LongHarness：通过有状态证据管理推进亿级 Token 语料库上的智能体推理

# Advancing Agentic Reasoning over100M-Token Corpora via Stateful Evidence Management
Source: [https://arxiv.org/html/2610.13075](https://arxiv.org/html/2610.13075)
\\addtolist

\[1\]Fudan University\\affiliationlist\\addtolist\[2\]Shanghai Artificial Intelligence Laboratory\\affiliationlist\\addtolist\[3\]Peking University\\affiliationlist\\addtolist\[4\]Shanghai Jiaotong University\\affiliationlist\\addtolist\[5\]University of Science and Technology of China\\affiliationlist\\addtolist\[6\]Harbin Institute of Technology\\affiliationlist\\contribution\[\*\]Corresponding author\\metadata\[Contact\]

###### Abstract

Corpus\-level long\-context reasoning requires agents to acquire and integrate evidence across document collections that exceed their context windows\. Existing agent harnesses enable incremental access to these corpora, but acquired evidence and evolving conclusions often remain entangled in a growing interaction history\. This implicit state management can lead to two coupled failure modes: unbounded exploration, where retrieval expands the interaction history without resolving decision\-relevant information gaps, and evidence\-state drift, where previously acquired evidence or subsequent corrections fail to inform the current conclusion\. We presentLongHarness, a model\-agnostic harness that makes the interaction between acquired evidence and evolving conclusions explicit\. Its core mechanism,*Evidence–Conclusion Coevolution*, uses unresolved conclusions to guide exploration toward missing information, while newly acquired evidence strengthens or revises the conclusions that directed the search\.LongHarnesscouples this process with*Evidence Lifecycle Management*, which keeps accumulated evidence current and traceable throughout the interaction\. Together, these mechanisms focus retrieval on remaining information needs and keep conclusions aligned with current evidence\. We evaluateLongHarnesson OfficeQA Pro V1 and V2, AA\-LCR, and LongBench v2 using GLM\-5\.3\-Flash and Qwen3\.8\-Flash\. Compared with direct reasoning using the same backbone,LongHarnessimproves GLM\-5\.3\-Flash exact match on OfficeQA Pro from 15\.8% to 60\.2% on V1 and from 4\.4% to 61\.1% on V2\. Against state\-of\-the\-art agent harnesses, it surpasses Claude Code by 12\.0 percentage points on V1 and Codex by 18\.9 points on V2\. With Qwen3\.8\-Flash,LongHarnessreaches 61\.7% on V1, outperforming Claude Code by 8\.3 points and Codex by 12\.0 points\.

![Refer to caption](https://arxiv.org/html/2610.13075v1/figure1.png)Figure 1:Evidence lifecycle management for corpus\-level long\-context reasoning\. The left panel illustrates how an agent loses an earlier constraint and continues searching, whileLongHarnessretains the evidence, uses an unresolved conclusion to guide retrieval, and revises its decision when new evidence arrives\. The right panel compares exact accuracy \(%\) on OfficeQA Pro V1 and V2 for direct models and agent harnesses using GLM\-5\.3\-Flash\.## 1Introduction

Real\-world document questions often require evidence scattered across large, evolving document collections\. Corpus\-level long\-context reasoning requires agents to locate and integrate this evidence even when the collection exceeds their context windows\. Longer windows expand the material available within a single invocation\([Gemini Team, 2024](https://arxiv.org/html/2610.13075#bib.bib10);[Yang et al\., 2025b](https://arxiv.org/html/2610.13075#bib.bib39);[Shang et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib31)\), but models remain sensitive to evidence position and distractors even when the relevant content fits\([Liu et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib23);[Hsieh et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib13);[Ling et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib22);[Zhuang et al\., 2025a](https://arxiv.org/html/2610.13075#bib.bib48)\)\. Agents must therefore search selectively, read relevant passages, and accumulate evidence across multiple interactions\. As these interactions unfold, the challenge extends from finding evidence to keeping it usable, revising it when needed, and carrying it into the final decision\.

We view this process at three levels\.*Evidence retrieval*\(Level 1\) locates relevant passages while filtering distractors\.*Evidence editing*\(Level 2\) selects, reconciles, and revises retrieved content\.*Evidence transformation*\(Level 3\) turns the resulting evidence into an answer\. Training methods supervise retrieval trajectories and memory edits\([Lin et al\., 2026b](https://arxiv.org/html/2610.13075#bib.bib21);[Chen et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib5);[Wang et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib36);[Gao and Marttinen, 2026](https://arxiv.org/html/2610.13075#bib.bib9)\), while agent harnesses enable adaptive document access and persistent storage\([Cao et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib4);[Chen et al\., 2023](https://arxiv.org/html/2610.13075#bib.bib6);[Lee et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib18);[Zhao et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib45);[Packer et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib27);[Karten et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib15);[Yu et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib41)\)\. These advances strengthen evidence acquisition and maintenance, yet the transition from accumulated evidence to an evolving conclusion often remains implicit\. The conclusions that determine what to search for next and what answer to produce remain embedded in session history, free\-form notes, or transient reasoning\.

At corpus scale, this implicit transition can lead to two coupled failure modes\. \(i\)*Unbounded exploration*occurs when retrieval continues without resolving decision\-relevant information gaps, potentially exhausting the interaction budget\. \(ii\)*Evidence\-state drift*occurs when previously acquired evidence or later corrections become disconnected from subsequent reasoning, leaving the final decision based on stale or incomplete evidence\. These failures can reinforce each other as drift obscures which information needs have already been resolved, prompting further retrieval that expands the history and makes relevant evidence harder to recover\.

We introduceLongHarness, a model\-agnostic harness that maintains source\-linked evidence and evolving conclusions as coupled persistent states\. Unresolved conclusion gaps guide subsequent search and reading, while accepted evidence informs revisions to the conclusions and their remaining gaps\. Its core mechanism,*Evidence–Conclusion Coevolution*, uses unresolved conclusions to guide exploration toward missing information, while newly acquired evidence strengthens or revises the conclusions that directed the search\. By making the remaining information needs explicit, this feedback directs retrieval toward gaps that matter to the decision and helps curb unbounded exploration\.LongHarnesscouples this loop with*Evidence Lifecycle Management*, which keeps accumulated evidence current and traceable throughout the interaction\. TheEvidenceBankretains source\-linked statements and tracks which remain active throughout the evidence lifecycle, helping reduce evidence\-state drift by keeping the basis for subsequent reasoning aligned with the latest findings\.

We evaluateLongHarnesson OfficeQA Pro V1 and V2, AA\-LCR, and the LongBench v2 Long subset using GLM\-5\.3\-Flash and Qwen3\.8\-Flash\. OfficeQA Pro requires reasoning over roughly 100M tokens in V1, while V2 doubles the corpus and introduces more demanding compositional questions\. Compared with direct reasoning using the same backbone,LongHarnessraises GLM\-5\.3\-Flash exact match from 15\.8% to 60\.2% on V1 and from 4\.4% to 61\.1% on V2\. Against state\-of\-the\-art agent harnesses, it surpasses Claude Code by 12\.0 percentage points on V1 and Codex by 18\.9 points on V2\. With Qwen3\.8\-Flash, it reaches 61\.7% on V1, outperforming Claude Code by 8\.3 points and Codex by 12\.0 points\. Trajectory analysis with GLM\-5\.3\-Flash further connects these gains to evidence management, with V2 success rates exceeding the stronger of Claude Code and Codex at each stage by 11\.1 points in retrieval, 2\.7 in evidence editing, and 9\.3 in transformation\. Our contributions are summarized as follows\.

- •We formulate corpus\-level long\-context reasoning as evidence lifecycle management, viewing the construction and continual refinement of the evidence supporting a decision as the central reasoning process\. This perspective unifies evidence retrieval, editing, and transformation through the evolving role of evidence in decision\-making\.
- •We developLongHarness, a model\-agnostic harness that operationalizes this formulation through*Evidence–Conclusion Coevolution*and*Evidence Lifecycle Management*\. It maintains evidence and conclusions as coupled persistent runtime states, allowing unresolved conclusions to direct evidence acquisition and newly acquired evidence to revise those conclusions, while anEvidenceBankexplicitly controls the evidence lifecycle and preserves the sources of active evidence\.
- •We evaluateLongHarnesson corpus\-level reasoning over roughly 100M\-token and doubled document collections, together with provided\-context benchmarks, using two model backbones\. With GLM\-5\.3\-Flash,LongHarnessoutperforms Claude Code by 12\.0 percentage points on OfficeQA Pro V1 and Codex by 18\.9 points on V2, while ablations and trajectory analyses reveal the contributions of conclusion\-guided exploration and evidence lifecycle management\.

