\-generation subset of ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. T2V, I2V, and V2V denote text\-to\-video, image\-to\-video, and video\-to\-video tasks, respectively\. GPT\-4o \+ OmniHarness uses the frozen policy snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}, learned through self\-directed inquiry in an image\-only capability space without exposure to video tasks during inquiry\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/2v.jpg)Figure 18:Qualitative cross\-modal generalization results of OmniHarness on representative video\-generation tasks from ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. Sampled frames illustrate generated content and source\-content preservation in T2V, I2V, and V2V tasks using the frozen policy snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}learned through image\-only inquiry\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/cross_modal_image_workflow_reuse_revised.png)Figure 19:Reuse of symbolic policies learned through image\-only inquiry during cross\-modal generalization on ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. The bars report invocation counts of workflow entries from the frozen image\-only workflow library for T2V, I2V, and V2V tasks\.As shown in Table[17](https://arxiv.org/html/2609.16057#A1.T17)and Figures[18](https://arxiv.org/html/2609.16057#A1.F18)and[19](https://arxiv.org/html/2609.16057#A1.F19), GPT\-4o \+ OmniHarness applies the frozen policy snapshot𝒦inquiry\\mathcal\{K\}\_\{\\mathrm\{inquiry\}\}, learned through image\-only inquiry, to video generation tasks not encountered during self\-directed inquiry\. The snapshot remains unchanged during evaluation\. OmniHarness achieves a 100\.0% Pass rate across T2V, I2V, and V2V, with Resolve rates of 84\.2%, 92\.0%, and 80\.0%, respectively\. Overall Resolve reaches 85\.9%, exceeding SymbOmni by 5\.1 percentage points\. The gains are 2\.6, 4\.0, and 13\.3 percentage points on T2V, I2V, and V2V, with the largest improvement on V2V tasks combining temporal processing and content\-preserving editing\. The sampled frames illustrate T2V generation, source\-conditioned I2V generation, and motion\-transfer and material\-editing tasks in V2V\. Invocation statistics document the reuse of image\-domain symbolic policies within video workflows\. Across all video tasks, OmniHarness invokes image workflows 90 times, including 48 calls for T2V, 27 for I2V, and 15 for V2V\. Photorealistic Generation and Reference\-Guided Generation contribute 33 and 28 calls, while Localized Image Editing and Image Super\-Resolution contribute 10 and 7 calls\. These policies support visual content construction, reference preservation, local editing, and quality enhancement, while video\-specific components handle temporal operations\. The results support reuse beyond image generation through task\-specific adaptation and composition of symbolic policies with video\-specific components, while the exported snapshot remains fixed\.

### A\.8User Study

![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/user_study.png)Figure 20:Web interface used in the 40\-task user study\. Participants assess an anonymized generation result against the task instruction and, when required, the input image\. Ratings cover task description alignment, input\-structure fidelity, photorealism, visual quality, and overall satisfaction on a five\-point Likert scale\.![Refer to caption](https://arxiv.org/html/2609.16057v1/Figure/user_study_grouped_bar_chart.png)Figure 21:User\-study comparison across five evaluation criteria\. Mean ratings over 40 ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]tasks are reported on a five\-point Likert scale, with higher scores indicating better performance\.We conduct a blind user study on 40 tasks selected from ComfyBench\[[65](https://arxiv.org/html/2609.16057#bib.bib3)\]\. A total of 36 participants evaluate anonymized outputs from OmniHarness, SymbOmni, BAGEL, and GPT\-Image\-1\. As shown in Figure[20](https://arxiv.org/html/2609.16057#A1.F20), each questionnaire presents the task instruction, the input image when required, and one anonymized generation result\. Participants rate each result on a five\-point Likert scale from 1 \(Strongly disagree\) to 5 \(Strongly agree\) across Task Description Alignment, Input\-Structure Fidelity, Photorealism, Visual Quality, and Overall Satisfaction\. Figure[21](https://arxiv.org/html/2609.16057#A1.F21)reports mean ratings over tasks and participants\. OmniHarness achieves the highest ratings for Task Description Alignment, Input\-Structure Fidelity, and Overall Satisfaction, reaching 4\.72, 4\.63, and 4\.62, respectively\. GPT\-Image\-1 receives higher Photorealism and Visual Quality ratings of 4\.68 and 4\.64, compared with 4\.50 and 4\.56 for OmniHarness\. Averaged across all five criteria, OmniHarness obtains the highest score of 4\.61, followed by GPT\-Image\-1, SymbOmni, and BAGEL at 4\.43, 4\.35, and 3\.89, respectively\. These results indicate favorable user assessments of OmniHarness, particularly for instruction alignment, input\-structure preservation, and overall satisfaction\.

