Screenshots or Tools? Eliciting Tool Use and Managing Multimodal Context in Hybrid GUI-MCP Computer-Use Agents

# Screenshots or Tools? Eliciting Tool Use and Managing Multimodal Context in Hybrid GUI–MCP Computer-Use Agents
Source: [https://arxiv.org/html/2608.03327](https://arxiv.org/html/2608.03327)
Siqi Fan1, Minghao Li2, Xiaoqian Ma2, Wenhui Tan3, Xiusheng Huang2, Juntong Wu4, Liujie Zhang2, Shuo Shang1, Weihang Chen2

###### Abstract

Hybrid computer\-use agents can act through screenshots or call text tools\. We find that having a tool available does not settle which way the effect goes\. Under one identical GUI–MCP harness on the OSWorld\-MCP benchmark \(309 tasks\), the same MCP tools improve a reasoning model by\+4\.0\+4\.0pp and degrade a non\-reasoning model by−5\.9\-5\.9pp \(5 runs each, both beyond22\\,SE\)\. What separates the two is tool\-decision behavior\. The non\-reasoning policy ignores, misnames, or falsely terminates around tools\. The reasoning model avoids these failures, yet still calls a tool on only 55/309 tasks, 23\.9% of the tool\-reachable ones\. We call this shortfall the*adoption gap*\. Both levels of the problem share one cause: the model already has a cheaper route and is never trained to take it\. Multi\-turn RL probes that cause\. At theaction level, a dense tool bonus raises spreadsheet adoption 0\.03→\\to0\.33 and carries into greedy decoding, but held\-out accuracy does not follow\. Behavior is steerable; competence is not\. The bottleneck lies in tool\-call*semantics*\. At thecontext level, a successful tool call often makes the next screenshot redundant\. Dropping it and halving image history cuts input tokens by about a third, at a small accuracy cost\. Retraining under the same observation rule removes that cost\. The compressed agent then reaches 37\.8% against 33\.0% for the uncompressed operating point, at 53% of the input cost, and closes the rich–lean gap on a pre\-registered degraded subset to zero\. Tools help when the model chooses and integrates them, and current hybrid agents leave many such choices unused\.

## 1Introduction

Computer\-use agents \(CUAs\)\(Anthropic[2024a](https://arxiv.org/html/2608.03327#bib.bib9); OpenAI[2025](https://arxiv.org/html/2608.03327#bib.bib10)\)have two ways to act on software\. They can drive the GUI through screenshots, clicking and typing at image coordinates\. This route is general, but costly and brittle\. Each frame consumes vision tokens, visual history grows over turns, and coordinates expire when the interface changes\. They can also call text\-level*tools*: MCP servers\(Anthropic[2024b](https://arxiv.org/html/2608.03327#bib.bib12)\), CLIs, and agent “skills”\(Jiaet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib3); Yanet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib4)\)\. Tools are cheap and precise, but they exist only for some applications and give no visual confirmation on their own\. The choice between the two routes also sets the serving cost of a deployed agent\. Screenshots dominate the token budget, so every frame kept or dropped is an economic decision as well as a behavioral one\.

Hybrid agents expose both routes\. The usual question is whether the tool set or the injection harness is good enough\. We ask a prior question\.*When a useful tool is present, does the model decide to use it?*Our answer is that availability alone does not fix the sign of the effect\. Under one identical retrieval\-injection harness on the OSWorld\-MCP benchmark\(Jiaet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib3)\)\(309 tasks\), the same MCP tools help a reasoning \(“Thinking”\) model by\+4\.0\+4\.0pp and hurt a non\-reasoning \(“Instruct”\) model by−5\.9\-5\.9pp \(Section[3\.1](https://arxiv.org/html/2608.03327#S3.SS1)\)\.

This qualifies the degradation reported by concurrent GUI\-tool work\(Huet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib17)\)\. In our setting, injection is neither universally harmful nor universally helpful\. Its sign tracks how the model handles the tool decision\. The non\-reasoning model often ignores, misnames, or falsely terminates around tools\. The reasoning model avoids those failures, but still uses tools sparingly\. Figure[1](https://arxiv.org/html/2608.03327#S1.F1)gives the high\-level picture\.