## 2Method

### 2\.1Overview

Given a questionqqand documentsD=\{d1,…,dn\}D=\\\{d\_\{1\},\\ldots,d\_\{n\}\\\}outside the prompt,LongHarnessincrementally acquires evidence through an external environmentℰD\\mathcal\{E\}\_\{D\}under a resource budget\. Its stateSt=\(Et,Ct,Ht\)S\_\{t\}=\(E\_\{t\},C\_\{t\},H\_\{t\}\)comprises source\-linked evidence, evidence\-linked conclusions and unresolved gaps, and prior reasoning, actions, and observations\. Givenqq, document mapMDM\_\{D\}, stateStS\_\{t\}, and remaining budgetBtB\_\{t\}, the policy selects an action and an optional conclusion update, and the environment returns the next state and observation\.

\(at,C^t\)\\displaystyle\(a\_\{t\},\\widehat\{C\}\_\{t\}\)∼πθ\(⋅∣q,MD,St,Bt\),\\displaystyle\\sim\\pi\_\{\\theta\}\(\\cdot\\mid q,M\_\{D\},S\_\{t\},B\_\{t\}\),\(1\)\(xt\+1,ot\+1\)\\displaystyle\(x\_\{t\+1\},o\_\{t\+1\}\)∼𝒫D\(⋅,⋅∣xt,at\),\\displaystyle\\sim\\mathcal\{P\}\_\{D\}\(\\cdot,\\cdot\\mid x\_\{t\},a\_\{t\}\),\(2\)Here,xtx\_\{t\}denotes the environment state, including registered source spans, evidence entries and their lifecycle status, and resource counters\. Its active evidence formsEtE\_\{t\}in the policy stateStS\_\{t\}\. The actionata\_\{t\}specifies search, bounded reading,EvidenceBankoperations, computation, or final submission, andC^t\\widehat\{C\}\_\{t\}is the optional conclusion update\. The controller validates action structure, registered source references, and budget constraints before exposing updates to subsequent turns\.

LongHarnesscouples two mechanisms \(Figure[2](https://arxiv.org/html/2610.13075#S2.F2)\)\. Evidence–Conclusion Coevolution uses unresolved conclusions to guide evidence acquisition and accepted evidence to strengthen or revise conclusions \(Section[2\.2](https://arxiv.org/html/2610.13075#S2.SS2)\)\. Evidence Lifecycle Management controls evidence admission, revision, retraction, and supersession while preserving source references \(Section[2\.3](https://arxiv.org/html/2610.13075#S2.SS3)\)\. A shared interface supports adaptive interleaving of these operations with document access, computation, and final submission \(Section[2\.4](https://arxiv.org/html/2610.13075#S2.SS4)\)\.

Figure 2:Overview ofLongHarness\. A single agent maintains anEvidenceBankand aConclusion Statewhile interacting with external documents\. Evidence–Conclusion Coevolution uses unresolved conclusions to guide exploration and accepted evidence to revise those conclusions, while Evidence Lifecycle Management governs which statements remain active and preserves their source references throughout the interaction\.
### 2\.2Evidence–Conclusion Coevolution

Separating evidence from inference\.TheEvidenceBankmaintains source\-linked statements extracted from passages the agent has read and serves as the persistent state in which evidence is admitted and revised, while theConclusion Staterecords the agent’s current interpretation of that evidence through established claims, supported hypotheses, and unresolved gaps\. Maintaining these representations separately distinguishes the acquired evidence from the inferences drawn from it, while explicit references connect each claim to its supporting evidence\. Each unresolved gap identifies information needed to establish, qualify, or revise the current conclusion\. At the level of a reasoning cycle, the two states coevolve through conclusion\-guided evidence acquisition and evidence\-grounded conclusion revision,

E\+\\displaystyle E^\{\+\}=𝒜θ​\(q,D,E,C\),\\displaystyle=\\mathcal\{A\}\_\{\\theta\}\(q,D,E,C\),\(3\)C\+\\displaystyle C^\{\+\}=ℛθ​\(q,E\+,C\)\.\\displaystyle=\\mathcal\{R\}\_\{\\theta\}\(q,E^\{\+\},C\)\.\(4\)where𝒜θ\\mathcal\{A\}\_\{\\theta\}represents exploration guided by the current conclusions and admission of the resulting statements into theEvidenceBank, whileℛθ\\mathcal\{R\}\_\{\\theta\}updates theConclusion Stateusing the active evidence in theEvidenceBank\. The same agent policy realizes both operations over one or more interactions, with the equations specifying their dependencies while allowing exploration and revision to be interleaved adaptively\.

Conclusion\-guided exploration\.The currentConclusion Stateguides what the agent searches for and reads next\. For each claim, the agent identifies observations that would establish, qualify, or refute it and uses these requirements to formulate queries and select reading scopes\. Competing hypotheses motivate searches that distinguish between them, while unresolved gaps direct attention to missing premises\. When an observation leaves a gap unresolved, the agent can reformulate the query or broaden the reading scope while keeping further exploration tied to the information needed to advance the conclusion\.

Evidence\-grounded revision\.After observing the result of anEvidenceBankupdate, the agent reassesses the claims in theConclusion Stateagainst the bank’s active evidence and determines whether they remain supported, require qualification, or must be revised\. It updates the affected claims together with their evidence references and unresolved gaps, keeping theConclusion Statealigned with the current evidence\. Before final submission, the agent incorporates any relevant evidence not yet reflected in this state and checks whether the remaining gaps could change the requested decision\.

### 2\.3Evidence Lifecycle Management

Evidence Lifecycle Management governs how source\-linked statements enter, persist in, and leave the active evidence state in theEvidenceBank\. A statement begins as a candidate extracted from a passage the agent has read, becomes active upon admission, and may subsequently be revised or retired as new evidence arrives\. This lifecycle requires \(i\) traceability to observed source passages, \(ii\) persistent availability across interactions, and \(iii\) explicit updates when a statement’s supporting evidence changes\.

TheEvidenceBankimplements these requirements through source\-linked admission, persistent storage, and explicit revision\. For traceability, each candidate statement must cite source references registered by prior read actions, with the controller checking reference validity and the agent judging whether the statement faithfully represents its sources\. For persistence, admitted statements remain available in theEvidenceBankacross subsequent interactions, allowing the agent to reuse them without reconstructing the evidence from earlier observations\. For explicit revision, a corrected statement is admitted as a new active entry and the previous entry is marked as superseded, while a retracted statement leaves the active state without a replacement\. The agent observes theseEvidenceBankupdates before revising the affected conclusions, and final references resolve to active statements and their underlying source passages\.

### 2\.4Agent–Environment Interaction

At each turn, the policy selects an actionata\_\{t\}and an optional conclusion updateC^t\\widehat\{C\}\_\{t\}under the formulation in Section[2\.1](https://arxiv.org/html/2610.13075#S2.SS1)\. The environment executes the action and returns an observation, while the harness enforces the interaction protocol, source\-reference constraints, and resource limits\. The shared action space lets the policy interleave evidence acquisition, maintenance, and use\.*Search*locates candidate documents or passages, and*Read*returns selected text and registers source references\. The agent proposes source\-linked evidence throughupsertand corrects or retracts entries throughrevise\(Section[2\.3](https://arxiv.org/html/2610.13075#S2.SS3)\)\.*Compute*evaluates numerical expressions supplied by the policy and validates their declared evidence references, and*Final*submits the answer with supporting references\.

EvidenceBankmutations defer conclusion updates until the agent observes their results\. The policy then revises affected claims and references through optional partial updates, with omitted fields retained\. Superseding or retracting evidence does not automatically rewrite dependent conclusions\. The controller checks declared evidence references at computation and final submission, while the policy remains responsible for source fidelity and whether the evidence supports its conclusions\. Reasoning, actions, and observations accumulate in the history, while evidence and conclusions persist across turns\. Final acceptance requires a valid answer structure and references resolving to active evidence and source passages\. Resource exhaustion or runtime failure may terminate the episode without an accepted answer\. Algorithm[1](https://arxiv.org/html/2610.13075#alg1)in Appendix[B\.5](https://arxiv.org/html/2610.13075#A2.SS5)summarizes the procedure\.

## 3Experiments

Table 1:Main results \(%\) across four benchmarks\. OfficeQA Pro reports exact match \(Ex\.\) and accuracy within 5% relative error \(@5%\), while AA\-LCR and the LongBench v2 Long subset report Exact accuracy\. Direct baselines use BM25 retrieval into a 256K\-token window on OfficeQA Pro\. “\-” denotes direct reasoning without a harness, and “Avg\.” averages the Exact scores across benchmarks\. Each harness result is from one run per setting\. Direct AA\-LCR scores are reported by Artificial Analysis; harness AA\-LCR scores are evaluated in this work\.### 3\.1Experimental Setup

#### Benchmarks\.

We evaluate on OfficeQA Pro V1 and V2\([Opsahl\-Ong et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib26)\), AA\-LCR\([Artificial Analysis Team, 2026](https://arxiv.org/html/2610.13075#bib.bib2)\), and the Long subset of LongBench v2\([Bai et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib3)\), covering both corpus\-level reasoning and reasoning over provided contexts\. OfficeQA Pro V1 contains 697 Treasury documents totaling approximately 100M tokens, while V2 expands the corpus to 1,435 documents and introduces composite statistical questions that require combining values from multiple sources\. Agents receive access to the full corpus without question\-specific gold documents, making source discovery part of the task\. AA\-LCR provides the relevant context for each of its 100 questions, while the LongBench v2 Long subset evaluates reasoning over the benchmark’s longest contexts\. We report exact match and accuracy within 5% relative error on OfficeQA Pro, and Exact accuracy on AA\-LCR and LongBench v2\.