### A\.9Limitations

OmniHarness combines symbolic policy learning, feedback\-guided execution, and self\-directed inquiry while keeping model parameters fixed\. Verified executions yield symbolic policies for visual generation task families, which are adapted and refined during downstream execution\. However, both self\-directed inquiry and downstream execution incur additional computation\. Self\-directed inquiry requires repeated task generation, workflow construction, execution, verification, and library consolidation\. Downstream tasks may trigger multiple planning, verification, and recovery cycles as the harness adapts and composes policies for new requirements\. These operations increase agent\-side token consumption and reasoning time, especially for complex tasks\. Future work should explore lighter execution frameworks, selective verification, adaptive recovery, efficient policy retrieval and composition, and reasoning budgets that adapt to task difficulty\.

### A\.10System Prompts

The following tables reproduce the two developer instructions and eight task\-input templates used by the implementation\. Braced Python expressions denote values inserted at runtime rather than literal prompt text\. Candidate validation and scoring, context retrieval, structural plan and workflow checks, compilation, execution, evidence aggregation, symbolic policy input abstraction, library updates, consolidation, and snapshot export are implemented programmatically without separate model prompts\. Candidate generation reuses a proposer thread across inquiry iterations\. Execution roles share a task\-scoped thread, while localized repair uses a fresh thread with the same generation\-team developer instructions\.

Table 18:Shared developer instructions for self\-directed inquiry\.Self\-Directed Inquiry Task ProposerYou are the Self\-Directed Inquiry Task Proposer inside OmniHarness\. Generate practice tasks before downstream objectives are specified\. Each task must use the supplied capability space, be feasible with current resources, differ from recent tasks, and be independently verifiable\. Use failure evidence and corrective strategies to avoid known antipatterns\. For I2I, use exactly one supplied image ID and state preservation constraints\. List applicable\_workflow\_ids only when the workflow’s preconditions hold for the proposed task and source image\. Exclude suspended workflows; return an empty list when none apply\. Use only the supplied context and library\. Do not inspect downstream task instructions, reference workflows, target outputs, or benchmark annotations\. Return only the requested JSON object\.

Table 19:Task\-input template for candidate task generation\.Candidate Task GenerationGenerate exactly\{candidate\_count\}candidate tasks\. Return \{"candidates": \[CandidateTask, …\]\}\. Each CandidateTask contains task\_id, description, modality \(T2I or I2I\), capability\_categories \(one to three exact names\), source\_image\_id \(null for T2I\), generation\_constraints, preservation\_constraints, success\_criteria, required\_resources, task\_signature, proposal\_rationale, safety\_flags \(empty for acceptable tasks\), and applicable\_workflow\_ids\.State:\{json\.dumps\(compact, ensure\_ascii=False, indent=2\)\}

Table 20:Shared developer instructions for feedback\-guided execution\.OmniHarness Generation TeamYou are the OmniHarness Generation Team\. Operate through explicit roles: Planner, Plan Verifier, Workflow Writer, Executor observer, Step Verifier, Goal Verifier, and Failure Diagnoser\. Retrieved workflows are symbolic policy templates for task families; retrieved failures contain evidence and corrective strategies\. Bind the current task’s inputs and adapt or compose templates before execution\. Respect preconditions, dependencies and component\-only retrieval roles\. Image components require video\-specific processing for video outputs\. Return only the requested JSON\. Never edit files, execute ComfyUI, change retry budgets, update memory, or declare technical success\. Code\-as\-Policy is a safe DSL with one assignment and one allow\-listed ComfyUI node call per line, keyword arguments only, literals or earlier variables as inputs, and no imports, attributes, control flow, nested calls, or arbitrary Python\.