![Refer to caption](https://arxiv.org/html/2608.03327v1/x1.png)Figure 1:*Screenshots or tools?*, asked at two levels for one example task \(“make the report heading red in Writer”\)\.\(a\)*Action level*: act through pixels or call a text tool; here a single call replaces GUI steps 2–5\.\(b\)*Context level*: after a successful tool call, whose result is already in context as text, choose whether to retain the following screenshot\. Token counts are illustrative\.#### The adoption gap\.

The reversal is only the surface\. Even the model that benefits*under\-adopts*\. It calls a tool on fewer than one task in five\. On VLC, where 16/17 tasks are tool\-reachable, it never calls one at all \(Section[3\.2](https://arxiv.org/html/2608.03327#S3.SS2)\)\. The model often*can*use tools but routinely does not, and this shortfall is the central object of the paper\. For practitioners it is a quiet failure mode: the cost of building and injecting a tool server is paid in full, while most of its benefit goes unrealized\.

#### Screenshots or tools, at two levels\.

We use one recurring question to study this gap:*screenshots or tools?*It appears twice\. At theaction level, the agent chooses whether to click through pixels or call a text tool \(Sections[3\.1](https://arxiv.org/html/2608.03327#S3.SS1)–[3\.2](https://arxiv.org/html/2608.03327#S3.SS2); probed with RL in Section[5\.2](https://arxiv.org/html/2608.03327#S5.SS2)\)\. At thecontext level, once a tool has succeeded, the agent chooses whether to keep the following screenshot or rely on the textual result \(Sections[4\.1](https://arxiv.org/html/2608.03327#S4.SS1)–[4\.2](https://arxiv.org/html/2608.03327#S4.SS2); probed in Section[5\.3](https://arxiv.org/html/2608.03327#S5.SS3)\)\.

The two levels are more than an analogy\. In both, a cheaper route is already available and the policy fails to take it, because nothing in training ever asked it to: the tool is present but unused, the compressed observation affordable but unfamiliar\. Vision cannot disappear either way, since about a quarter of the tasks are tool\-unreachable and require pixels \(CAPTCHAs, slide recoloring, heavy in\-browser interaction\)\(Jiaet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib3); Brand and Burnham[2025](https://arxiv.org/html/2608.03327#bib.bib1)\)\. The question is therefore not whether to replace screenshots with tools, but when to use each\. The two probes answer differently\. The tool decision is easy to steer, but steering it adds no accuracy\. Matching the observation rule between training and inference does pay off, and makes the compressed configuration the better deployment point at half the input cost\. The next gains in hybrid agents therefore lie less in adding tools or reward terms than in training signals that teach tool semantics and in observation rules that match deployment\.

#### Contributions\.

We report a finding, a mechanism for it, and two RL probes that test the mechanism at each level\.

- •Diagnosis: conditional tool effectiveness and the adoption gap\.With the same tools and harness, MCP injection helps the reasoning model and hurts the non\-reasoning one, and the sign tracks tool\-decision behavior\. Even the model that benefits invokes a tool on fewer than a quarter of tool\-reachable tasks \(Sections[3\.1](https://arxiv.org/html/2608.03327#S3.SS1)–[3\.2](https://arxiv.org/html/2608.03327#S3.SS2)\)\.
- •Action\-level probe\.A dense post\-normalization bonus lifts adoption by an order of magnitude, and the shift survives into greedy decoding\. Held\-out accuracy does not follow, and a broad sweep does not change that\. RL reaches the tool*decision*but not tool*competence*\(Section[5\.2](https://arxiv.org/html/2608.03327#S5.SS2)\)\.
- •Context\-level probe\.After a successful tool call the next screenshot is often redundant\. Retraining under the deployment\-time observation rule converts the compression discount into a half\-cost operating point with no out\-of\-distribution accuracy loss \(Sections[4\.2](https://arxiv.org/html/2608.03327#S4.SS2)and[5\.3](https://arxiv.org/html/2608.03327#S5.SS3)\)\.