#### Models and harnesses\.

Our direct\-reasoning baselines include Qwen3\.8\-27B, Qwen3\.7\-Max, and Qwen3\.8\-Flash\([Yang et al\., 2025a](https://arxiv.org/html/2610.13075#bib.bib38)\); DeepSeek\-V4\-Flash\-0731 and DeepSeek\-V4\-Pro\-0813\([DeepSeek\-AI et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib7)\); Gemini\-3\.8\-Flash\([Google DeepMind, 2026](https://arxiv.org/html/2610.13075#bib.bib12)\); GPT\-5\.6\-Luna and GPT\-5\.6\-Sol\([Singh et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib33)\); GLM\-5\.3 and GLM\-5\.3\-Flash\([GLM\-5\-Team et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib11)\); and Kimi\-K3\([Kimi Team et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib17)\)\. Citations refer to the corresponding model families or release\-specific model cards\. On OfficeQA Pro, these models receive question\-relevant passages retrieved with BM25 into a 256K\-token context window, while LongBench v2 contexts follow the official truncation procedure with the same token limit\. We compare Codex\([OpenAI, 2026](https://arxiv.org/html/2610.13075#bib.bib25)\), Claude Code\([Anthropic, 2026](https://arxiv.org/html/2610.13075#bib.bib1)\), RLM\([Zhang et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib42)\), andLongHarnessusing GLM\-5\.3\-Flash and Qwen3\.8\-Flash as matched backbones, holding model identity fixed within each comparison\. All harnesses receive the same questions and document corpus without access to reference answers or gold source annotations\.

#### Implementation details\.

All backbone models use temperature 1, top\-p 0\.95, maximum reasoning effort, and an output limit of 32,768 tokens per request\. Codex and Claude Code each allow up to 200 model turns and 200 tool calls, while RLM allows 30 iterations andLongHarnessallows 200 steps subject to a cumulative read budget of 1,000,000 tokens and a limit of 60 search calls\. We grade AA\-LCR harness responses with GPT\-5\.6\-Luna\. Due to computational resource constraints, the main comparison uses one run per harness and evaluation setting\. To assess variability across runs, Appendix[D](https://arxiv.org/html/2610.13075#A4)reports three\-run results with GLM\-5\.3\-Flash, including individual scores, means, standard deviations, and maxima\. Missing answers and agent termination failures are counted as incorrect\.

### 3\.2Main Results

Table[1](https://arxiv.org/html/2610.13075#S3.T1)shows that direct baselines reach at most 23\.2% exact match on OfficeQA Pro V1 and 5\.6% on V2 despite receiving question\-relevant passages in a 256K\-token window\. At corpus scale, this fixed window cannot accommodate all potentially relevant evidence, and passages selected before reasoning may omit information needed to resolve intermediate conclusions\.LongHarnessuses these conclusions to guide further retrieval and accumulates the resulting evidence across interactions, raising GLM\-5\.3\-Flash exact match from 15\.8% to 60\.2% on V1 and from 4\.4% to 61\.1% on V2\.

With GLM\-5\.3\-Flash,LongHarnessoutperforms the strongest competing harness by 12\.0 percentage points on OfficeQA Pro V1 and 18\.9 points on V2\. The larger advantage on V2 is consistent with maintaining evidence across interactions as more scattered values must be located, reconciled, and combined\. With Qwen3\.8\-Flash,LongHarnessalso leads on both versions, reaching 61\.7% on V1 and 51\.1% on V2\. Its advantage over the strongest competing harness is 8\.3 points on V1 but narrows to 1\.1 points over Codex on V2, showing that the size of the gain varies with the backbone and evaluation setting\. On the LongBench v2 Long subset,LongHarnessreaches 75\.9% with GLM\-5\.3\-Flash and 76\.9% with Qwen3\.8\-Flash, exceeding the next\-best harness by 6\.5 and 7\.4 percentage points, respectively\. On AA\-LCR, it scores 84\.0% and 87\.0%, leading the strongest competing harness by 5\.0 and 4\.0 points\.LongHarnessachieves the highest average Exact score across the four evaluation settings with both backbones, reaching 70\.3% and 69\.2%, respectively\.

Table 2:Ablation results with GLM\-5\.3\-Flash on OfficeQA Pro V1 and V2\. Ex\. denotes exact\-match accuracy \(%\), andΔ\\Deltadenotes the decrease from fullLongHarnessin percentage points, computed before rounding\.
### 3\.3Ablation Study

We ask whether evidence–conclusion coevolution and evidence lifecycle management are critical to the effectiveness ofLongHarness, and examine this question through five ablations in Table[2](https://arxiv.org/html/2610.13075#S3.T2)\.\(i\) w/o Persistent Statesremoves theEvidenceBankandConclusion Statewhile retaining interaction history\.\(ii\) w/o Evidence–Conclusion Coevolutionretains theEvidenceBankbut removes theConclusion Stateand its retrieval feedback\.\(iii\) Naive Evidence Accumulationpreserves conclusion feedback but permits only append\-only evidence updates\.\(iv\) w/o Explicit Conclusion\-Guided Explorationretains both states but removes the instructions to guide retrieval using conclusion gaps\.\(v\) Generic Working Notesretains theEvidenceBankand replaces structured conclusions with editable free\-text notes that can record conclusions, unresolved questions, and evidence references\.

Removing both persistent states reduces exact match by 4\.5 percentage points on V1 and 7\.8 on V2, showing that explicit evidence and conclusion states contribute beyond access to the interaction history\. Removing the maintainedConclusion Stateand its retrieval feedback reduces accuracy by 5\.3 and 9\.4 points\. Restricting theEvidenceBankto append\-only accumulation yields 56\.4% on V1 and 51\.1% on V2, corresponding to decreases of 3\.8 and 10\.0 points\. These results support the value of revising and retiring evidence as new observations arrive, even when conclusion feedback is preserved\. Removing the explicit conclusion\-guided exploration instructions while retaining both persistent states reduces accuracy by 4\.5 and 7\.8 points, supporting the value of instructing the policy to use conclusion gaps to guide retrieval beyond making those states available\. Replacing structured conclusions with generic working notes reduces accuracy by 6\.8 and 7\.8 points, showing that the full design outperforms editable free\-text notes that can also retain conclusions, open questions, and evidence references\. All five ablations incur larger losses on V2, consistent with the greater importance of maintaining and reconciling evidence when questions require combining values from multiple sources\.

Table 3:Success rates \(%\) for evidence retrieval, editing, and transformation on OfficeQA Pro V1 and V2 with GLM\-5\.3\-Flash\. Retrieval measures whether the required evidence is found, editing measures whether retrieved evidence is retained and adopted in the final decision, and transformation measures whether the adopted evidence yields an exactly correct answer\. Editing and transformation rates are calculated only over cases that pass the preceding stages\. Each harness is evaluated on one run per split\.
### 3\.4Interaction Length and Trajectory Analysis

We ask where agent harnesses fail across evidence retrieval, editing, and transformation, and how these failures change from OfficeQA Pro V1 to V2\. For each harness, we use GPT\-5\.6\-Luna at temperature 0 to audit incorrect episodes and assess evidence retrieval, evidence adoption, and answer transformation\. Exact\-correct episodes are counted as passing the first two stages\. LetE1E\_\{1\}denote the presence of required evidence in the trajectory,E2E\_\{2\}its retention and reconciliation for use in the final decision, andE3E\_\{3\}an exactly correct answer\. We measure retrieval coverage and conditional editing and transformation success as

r1=P⁡\(E1\),r2=P⁡\(E2∣E1\),r3=P⁡\(E3∣E1,E2\)\.r\_\{1\}=P\(E\_\{1\}\),\\qquad r\_\{2\}=P\(E\_\{2\}\\mid E\_\{1\}\),\\qquad r\_\{3\}=P\(E\_\{3\}\\mid E\_\{1\},E\_\{2\}\)\.\(5\)
Figure 3:Distributions of model turns per question on OfficeQA Pro V1 and V2 with GLM\-5\.3\-Flash\. Each panel compares the two versions for one harness\. Curves show kernel density estimates, and circles indicate medians\.Figure 4:Evidence lifecycle comparison with GLM\-5\.3\-Flash on OfficeQA Pro V1\. For the geometric mean of 80 monthly Treasury expenditures, Codex continues searching and omits a recognized correction;LongHarnessupdates its evidence and returns the reference answer, 4962\.46 million dollars\. Circles mark Codex’s incorrect input, correction, and final answer\.Table[3](https://arxiv.org/html/2610.13075#S3.T3)shows thatLongHarnessmatches Claude Code’s retrieval coverage on V1 and achieves higher conditional evidence\-adoption and answer\-conversion rates\. On V2, its retrieval, evidence\-adoption, and answer\-conversion rates are 87\.8%, 87\.3%, and 79\.7%, respectively\. The conditional adoption and conversion rates remain close to their V1 values of 87\.6% and 80\.8%, indicating similar evidence\-use outcomes across the two evaluated settings\. The conditional rates are computed over the questions that pass the preceding stages for each harness and describe stage outcomes within the analyzed runs\.