Table 21:Task\-input template for planning\.PlannerCreate a dependency\-consistent ordered plan with explicit verification for every step\. For each selected workflow, check its input roles, preconditions, and dependencies against the request and planned intermediate inputs\. In task\_analysis\.policy\_applicability return \{workflow\_id:\{satisfied:true,evidence:\[specific condition checks\]\}\}\. Select a policy only when its conditions can be met by the plan\. Choose reuse, adapt, compose, or build\. Return \{plan\_id, task\_analysis, strategy, selected\_workflow\_ids, steps:\[\{step\_id, objective, inputs, outputs, required\_nodes, verification:\{type,criteria\}, expected\_output\_nodes, fallback\}\], rationale\}\.Request:\{\_json\(request\)\}Context:\{\_json\(context\.compact\(include\_prompts=False\)\)\}

Table 22:Conditional task\-input template for plan repair\.Plan RepairRepair this plan and return the same schema\.Request:\{\_json\(request\)\}Plan:\{\_json\(plan\)\}Issues:\{\_json\(issues\)\}Context:\{\_json\(context\.compact\(include\_prompts=False\)\)\}

Table 23:Task\-input template for workflow writing and policy adaptation\.Workflow WriterConvert the verified plan to Code\-as\-Policy\. Use $SOURCE\_MEDIA\_0 placeholders for media inputs; $SOURCE\_IMAGE\_0 remains accepted for compatibility\. Retrieved graphs are templates: replace every $POLICY\_\* placeholder with task\-specific text, numeric parameters, output names, or $SOURCE\_MEDIA\_n bindings\. Check input\-role preconditions and compose component\_only image policies with video\-specific nodes\. Set \_meta\_title on every Code\-as\-Policy node to its owning plan step\_id\. Preserve existing assignment variable names during localized repairs\. Each step must emit an inspectable intermediate output before dependent steps execute; expected\_output\_nodes should identify actual numeric node IDs of those output nodes, matching their one\-based Code\-as\-Policy statement positions\. A step title refers only to that step’s output nodes; it does not require observing every internal node\. Every node must contribute to at least one declared step output\. If no applicable policy is available, construct a workflow from node knowledge\. Describe the reusable task\-family procedure, not the particular objects, prompt, or source filename of this execution, in workflow\_name, description, preconditions and expected\_effects\. Return \{ir\_code, workflow\_name, description, preconditions, expected\_effects, dependencies:\{Nodes,Models,Parameters\}\}\.Request:\{\_json\(request\)\}Plan:\{\_json\(plan\)\}Context:\{\_json\(context\.compact\(include\_prompts=True\)\)\}

Table 24:Conditional task\-input template for step verification\.Step VerifierInspect the supplied step output artifacts using image or video tools\. Verify each listed criterion against these artifacts and the request\. Do not infer an intermediate effect from the final result\. Return \{criteria:\[\{criterion,status:pass\|fail\|uncertain,evidence:\[string\]\}\]\}\. Copy each criterion exactly and use uncertain if its evidence cannot be inspected\.Request:\{\_json\(request\)\}Step:\{\_json\(step\)\}Step outputs:\{\_json\(\[item\.path for item in result\.outputs\]\)\}

Table 25:Conditional task\-input template for goal verification\.Goal VerifierInspect every output path with the appropriate image or video tools\. Evaluate every success and preservation criterion\. Use uncertain when evidence is insufficient\. Return \{status:pass\|fail\|uncertain, verifier, scores, evidence, failed\_criteria\}\.Request:\{\_json\(request\)\}Plan:\{\_json\(plan\)\}Outputs:\{\_json\(\[item\.path for item in result\.outputs\]\)\}

Table 26:Conditional task\-input template for failure diagnosis\.Failure DiagnoserIdentify the smallest failed stage\. Return \{failure\_class:planning\_failure\|compilation\_failure\|infrastructure\_failure\|execution\_failure\|verifier\_failure\|semantic\_failure, failed\_stage, root\_cause, workflow\_antipattern, remedy, applicable\_scope, retry\_scope\}\.Request:\{\_json\(request\)\}Plan:\{\_json\(plan\)\}Draft:\{\_json\(draft\)\}Error:\{\_json\(error\)\}Verification:\{\_json\(verification\)\}

Table 27:Conditional task\-input template for localized repair\.Failure SubagentAct as an isolated Failure Subagent\. Repair only the diagnosed step and preserve unrelated workflow logic\. Preserve the assignment variable names and node inputs outside that step and its dependent downstream nodes\. Return \{failed\_step\_id, root\_cause, subworkflow\_ir, revised\_ir\_code, reusable\_name, rationale\}\. The full revised\_ir\_code must integrate the localized component\.Request:\{\_json\(request\)\}Plan:\{\_json\(plan\)\}Draft:\{\_json\(draft\)\}Diagnosis:\{\_json\(diagnosis\)\}Recovery context:\{\_json\(context\.compact\(include\_prompts=True\)\)\}