## 2Related Work

#### Computer\-use agents & benchmarks\.

OSWorld\(Xieet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib2)\)provides an execution\-based desktop testbed\. OSWorld\-MCP\([2026](https://arxiv.org/html/2608.03327#bib.bib3)\)adds verified MCP tools and reports that even strong models invoke tools on only36\.3%36\.3\\%of tasks\. That low rate is our starting point: we ask why visible tools go unused, and when using them helps\. MCPWorld\(Yanet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib4)\)studies API/GUI/hybrid evaluation, and OSWorld\-Human\(Abhyankaret al\.[2025](https://arxiv.org/html/2608.03327#bib.bib5)\)temporal efficiency\. Mind2Web\(Denget al\.[2023](https://arxiv.org/html/2608.03327#bib.bib7)\), WebArena\(Zhouet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib6)\), and AndroidWorld\(Rawleset al\.[2025](https://arxiv.org/html/2608.03327#bib.bib8)\)pose the same control problem over web and mobile interfaces\. OpenCUA\(Wanget al\.[2025b](https://arxiv.org/html/2608.03327#bib.bib19)\), ScaleCUA\(Liuet al\.[2025b](https://arxiv.org/html/2608.03327#bib.bib20)\), UI\-TARS\(Qinet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib21)\), and Agent S2\(Agasheet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib22)\)scale CUA data, models, and GUI planning systems\(Wanget al\.[2024](https://arxiv.org/html/2608.03327#bib.bib11)\)\. We use these settings to study behavior and token cost\.

WorkRouteRLAction:tool decisionContext:managementOSWorld\-MCP\([2026](https://arxiv.org/html/2608.03327#bib.bib3)\)GUI\+MCP✗✓measured✗ToolCUA\([2026](https://arxiv.org/html/2608.03327#bib.bib17)\)GUI\+tool✓✓learned✔✗windowUltraCUA\([2025](https://arxiv.org/html/2608.03327#bib.bib18)\)GUI\+prog\.✓✓learned✔✗windowComputerRL\([2025](https://arxiv.org/html/2608.03327#bib.bib29)\)GUI\+API✓✔✗implicit✔✗windowACON\([2025](https://arxiv.org/html/2608.03327#bib.bib49)\)text✗✗✓compressCAT\([2025a](https://arxiv.org/html/2608.03327#bib.bib50)\)tool calls✗✗✓callableContext\-Folding\([2025](https://arxiv.org/html/2608.03327#bib.bib51)\)tool calls✓✗✓foldingThis workGUI\+MCP✓✓diagnosed\+ probe✓matchedruleTable 1:Positioning against the closest work, along the two levels of the screenshots\-or\-tools question\. The cross\-model*sign*of tool injection and train–inference context matching are, to our knowledge, examined only here\.✗: not addressed;✔✗: passive or implicit;✓: explicitly studied or learned\.
#### GUI\-tool hybrid agents & modality\.

ToolCUA\(Huet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib17)\)is the most directly concurrent work\. It trains GUI\-tool path orchestration with heavy RFT/RL \(46\.85%46\.85\\%on OSWorld\-MCP\) and reports that naive MCP injection can hurt a GUI agent\. Our result narrows that observation\. The sign of injection tracks the base model’s tool\-decision behavior, which our RL experiment then isolates\. UltraCUA\(Yanget al\.[2025](https://arxiv.org/html/2608.03327#bib.bib18)\)likewise trains hybrid GUI/programmatic actions at scale, and the tool route itself builds on LLM tool learning\(Schicket al\.[2023](https://arxiv.org/html/2608.03327#bib.bib13); Qinet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib14); Patilet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib15)\)\. A separate line improves the screenshot side of the interface, covering grounding, GUI perception, and action modeling: CogAgent\(Honget al\.[2024](https://arxiv.org/html/2608.03327#bib.bib23)\)UGround\(Gouet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib24)\), Aguvis\(Xuet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib25)\), MP\-GUI\(Wanget al\.[2025c](https://arxiv.org/html/2608.03327#bib.bib26)\), and OS\-Atlas\(Wuet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib27)\)\.