Figure[4](https://arxiv.org/html/2610.13075#S3.F4)shows that median turns increase from 22 to 30 forLongHarness, from 21 to 116 for Claude Code, and from 19 to 43 for Codex between V1 and V2\. Codex exhausts its recorded model\-turn budget on 3\.8% of V1 questions and 21\.1% of V2 questions\. Claude Code’s distribution also concentrates more heavily near 200 turns on V2\. V2 combines a larger corpus with questions requiring values from multiple sources\.LongHarnessrecords unresolved information needs in theConclusion Stateand retains acquired evidence in theEvidenceBank, allowing subsequent retrieval to focus on missing information while reusing values already established\. As questions require more cross\-source checks, this organization helps avoid repeatedly searching for or re\-establishing the same facts\. This offers a plausible explanation for the smaller shift inLongHarness’s turn distribution from V1 to V2\. Figure[4](https://arxiv.org/html/2610.13075#S3.F4)compares how Codex andLongHarnesscarry evidence corrections into the final computation, with a detailed case analysis in Appendix[E](https://arxiv.org/html/2610.13075#A5)\.

## 4Related Work

### 4\.1Long\-Context Reasoning

Post\-training supervises evidence acquisition and use\. LongRLVR rewards correct identification of evidence chunks\([Chen et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib5)\), LongPAS assigns advantages to reasoning steps\([Peng et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib29)\), and LongR measures document consultation through information gain\([Ping et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib30)\)\. Training data include key\-fact chains\([Wang et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib36)\), distractor\-augmented search trajectories\([Lin et al\., 2026b](https://arxiv.org/html/2610.13075#bib.bib21)\), capability\-aligned task mixtures\([Lv et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib24)\), and verifiable multi\-hop or memory\-management tasks\([Wan et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib35);[Shen et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib32)\)\. Training also uses agent\-generated data and supervision for memory operations\. ACC compiles agent trajectories into long\-context QA training examples\([Su et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib34)\), AgenticLU distills self\-generated clarification trajectories without external teachers\([Zhuang et al\., 2025b](https://arxiv.org/html/2610.13075#bib.bib49)\), MAVEN rewards evidence\-memory edits\([Gao and Marttinen, 2026](https://arxiv.org/html/2610.13075#bib.bib9)\), and Memory\-R2 credits individual memory operations\([Yan et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib37)\)\.LongHarnessinstead couples evidence and conclusions at runtime, linking retrieval, editing, and transformation without additional policy training\.

### 4\.2Agentic Harnesses

ReAct couples reasoning with action\([Yao et al\., 2023](https://arxiv.org/html/2610.13075#bib.bib40)\), recursive language models process context under program control\([Zhang et al\., 2025](https://arxiv.org/html/2610.13075#bib.bib42)\), and Prime Agent maintains persistent REPL state\([Karten et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib15)\)\. Codex and Claude Code support long\-context reasoning\([Cao et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib4)\); other harnesses develop bounded long\-horizon decomposition, harness maintenance, and persistent memory\([Zheng et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib47);[Lin et al\., 2026a](https://arxiv.org/html/2610.13075#bib.bib20);[Packer et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib27);[Yu et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib41);[Zhang et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib43);[Pan et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib28)\)\. MemWalker uses memory trees\([Chen et al\., 2023](https://arxiv.org/html/2610.13075#bib.bib6)\), ReadAgent uses gist representations\([Lee et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib18)\), GraphReader uses chunk graphs\([Li et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib19)\), and multi\-agent systems coordinate readers\([Zhao et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib45);[Zhang et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib44)\)\. SLIDERS reconciles evidence under an induced schema\([Joshi et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib14)\), ReContext replays found\-but\-unused evidence\([Zhao et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib46)\), and TRACE evaluates tool\-augmented reasoning trajectories\([Kim et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib16)\)\. LedgerMind uses grounding checks and event\-triggered repair to address unsupported or inconsistent claims\([Du et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib8)\)\. However, existing claims can remain grounded while the evidence needed to answer the question is still incomplete\. In contrast,LongHarnessfocuses on corpus\-level long\-context reasoning, maintaining source\-linked evidence alongside evolving conclusions and unresolved information needs\. These needs guide subsequent search and reading, while newly acquired evidence informs revisions to conclusions and remaining gaps\.

## 5Conclusion

We proposeLongHarness, a model\-agnostic harness that unifies evidence retrieval, editing, and transformation for corpus\-level long\-context reasoning\. Evidence–Conclusion Coevolution couples exploration with conclusion revision, while Evidence Lifecycle Management keeps evidence current and traceable\. Evaluations on OfficeQA Pro, AA\-LCR, and LongBench v2 show gains with two backbones\. With GLM\-5\.3\-Flash,LongHarnessexceeds Claude Code by 12\.0 percentage points on OfficeQA Pro V1 and Codex by 18\.9 points on V2\. Ablations support the complementary roles of persistent states, conclusion\-guided retrieval, and evidence revision\. Trajectory analysis shows similar conditional evidence\-use rates across V1 and V2 and a smaller increase in interaction length forLongHarnessthan for the competing harnesses\.

## References

- Anthropic \(2026\)Anthropic\.Claude Code\.Software, 2026\.[https://github\.com/anthropics/claude\-code](https://github.com/anthropics/claude-code)\.
- Artificial Analysis Team \(2026\)Artificial Analysis Team\.Artificial analysis long context reasoning \(AA\-LCR\) dataset\.Hugging Face Datasets, 2026\.[https://huggingface\.co/datasets/ArtificialAnalysis/AA\-LCR](https://huggingface.co/datasets/ArtificialAnalysis/AA-LCR)\.
- Bai et al\. \(2024\)Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li\.LongBench v2: Towards deeper understanding and reasoning on realistic long\-context multitasks\.*arXiv preprint arXiv:2412\.15204*, 2024\.[https://arxiv\.org/abs/2412\.15204](https://arxiv.org/abs/2412.15204)\.
- Cao et al\. \(2026\)Weili Cao, Xunjian Yin, Bhuwan Dhingra, and Shuyan Zhou\.Coding agents are effective long\-context processors\.*arXiv preprint arXiv:2603\.20432*, 2026\.[https://arxiv\.org/abs/2603\.20432](https://arxiv.org/abs/2603.20432)\.
- Chen et al\. \(2026\)Guanzheng Chen, Michael Qizhe Shieh, and Lidong Bing\.Longrlvr: Long\-context reinforcement learning requires verifiable context rewards\.*arXiv preprint arXiv:2603\.02146*, 2026\.[https://arxiv\.org/abs/2603\.02146](https://arxiv.org/abs/2603.02146)\.
- Chen et al\. \(2023\)Howard Chen, Ramakanth Pasunuru, Jason Weston, and Asli Celikyilmaz\.Walking down the memory maze: Beyond context limit through interactive reading\.*arXiv preprint arXiv:2310\.05029*, 2023\.[https://arxiv\.org/abs/2310\.05029](https://arxiv.org/abs/2310.05029)\.
- DeepSeek\-AI et al\. \(2026\)DeepSeek\-AI et al\.DeepSeek\-V4: Towards Highly Efficient Million\-Token Context Intelligence, 2026\.[https://arxiv\.org/abs/2606\.19348](https://arxiv.org/abs/2606.19348)\.
- Du et al\. \(2026\)Enjun Du, Hange Zhou, Chenxu Du, Siyi Liu, Zirong Chen, Ziyu Zheng, and Yongqi Zhang\.LEDGERMIND: Provenance\-constrained multimodal agentic reasoning with a structured evidence ledger\.*arXiv preprint arXiv:2607\.28374*, 2026\.[https://arxiv\.org/abs/2607\.28374](https://arxiv.org/abs/2607.28374)\.
- Gao and Marttinen \(2026\)Ya Gao and Pekka Marttinen\.Evidence\-state rewards for long\-context reasoning\.*arXiv preprint arXiv:2607\.02073*, 2026\.[https://arxiv\.org/abs/2607\.02073](https://arxiv.org/abs/2607.02073)\.
- Gemini Team \(2024\)Gemini Team\.Gemini 1\.5: Unlocking multimodal understanding across millions of tokens of context\.*arXiv preprint arXiv:2403\.05530*, 2024\.[https://arxiv\.org/abs/2403\.05530](https://arxiv.org/abs/2403.05530)\.
- GLM\-5\-Team et al\. \(2026\)GLM\-5\-Team et al\.GLM\-5: from Vibe Coding to Agentic Engineering, 2026\.[https://arxiv\.org/abs/2602\.15763](https://arxiv.org/abs/2602.15763)\.
- Google DeepMind \(2026\)Google DeepMind\.Gemini 3\.8 Flash Model Card, September 2026\.[https://deepmind\.google/models/model\-cards/gemini\-3\-8\-flash/](https://deepmind.google/models/model-cards/gemini-3-8-flash/)\.
- Hsieh et al\. \(2024\)Cheng\-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg\.RULER: What’s the real context size of your long\-context language models?*Proceedings of the Conference on Language Modeling*, 2024\.[https://arxiv\.org/abs/2404\.06654](https://arxiv.org/abs/2404.06654)\.
- Joshi et al\. \(2026\)Harshit Joshi, Priyank Shethia, Jadelynn Dao, and Monica S\. Lam\.SLIDERS: Systematic reviews via automated evidence synthesis and reconciliation\.*arXiv preprint arXiv:2604\.22294*, 2026\.[https://arxiv\.org/abs/2604\.22294](https://arxiv.org/abs/2604.22294)\.
- Karten et al\. \(2026\)Seth Karten, Alex L\. Zhang, Kevin Thomas, Sebastian Mueller, Elie Bakouch, Daniel Auras, Mika Senghaas, Fares Obeid, Konstantin Dunas, Johannes Hagemann, and Sami Jaghouar\.Prime agent: A self\-improving rlm harness\.*arXiv preprint arXiv:2608\.23552*, 2026\.[https://arxiv\.org/abs/2608\.23552](https://arxiv.org/abs/2608.23552)\.
- Kim et al\. \(2026\)Wonjoong Kim, Sangwu Park, Yeonjun In, Sein Kim, Dongha Lee, and Chanyoung Park\.Beyond the final answer: Evaluating the reasoning trajectories of tool\-augmented agents\.*Proceedings of the 43rd International Conference on Machine Learning*, 2026\.[https://arxiv\.org/abs/2510\.02837](https://arxiv.org/abs/2510.02837)\.
- Kimi Team et al\. \(2026\)Kimi Team et al\.Kimi K3: Open Frontier Intelligence, 2026\.[https://arxiv\.org/abs/2607\.24653](https://arxiv.org/abs/2607.24653)\.
- Lee et al\. \(2024\)Kuang\-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer\.A human\-inspired reading agent with gist memory of very long contexts\.*arXiv preprint arXiv:2402\.09727*, 2024\.[https://arxiv\.org/abs/2402\.09727](https://arxiv.org/abs/2402.09727)\.
- Li et al\. \(2024\)Shilong Li, Yancheng He, Hangyu Guo, Xingyuan Bu, Ge Bai, Jie Liu, Jiaheng Liu, Xingwei Qu, Yangguang Li, Wanli Ouyang, Wenbo Su, and Bo Zheng\.GraphReader: Building graph\-based agent to enhance long\-context abilities of large language models\.In*Findings of the Association for Computational Linguistics: EMNLP 2024*, pages 12758–12786, 2024\.[https://aclanthology\.org/2024\.findings\-emnlp\.746/](https://aclanthology.org/2024.findings-emnlp.746/)\.
- Lin et al\. \(2026a\)Jiahang Lin, Shichun Liu, Chengjun Pan, Lizhi Lin, Shihan Dou, Zhiheng Xi, Xuanjing Huang, Hang Yan, Zhenhua Han, Tao Gui, and Yu\-Gang Jiang\.Agentic harness engineering: Observability\-driven automatic evolution of coding\-agent harnesses\.*arXiv preprint arXiv:2604\.25850*, 2026a\.[https://arxiv\.org/abs/2604\.25850](https://arxiv.org/abs/2604.25850)\.
- Lin et al\. \(2026b\)Nianyi Lin, Jiajie Zhang, Lei Hou, and Juanzi Li\.Longtracerl: Learning long\-context reasoning from search agent trajectories with rubric rewards\.*arXiv preprint arXiv:2605\.31584*, 2026b\.[https://arxiv\.org/abs/2605\.31584](https://arxiv.org/abs/2605.31584)\.
- Ling et al\. \(2025\)Zhan Ling, Kang Liu, Kai Yan, Yifan Yang, Weijian Lin, Ting\-Han Fan, Lingfeng Shen, Zhengyin Du, and Jiecao Chen\.Longreason: A synthetic long\-context reasoning benchmark via context expansion\.*arXiv preprint arXiv:2501\.15089*, 2025\.[https://arxiv\.org/abs/2501\.15089](https://arxiv.org/abs/2501.15089)\.
- Liu et al\. \(2024\)Nelson F\. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang\.Lost in the middle: How language models use long contexts\.*Transactions of the Association for Computational Linguistics*, 12:157–173, 2024\.[https://aclanthology\.org/2024\.tacl\-1\.9/](https://aclanthology.org/2024.tacl-1.9/)\.
- Lv et al\. \(2026\)Minxuan Lv, Tiehua Mei, Tanlong Du, Junmin Chen, Zhenpeng Su, Ziyang Chen, Ziqi Wang, Zhennan Wu, Ruotong Pan, Jian Liang, Ruiming Tang, and Han Li\.GoLongRL: Capability\-oriented long context reinforcement learning with multitask alignment\.*arXiv preprint arXiv:2605\.19577*, 2026\.[https://arxiv\.org/abs/2605\.19577](https://arxiv.org/abs/2605.19577)\.
- OpenAI \(2026\)OpenAI\.Codex CLI\.Software, 2026\.[https://github\.com/openai/codex](https://github.com/openai/codex)\.
- Opsahl\-Ong et al\. \(2026\)Krista Opsahl\-Ong, Arnav Singhvi, Jasmine Collins, Ivan Zhou, Cindy Wang, Ashutosh Baheti, Owen Oertell, Jacob Portes, Sam Havens, Erich Elsen, Michael Bendersky, Matei Zaharia, and Xing Chen\.OfficeQA Pro: An enterprise benchmark for end\-to\-end grounded reasoning, 2026\.[https://arxiv\.org/abs/2603\.08655](https://arxiv.org/abs/2603.08655)\.
- Packer et al\. \(2024\)Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G\. Patil, Ion Stoica, and Joseph E\. Gonzalez\.Memgpt: Towards llms as operating systems, 2024\.[https://arxiv\.org/abs/2310\.08560](https://arxiv.org/abs/2310.08560)\.
- Pan et al\. \(2026\)Zhuoshi Pan, Qizhi Pei, Junru Lu, Honglin Lin, H\. Vicky Zhao, Di Yin, and Xing Sun\.ContextPilot: Teaching agents for proactive context management via fine\-grained RL\.In*Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing*, 2026\.[https://arxiv\.org/abs/2608\.28476](https://arxiv.org/abs/2608.28476)\.
- Peng et al\. \(2026\)Miao Peng, Weizhou Shen, Nuo Chen, Chenliang Li, Ming Yan, and Jia Li\.Incentivizing in\-depth reasoning over long contexts with process advantage shaping\.*arXiv preprint arXiv:2601\.12465*, 2026\.[https://arxiv\.org/abs/2601\.12465](https://arxiv.org/abs/2601.12465)\.
- Ping et al\. \(2026\)Bowen Ping, Zijun Chen, Yiyao Yu, Tingfeng Hui, Junchi Yan, and Baobao Chang\.LongR: Unleashing long\-context reasoning via reinforcement learning with dense utility rewards\.*arXiv preprint arXiv:2602\.05758*, 2026\.[https://arxiv\.org/abs/2602\.05758](https://arxiv.org/abs/2602.05758)\.
- Shang et al\. \(2025\)Ning Shang, Li Lyna Zhang, Siyuan Wang, Gaokai Zhang, Gilsinia Lopez, Fan Yang, Weizhu Chen, and Mao Yang\.LongRoPE2: Near\-lossless LLM context window scaling\.*arXiv preprint arXiv:2502\.20082*, 2025\.[https://arxiv\.org/abs/2502\.20082](https://arxiv.org/abs/2502.20082)\.
- Shen et al\. \(2025\)Weizhou Shen, Ziyi Yang, Chenliang Li, Zhiyuan Lu, Miao Peng, Huashan Sun, Yingcheng Shi, Shengyi Liao, Shaopeng Lai, Bo Zhang, Dayiheng Liu, Fei Huang, Jingren Zhou, and Ming Yan\.QwenLong\-L1\.5: Post\-training recipe for long\-context reasoning and memory management\.*arXiv preprint arXiv:2512\.12967*, 2025\.[https://arxiv\.org/abs/2512\.12967](https://arxiv.org/abs/2512.12967)\.
- Singh et al\. \(2026\)Aaditya Singh, Adam Fry, Adam Perelman, et al\.OpenAI GPT\-5 System Card, 2026\.[https://arxiv\.org/abs/2601\.03267](https://arxiv.org/abs/2601.03267)\.
- Su et al\. \(2026\)Qisheng Su, Zhen Fang, Shiting Huang, Yu Zeng, Yiming Zhao, Kou Shi, Ziao Zhang, Lin Chen, Zehui Chen, Lijun Wu, and Feng Zhao\.Acc: Compiling agent trajectories for long\-context training\.*arXiv preprint arXiv:2605\.21850*, 2026\.[https://arxiv\.org/abs/2605\.21850](https://arxiv.org/abs/2605.21850)\.
- Wan et al\. \(2025\)Fanqi Wan, Weizhou Shen, Shengyi Liao, Yingcheng Shi, Chenliang Li, Ziyi Yang, Ji Zhang, Fei Huang, Jingren Zhou, and Ming Yan\.Qwenlong\-l1: Towards long\-context large reasoning models with reinforcement learning, 2025\.[https://arxiv\.org/abs/2505\.17667](https://arxiv.org/abs/2505.17667)\.
- Wang et al\. \(2025\)Siyuan Wang, Gaokai Zhang, Li Lyna Zhang, Ning Shang, Fan Yang, Dongyao Chen, and Mao Yang\.Loongrl: Reinforcement learning for advanced reasoning over long contexts\.*arXiv preprint arXiv:2510\.19363*, 2025\.[https://arxiv\.org/abs/2510\.19363](https://arxiv.org/abs/2510.19363)\.
- Yan et al\. \(2026\)Sikuan Yan, Ahmed Bahloul, Ercong Nie, Susanna Schwarzmann, Riccardo Trivisonno, Volker Tresp, and Yunpu Ma\.Memory\-R2: Fair credit assignment for long\-horizon memory\-augmented LLM agents\.*arXiv preprint arXiv:2605\.21768*, 2026\.[https://arxiv\.org/abs/2605\.21768](https://arxiv.org/abs/2605.21768)\.
- Yang et al\. \(2025a\)An Yang, Anfeng Li, Baosong Yang, et al\.Qwen3 Technical Report, 2025a\.[https://arxiv\.org/abs/2505\.09388](https://arxiv.org/abs/2505.09388)\.
- Yang et al\. \(2025b\)An Yang, Bowen Yu, Chengyuan Li, et al\.Qwen2\.5\-1M technical report\.*arXiv preprint arXiv:2501\.15383*, 2025b\.[https://arxiv\.org/abs/2501\.15383](https://arxiv.org/abs/2501.15383)\.
- Yao et al\. \(2023\)Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao\.ReAct: Synergizing reasoning and acting in language models\.In*International Conference on Learning Representations*, 2023\.[https://openreview\.net/forum?id=WE\_vluYUL\-X](https://openreview.net/forum?id=WE_vluYUL-X)\.
- Yu et al\. \(2026\)Zhaochen Yu, Yingcheng Wu, Zhenfei Yin, Kaiyuan Chen, Zhe Zhao, Mengdi Wang, Shuicheng Yan, and Ling Yang\.Recursive experiential–working memory evolution for long\-horizon agent harnesses\.*arXiv preprint arXiv:2608\.24876*, 2026\.[https://arxiv\.org/abs/2608\.24876](https://arxiv.org/abs/2608.24876)\.
- Zhang et al\. \(2025\)Alex L\. Zhang, Tim Kraska, and Omar Khattab\.Recursive language models\.*arXiv preprint arXiv:2512\.24601*, 2025\.[https://arxiv\.org/abs/2512\.24601](https://arxiv.org/abs/2512.24601)\.
- Zhang et al\. \(2026\)Kehao Zhang, Shangtong Gui, Sheng Yang, Wei Chen, and Yang Feng\.Learning to remember: End\-to\-end training of memory agents for long\-context reasoning\.*arXiv preprint arXiv:2602\.18493*, 2026\.[https://arxiv\.org/abs/2602\.18493](https://arxiv.org/abs/2602.18493)\.
- Zhang et al\. \(2024\)Yusen Zhang, Ruoxi Sun, Yanfei Chen, Tomas Pfister, Rui Zhang, and Sercan Ö\. Arik\.Chain of agents: Large language models collaborating on long\-context tasks\.*arXiv preprint arXiv:2406\.02818*, 2024\.[https://arxiv\.org/abs/2406\.02818](https://arxiv.org/abs/2406.02818)\.
- Zhao et al\. \(2024\)Jun Zhao, Can Zu, Xu Hao, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang\.LONGAGENT: Achieving question answering for 128k\-token\-long documents through multi\-agent collaboration\.In*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing*, pages 16310–16324, 2024\.[https://aclanthology\.org/2024\.emnlp\-main\.912/](https://aclanthology.org/2024.emnlp-main.912/)\.
- Zhao et al\. \(2026\)Yanjun Zhao, Ruizhong Qiu, Tianxin Wei, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, and Jingrui He\.Recontext: Recursive evidence replay as LLM harness for long\-context reasoning\.*arXiv preprint arXiv:2607\.02509*, 2026\.[https://arxiv\.org/abs/2607\.02509](https://arxiv.org/abs/2607.02509)\.
- Zheng et al\. \(2026\)Jingsheng Zheng, Xinyuan Fang, Jintian Zhang, Zhengke Gui, Huajun Chen, and Ningyu Zhang\.OneDayAgent: Towards a long\-horizon harness for autonomous agents\.*arXiv preprint arXiv:2608\.05013*, 2026\.[https://arxiv\.org/abs/2608\.05013](https://arxiv.org/abs/2608.05013)\.
- Zhuang et al\. \(2025a\)Tianyi Zhuang, Chuqiao Kuang, Xiaoguang Li, Yihua Teng, Jihao Wu, Yasheng Wang, and Lifeng Shang\.Docpuzzle: A process\-aware benchmark for evaluating realistic long\-context reasoning capabilities\.*arXiv preprint arXiv:2502\.17807*, 2025a\.[https://arxiv\.org/abs/2502\.17807](https://arxiv.org/abs/2502.17807)\.
- Zhuang et al\. \(2025b\)Yufan Zhuang, Xiaodong Yu, Jialian Wu, Ximeng Sun, Ze Wang, Jiang Liu, Yusheng Su, Jingbo Shang, Zicheng Liu, and Emad Barsoum\.Self\-taught agentic long\-context understanding\.*arXiv preprint arXiv:2502\.15920*, 2025b\.[https://arxiv\.org/abs/2502\.15920](https://arxiv.org/abs/2502.15920)\.