#### RL for agents & credit assignment\.

ComputerRL\(Laiet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib29)\)scales online RL to48\.9%48\.9\\%on OSWorld with a single large run over an API–GUI hybrid action space\. DigiRL\(Baiet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib30)\)and WebRL\(Qiet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib31)\)train device\-control and web agents with online RL\. GUI\-R1, UI\-R1, and UI\-TARS\-2\(Luoet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib33); Luet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib32); Wanget al\.[2025a](https://arxiv.org/html/2608.03327#bib.bib34)\)apply RL to GUI action prediction and multi\-turn control\. Our probe targets a different question: whether an outcome\-level signal can make the policy leave the GUI path for a tool it already has\. Our advantage follows the group\-relative \(R1\-zero\) line\(Guoet al\.[2025a](https://arxiv.org/html/2608.03327#bib.bib38); Shaoet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib37); Liuet al\.[2025c](https://arxiv.org/html/2608.03327#bib.bib35); Yuet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib36)\), a critic\-free simplification of PPO\(Schulmanet al\.[2017](https://arxiv.org/html/2608.03327#bib.bib28)\), with a per\-step broadcast that keeps long trajectories from dominating the gradient\. Turn\- and segment\-level credit assignment\(Weiet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib40); Guoet al\.[2025b](https://arxiv.org/html/2608.03327#bib.bib41); Liet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib42); Fenget al\.[2025](https://arxiv.org/html/2608.03327#bib.bib43); Fuet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib39)\)remains complementary\.

#### Context management for long\-horizon agents\.

Long\-horizon agents manage bounded context with memory hierarchies\(Packeret al\.[2023](https://arxiv.org/html/2608.03327#bib.bib47)\)and prompt compression\(Jianget al\.[2023](https://arxiv.org/html/2608.03327#bib.bib48)\)\. ACON\(Kanget al\.[2025](https://arxiv.org/html/2608.03327#bib.bib49)\)compresses single\-modality*text*context\. We study the same pressure in a dual\-modality setting, where old information can live as pixels or as tool text\. CAT\(Liuet al\.[2025a](https://arxiv.org/html/2608.03327#bib.bib50)\)and Context\-Folding\(Sunet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib51)\)make context maintenance a callable or learned operation\. We instead keep the rule fixed, which isolates the train/inference mismatch before any context policy is learned\. Table[1](https://arxiv.org/html/2608.03327#S2.T1)summarizes the closest comparisons\.

## 3Action Level: Does the Model Use Its Tools?

#### Setup\.

We compare two checkpoints of the same 8B backbone, Qwen3\-VL\-8B\-Thinking \(reasoning\) and Qwen3\-VL\-8B\-Instruct \(non\-reasoning\)\(Baiet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib52)\), which differ in whether they emit an explicit reasoning trace\. The benchmark istest\_all\_no\_internet\(309 tasks\) from OSWorld\-MCP\([2026](https://arxiv.org/html/2608.03327#bib.bib3)\), built on OSWorld\(Xieet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib2)\)\. Everything else is held fixed: harness, retriever, prompt template, and tool set\. The MCP inventory spans 120 tools in 9 application namespaces, exposed through BM25\(Robertson and Zaragoza[2009](https://arxiv.org/html/2608.03327#bib.bib16)\)top\-18 retrieval with one call per step and structured error feedback\. Retrieval is keyed to the active application, so multi\-app tasks see a toolset that changes as the agent switches apps\. Qwen3\-VL emits relative coordinates on a 1000\-grid, and the harness applies the required resize \(Appendix A\)\. All evaluations use greedy decoding,max\_steps=50=50, and five repeated runs\. We call a difference significant only when\|Δ\|\>2​SE\|\\Delta\|\>2\\,\\mathrm\{SE\}\.

#### Unified hybrid action space\.

GUI actions and MCP tools are presented through the same call surface\. At each step the model emits one<tool\_call\>object: eithercomputer\_usewith one of its 11 primitive actions \(click, double\-click, drag, scroll, key, type, wait, terminate,*etc\.*\), or one of the retrieved MCP tools\. Both appear in the same<tools\>block\. No external controller routes the model\. It must choose inside one action head whether to act through pixels or through a text tool\(Yaoet al\.[2023](https://arxiv.org/html/2608.03327#bib.bib44); Schicket al\.[2023](https://arxiv.org/html/2608.03327#bib.bib13)\)\.