## Appendix ABenchmark Details

Table[4](https://arxiv.org/html/2610.13075#A1.T4)summarizes the four evaluation settings, covering external\-corpus retrieval in OfficeQA Pro and reasoning over provided contexts in AA\-LCR and LongBench v2\.

Table 4:Benchmark sizes, evidence access, and evaluation metrics\. The AA\-LCR document count covers all document sets, and LongBench v2 lengths are measured in words from the preprocessed contexts used in our experiments\.#### OfficeQA Pro\.

OfficeQA Pro V1\([Opsahl\-Ong et al\., 2026](https://arxiv.org/html/2610.13075#bib.bib26)\)contains 133 questions over 697 Treasury documents totaling approximately 100M tokens, while V2 contains 90 questions over an expanded corpus of 1,435 documents\. V2 places greater emphasis on composite numerical questions requiring values from different series, versions, or documents\. Both versions expose the full corpus without question\-specific gold documents, requiring the system to identify the sources needed for each answer\. We report exact match, which requires the requested value and format, and accuracy within 5% relative error\.

#### AA\-LCR\-1\.1\.

We use AA\-LCR\-1\.1\([Artificial Analysis Team, 2026](https://arxiv.org/html/2610.13075#bib.bib2)\),111[https://huggingface\.co/datasets/ArtificialAnalysis/AA\-LCR](https://huggingface.co/datasets/ArtificialAnalysis/AA-LCR)which contains 100 questions associated with 30 document sets comprising 234 documents in total, with an average of 99,325 tokens per set\. The documents include company and industry reports, government consultations, academic papers, legal and marketing materials, and survey reports\. Questions require synthesis, temporal reasoning, conditional logic, or numerical computation across documents rather than direct retrieval from a single passage\. Each question supplies its relevant document set, allowing the evaluation to focus on evidence use within the provided context\. For direct\-reasoning models, we report the AA\-LCR scores published by Artificial Analysis in its official evaluation results\.

#### LongBench v2\.

LongBench v2\([Bai et al\., 2024](https://arxiv.org/html/2610.13075#bib.bib3)\)contains 503 four\-option questions covering single\-document QA, multi\-document QA, long in\-context learning, long\-dialogue history understanding, code\-repository understanding, and long structured\-data understanding\. The full benchmark spans contexts of 8K–2M words, most below 128K words\. We evaluate the official Long subset of 108 questions, whose contexts in the release used here range from approximately 132K to 1\.38M words with a median of 204K words\. This subset contains 29 code\-repository understanding questions, 26 long in\-context learning questions, 24 single\-document QA questions, 23 multi\-document QA questions, and 6 structured\-data understanding questions\. Each question provides a long context and four answer options, and performance is measured by exact accuracy on the selected option\.

## Appendix BMethod Implementation Details

This appendix specifies the interaction interface,EvidenceBankoperations, state contract, and resource settings used in the reportedLongHarnessevaluations\.

### B\.1Interaction Interface

Table 5:LongHarnessinteraction interface, showing the purpose and returned observation of each operation\.The interface distinguishes candidate discovery from source inspection, withsearchreturning candidate locations andreadreturning source text together with registered references\. Search snippets guide subsequent reading, while statements proposed to theEvidenceBankmust cite references registered byread\.

The policy selects one operation per turn, and the controller validates its structure, source references, and resource requirements before execution\. Invalid actions return an error observation without modifying the persistent states\.

### B\.2EvidenceBank Operations

Table 6:EvidenceBankoperations for admitting, correcting, and retracting source\-linked statements\.Anupsertvalidates each proposed statement independently by checking that its source references were registered by earlierreadactions, so a batch may be partially accepted\. The controller validates reference registration, while the policy judges whether the statement faithfully represents its sources\.

AcceptedEvidenceBankentries retain immutable identifiers, and corrections are recorded throughreviserather than in\-place edits\. A non\-null replacement creates a corrected active entry and marks the previous entry as superseded, while a null replacement retracts the previous entry without creating a replacement\. Superseded and retracted entries remain in the revision record but cannot be cited in compute or final requests\. Existing conclusions may retain their references until the policy updates them\.

### B\.3State Contract

TheConclusion Statecontains established claims and supported hypotheses linked to activeEvidenceBankentries, together with unresolved gaps describing information that could change those conclusions\. This representation separates the agent’s current inferences from the source\-linked statements maintained in theEvidenceBank\.

When an action updates theEvidenceBank, conclusion revision follows observation of the update result so that claims reference accepted evidence\. For other actions, the proposed conclusion update is merged before execution\. The complete interaction history remains available alongside the persistent evidence and conclusion states throughout the episode, and final supporting references must resolve to activeEvidenceBankentries and their underlying source text\. Conclusion updates replace the supplied lists of established claims, hypotheses, or open gaps\. Omitted lists remain unchanged\. When evidence becomes inactive, transition logs flag affected conclusion references, but the runtime does not automatically delete or rewrite those claims or block subsequent search and reading\. Compute and final validate the evidence IDs declared in their requests, rather than all references retained in theConclusion State\. Compute also does not establish that literal operands match the cited source values\.

### B\.4Runtime Settings

Table 7:Interaction and reading limits used in theLongHarnessevaluations\.At each step, the policy receives the system instructions and task specification, followed by the complete interaction history and the current runtime state\. The history includes all previous reasoning, actions, and tool observations\. Historical observations are retained in full rather than replaced with summaries or placeholders\. The runtime state presents theEvidenceBank,Conclusion State, and remaining resource budgets\. These explicit states provide an organized record of acquired evidence and unresolved information needs alongside the full interaction history\. The cumulative read budget counts only source text returned by thereadoperation across the episode\.

The cumulative reading budget limits source acquisition over the episode; each model request is separately subject to the endpoint’s context limit\. The client does not automatically truncate or compact the interaction history\. A request failure that persists after the model client’s retry handling, including rejection of an oversized request, is recorded asmodel\_request\_failedand terminates the episode\. Episodes that terminate without an accepted final answer are scored as incorrect\.

The same state and action contract applies across benchmarks, with OfficeQA Pro exposing the full external corpus and AA\-LCR and LongBench v2 exposing each question’s provided context as the available document collection\. Reference answers, gold source annotations, and question\-specific document labels are withheld from the agent\. All backbone models use temperature 1, top\-pp0\.95, maximum reasoning effort, and an output limit of 32,768 tokens per request\.

### B\.5Interaction Algorithm

Algorithm[1](https://arxiv.org/html/2610.13075#alg1)summarizes the interaction process, withMerge\\operatorname\{Merge\}replacing each supplied conclusion\-state list and preserving omitted lists \(Appendix[B\.3](https://arxiv.org/html/2610.13075#A2.SS3)\)\. For example, an update containing onlyopenreplaces the unresolved gaps while retaining the existingestablishedandhypotheseslists; an explicitly empty list clears that field\.

Algorithm 1LongHarnesswith Evidence–Conclusion Coevolution1:Question

qq, documents

DD, policy

πθ\\pi\_\{\\theta\}, budget

B0B\_\{0\}
2:Accepted final output, or

⊥\\botif none is accepted

3:Initialize document environment

ℰD\\mathcal\{E\}\_\{D\}, document map

MDM\_\{D\}, and state

x0x\_\{0\}
4:

E0←∅E\_\{0\}\\leftarrow\\emptyset,

C0←∅C\_\{0\}\\leftarrow\\emptyset,

H0←∅H\_\{0\}\\leftarrow\\emptyset,

t←0t\\leftarrow 0
5:while

BtB\_\{t\}permits another turn and the environment remains operationaldo

6:

St←\(Et,Ct,Ht\)S\_\{t\}\\leftarrow\(E\_\{t\},C\_\{t\},H\_\{t\}\)
7:

\(at,C^t\)∼πθ\(⋅∣q,MD,St,Bt\)\(a\_\{t\},\\widehat\{C\}\_\{t\}\)\\sim\\pi\_\{\\theta\}\(\\cdot\\mid q,M\_\{D\},S\_\{t\},B\_\{t\}\)
8:⊳\\trianglerightSelect action and optional conclusion update

9:

Ct\+1←CtC\_\{t\+1\}\\leftarrow C\_\{t\}
10:if

ata\_\{t\}violates protocol, source\-reference, or budget constraintsthen

11:

xt\+1←xtx\_\{t\+1\}\\leftarrow x\_\{t\}
12:

ot\+1←o\_\{t\+1\}\\leftarrowvalidation feedback

13:⊳\\trianglerightPreserve persistent states on invalid actions

14:else

15:if

ata\_\{t\}is neitherupsertnorrevisethen

16:

Ct\+1←Merge⁡\(Ct,C^t\)C\_\{t\+1\}\\leftarrow\\operatorname\{Merge\}\(C\_\{t\},\\widehat\{C\}\_\{t\}\)
17:⊳\\trianglerightMerge conclusion updates before other actions

18:endif

19:

\(xt\+1,ot\+1\)∼𝒫D\(⋅,⋅∣xt,at\)\(x\_\{t\+1\},o\_\{t\+1\}\)\\sim\\mathcal\{P\}\_\{D\}\(\\cdot,\\cdot\\mid x\_\{t\},a\_\{t\}\)
20:⊳\\trianglerightEvidence updates are observed before conclusion revision

21:endif

22:

Et\+1←ActiveEvidence⁡\(xt\+1\)E\_\{t\+1\}\\leftarrow\\operatorname\{ActiveEvidence\}\(x\_\{t\+1\}\)
23:Extend

HtH\_\{t\}with the turn’s reasoning, action, and observation to obtain

Ht\+1H\_\{t\+1\}
24:

Bt\+1←B\_\{t\+1\}\\leftarrowremaining budget after the turn

25:if

ot\+1o\_\{t\+1\}confirms an accepted final submissionthen

26:returnthe final output and resolved source references

27:endif

28:

t←t\+1t\\leftarrow t\+1
29:endwhile

30:return

⊥\\bot

## Appendix CExperimental Protocol Details

### C\.1Benchmark Preparation

OfficeQA Pro exposes the complete V1 or V2 corpus as an external document environment, with a document map containing identifiers, titles, and section structure\. Agents must select relevant sources from this corpus without access to question\-specific gold documents\. For OfficeQA Pro V1 and V2, we use the Parsed Documents provided by the benchmark authors\.

AA\-LCR\-1\.1 retains the document set supplied with each question, including its official grouping and document order\. Direct\-reasoning baselines receive these documents in the prompt, while agent harnesses access the same collection through search and bounded reading\. For LongBench v2, direct\-reasoning baselines follow the official truncation procedure with a 256K\-token limit, while harnesses access the original Long\-subset context through the external document interface\.

Reference answers, gold evidence annotations, and question\-specific document labels are excluded from agent\-facing inputs, and document filtering, ranking, and truncation do not use answer information or the final prediction\.

#### Document structure and retrieval\.

LongHarnessconstructs its document environment from the benchmark text, preserving source character offsets for subsequent evidence references\. Existing section annotations are retained\. Otherwise, sections are identified from Markdown headings, with a single root section used for documents without headings\. Searchable blocks follow paragraph boundaries, with oversized paragraphs split into non\-overlapping blocks of at most 1,024 estimated tokens, preferably at newline boundaries\. Token estimates count word sequences and individual non\-whitespace punctuation symbols\.

The search interface supports lexical and regular\-expression queries, optionally restricted to specified documents or sections\. Lexical queries use BM25\-style ranking over the text blocks\. Search results provide source locators and short snippets, which guide subsequentreadoperations\. Read operations return source text together with registered source\-span identifiers, allowingEvidenceBankentries to refer back to the corresponding document offsets\.

### C\.2Harness Configuration

All harnesses operate in an isolated workspace containing the question, available documents, and an output directory\. Tool\-level network access is disabled, and file access is confined to the workspace, excluding reference answers, scoring scripts, and external run metadata\.

Table 8:Available operations and resource limits for direct reasoning and the agent harnesses in the main comparison\.We use Codex CLI version 0\.135\.0 and Claude Code version 2\.1\.216\. Claude Code and Codex inspect and modify workspace files through their native tools and local shell commands\. RLM accesses documents programmatically through its Python REPL\.LongHarnessaccesses documents through its defined search and reading operations and maintains evidence and conclusion states through explicit update actions\. Table[8](https://arxiv.org/html/2610.13075#A3.T8)summarizes these interfaces and resource limits\. Each comparison uses the same questions, available documents, and backbone model, with the harness interface and execution protocol determining how the model interacts with the corpus\. RLM operates in a local Python REPL whose variables persist across iterations within an episode\. The model can inspect corpus files, process retrieved text programmatically, and invoke model subcalls through the REPL\. The task prompt specifies the question, corpus location, and required answer format\. Our benchmark adapter converts document paths to absolute paths accessible from the REPL and adds the benchmark\-specific output contract\. The root loop allows up to 30 iterations\. The driver uses the library’s default maximum recursion depth of one, at which subcalls receive plain model completions rather than launching further REPL loops\. Root and subcalls use the same backbone\. If the returned answer does not satisfy the required output format, the adapter makes one additional call to the same model, instructed to copy the answer value verbatim into the required wrapper\.

#### Evidence\-pipeline audit\.

We use GPT\-5\.6\-Luna with temperature 0 to assess incorrect episodes using the recorded reasoning and available tool outputs or evidence\-state records\. The audit distinguishes failures to retrieve the required evidence, failures to select or retain that evidence, and failures to transform the adopted evidence into a correct answer\. Answer correctness is determined by the benchmark scorer at tolerance 0; correct episodes are counted as passing the first two stages by protocol\. Retrieval coverage is computed over all evaluated questions, evidence adoption over questions passing retrieval, and answer conversion over questions passing both preceding stages\. These conditional rates characterize stage outcomes within the analyzed runs\.

### C\.3LongHarness Prompt Template

The template below presents the instructions used byLongHarnessin the terminology of the paper\. A task\-specific block supplies the question, answer requirements, document map, tool schemas, and resource budget, with the available operations defined in Appendix[B\.1](https://arxiv.org/html/2610.13075#A2.SS1)\.

#### Runtime request construction\.

The prompt template supplies the policy instructions\. Each model request additionally contains a task block specifying the question, answer requirements, document map, and action schemas\. This is followed by all previous assistant responses and tool observations in chronological order, then a current\-state block containing theEvidenceBank,Conclusion State, and remaining budgets\. TheConclusion Stateis serialized under thestatefield, withestablished,hypotheses, andopenlists\. Historical observations remain available alongside these explicit states\.

LongHarness Prompt Template[⬇](data:text/plain;base64,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)YouareaReActagentforevidence\-groundedlong\-contextquestionanswering\.Documentsremainexternaltotheprompt\.Investigatethemovermultipleturnsusingsearch,read,compute,EvidenceBank,andfinal\.Youdecidewhattheevidencemeans,whatremainsuncertain,andwhentheanswerissufficient\.Followonerevisableloop:1\.Selecttheunresolvedgapwhoseresolutionismostlikelytochangetheanswer\.2\.Usesearchtolocaterelevantdocumentsandpassages\.Searchsnippetsarenavigationaids,notcitableevidence\.3\.Usereadtoinspectsourcetextandregistersourcereferences\.OnlyasourcereferencereturnedbyreadmaysupportanEvidenceBankstatement\.4\.Saveusefulsourcefactsasatomic,source\-linkedEvidenceBankstatements\.AnEvidenceBankidentifierbecomesusableonlyafterasuccessfulEvidenceBankobservation\.5\.OnlyciteEvidenceBankidentifiersthatwereactivebeforethecurrentturn\.EvidencecreatedbythecurrentEvidenceBankactionbecomescitableonthenextturn\.6\.RevisetheConclusionStatewhennewevidencesupports,qualifies,contradicts,orreplacesearlierreasoning\.TheEvidenceBankrecordswhatthedocumentsstate,notwhetheranansweriscorrect\.Keepstatementsatomicandsource\-faithful\.Retryonlyrejectedstatements\.Ifanacceptedstatementisinvalid,useEvidenceBankrevise:provideacorrectedreplacement,oruseanullreplacementtowithdrawit\.Forcompute,placethesourcevaluesinonedeterministicexpressionandcitetheactiveEvidenceBankidentifiersthatsupportthosevalues\.Aftercompute,verifydirection,unit,andprecision\.ExpressthederivedresultintheConclusionStatetogetherwiththeEvidenceBankidentifierssupportingeveryoperand\.MaintainacompactConclusionStatewiththreerevisablelists:\-established:conclusionsjustifiedbyactiveEvidenceBankentries;\-supportedhypotheses:plausiblebutunsettledinterpretationssupportedbyactiveEvidenceBankentries;\-unresolved:concretemissinginformationthatcouldchangetheanswer\.EstablishedconclusionsandsupportedhypothesesmustciteactiveEvidenceBankentries\.Unsupportedpossibilitiesbelongunderunresolved\.Statetheinference,notmerelydisconnectedsourcefacts\.Searchforthemissingdistinction,notonlyforthequestiontopic\.Ifretrievalisuninformative,changethequeryorreadingstrategyinsteadofrepeatingit\.Readenoughcontexttointerpretthesourcecorrectly\.Beforeadoptingaconclusion,verifythatitsentities,relationships,scope,andrequestedoutputmatchthequestion\.Aderivedoraggregateconclusionmustciteeverydecision\-criticalpremiseandusecomputewhenarithmeticisneeded\.Evidencerelevancealoneisnotsupport\.Beforefinal,reconstructtherequestedansweronlyfromcurrentestablishedconclusionsandtheirsources\.Revisestaleconclusions\.Checkaplausiblealternativeonlywhenitcouldchangetheanswer\.Continueonlywhileresolvingaremainingunresolvedgapcouldchangetheresult\.Donotsearchmerelybecauseunrelateduncertaintyremains\.Placetheanswer\-bearingconclusionunderestablished\.SubmitfinalwithitsactiveEvidenceBankidentifiers\.Followtherequestedchoicelabel,units,precision,ordering,andoutputformat\.Invokeexactlyoneoperationperturn\.ConclusionStateupdatesmaybepartial,andomittedlistsretaintheirpreviousvalues\.EvidencemutationturnsdeferConclusionStateupdatesuntiltheirobservationshavebeenseen\.Usefinalastheonlywaytosubmittheanswer\.

## Appendix DRepeated\-Run Results

Table[9](https://arxiv.org/html/2610.13075#A4.T9)reports three runs for each harness and evaluation setting with GLM\-5\.3\-Flash\. Summary statistics are computed from the three reported scores\. Std\. denotes the sample standard deviation, and Max\. is the highest score across the three runs\.

Table 9:Three\-run Exact accuracy with GLM\-5\.3\-Flash\. Run scores, means, and maxima are percentages\. Standard deviations are in percentage points\.
## Appendix ECase Study

Figure[4](https://arxiv.org/html/2610.13075#S3.F4)compares GLM\-5\.3\-Flash trajectories for computing the geometric mean of 80 monthly Treasury expenditures in OfficeQA Pro V1\. We randomly selected this case from Codex trajectories that failed the exact\-match criterion\. Steps in Figure[4](https://arxiv.org/html/2610.13075#S3.F4)denote shell\-command calls for Codex and policy actions forLongHarness\. Codex locates the relevant table within 22 command calls but leaves missing values implicit, later making 103 consecutive backward reads \(*unbounded exploration*\)\. It computes 4958\.09 using 4100 for May 1942, then recognizes the corrected value 4400 but retains the earlier result in its final answer\. Replacing this operand yields the reference answer 4962\.46\. Codex therefore fails exact match but passes the 5% tolerance criterion, with a relative error of approximately 0\.088%\. This*evidence\-state drift*illustrates a recognized correction that was not incorporated into the final computation\.

After recording 4400 in theEvidenceBank,LongHarnesstracks cross\-issue consistency as an unresolvedConclusion Stategap\. Further reading corroborates the value and corrects two other entries\. Evidence and conclusions incorporate these revisions before computation\.LongHarnessreturns 4962\.46 in 10 policy actions, showing how conclusion\-guided retrieval and evidence maintenance carry corrections into the final decision\.