#### Context construction\.

A pure\-GUI agent keeps its full textual action history plus a sliding window over the lastkkscreenshots, the standard bounded\-memory recipe for CUAs\(Xieet al\.[2024](https://arxiv.org/html/2608.03327#bib.bib2); Qinet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib21)\)\. Adding MCP tools leaves this skeleton unchanged but opens a second, textual observation channel\. Formally, the step\-iiassistant outputai=\(thinki,Actioni,ci\)a\_\{i\}=\(\\texttt\{think\}\_\{i\},\\,\\text\{Action\}\_\{i\},\\,c\_\{i\}\)consists of an optional reasoning trace \(Thinking model only\), a one\-line action summary, and exactly one callcic\_\{i\}\. Each user turn isui=ρ​\(ri\)⊕o~iu\_\{i\}=\\rho\(r\_\{i\}\)\\oplus\\tilde\{o\}\_\{i\}, whererir\_\{i\}is the \(possibly empty\) result ofci−1c\_\{i\-1\}truncated toL=1500L\{=\}1500characters, ando~i\\tilde\{o\}\_\{i\}is the retained screenshot\. The context at stepttis

Ct=σ​\(𝒯t\)⊕⨁i=τtt−1\(ui⊕ai\)⏟windowed visual memory⊕ut⊕I⊕Ht⏟current turn,τt=max⁡\(1,t−k\+1\)\.\\begin\{split\}C\_\{t\}\\;=\\;\\sigma\(\\mathcal\{T\}\_\{t\}\)\\;\\oplus\\;&\\underbrace\{\\bigoplus\_\{i=\\tau\_\{t\}\}^\{t\-1\}\\\!\\bigl\(u\_\{i\}\\oplus a\_\{i\}\\bigr\)\}\_\{\\text\{windowed visual memory\}\}\\\\\[2\.0pt\] \\;\\oplus\\;&\\underbrace\{u\_\{t\}\\oplus I\\oplus H\_\{t\}\}\_\{\\text\{current turn\}\},\\qquad\\tau\_\{t\}=\\max\(1,\\,t\{\-\}k\{\+\}1\)\.\\end\{split\}\(1\)Hereσ​\(𝒯t\)\\sigma\(\\mathcal\{T\}\_\{t\}\)is the system message, holding the format rules and the BM25\-retrieved tool set𝒯t\\mathcal\{T\}\_\{t\}, andIIis the task instruction\. The text traceHt=⨁i<tActioniH\_\{t\}=\\bigoplus\_\{i<t\}\\text\{Action\}\_\{i\}keeps the one\-line summary of every prior step\(Yaoet al\.[2023](https://arxiv.org/html/2608.03327#bib.bib44); Shinnet al\.[2023](https://arxiv.org/html/2608.03327#bib.bib45)\)\. A raw screenshotoio\_\{i\}costs roughly 2K vision tokens, ando~i\\tilde\{o\}\_\{i\}is the version kept after the retention rule in Eq\.[2](https://arxiv.org/html/2608.03327#S3.E2)\(rendered prompt in Appendix C\)\. The two memories run at different timescales\. Pixels are windowed at depthkk\(defaultk=4k\{=\}4\) while text persists in full, so old pixels can drop out while their semantic trace and tool resultsrir\_\{i\}remain\. This asymmetry motivates the two knobs of Section[4](https://arxiv.org/html/2608.03327#S4)\.

ThinkingInstructAdoption†GUIGUI\+MCP, by context policyGUIGUI\+MCP, by context policyThinkInstDomainwin\. 4win\. 4\(op\.\)win\. 4\+\+dropwin\. 2win\. 2\+\+drop\(ctx\_opt\)win\. 4win\. 4win\. 4\+\+dropwin\. 2win\. 2\+\+dropcalc \(spreadsheet\)17\.918\.314\.912\.812\.317\.07\.25\.53\.84\.721%0%writer \(document\)51\.345\.238\.337\.438\.329\.914\.813\.95\.29\.652%22%impress \(slides\)28\.526\.820\.921\.720\.018\.515\.317\.916\.610\.640%40%vs\_code57\.168\.664\.862\.964\.849\.538\.139\.043\.845\.714%19%os51\.757\.557\.553\.354\.241\.745\.043\.346\.747\.58%4%multi\_apps8\.314\.914\.113\.613\.67\.09\.910\.99\.69\.111%4%*Zero\-adoption domains \(accuracy changes not tool\-attributable\):*gimp60\.867\.770\.069\.266\.957\.734\.638\.550\.846\.90%0%thunderbird6\.710\.712\.08\.012\.06\.74\.04\.06\.76\.70%0%vlc24\.740\.034\.131\.831\.827\.116\.525\.925\.921\.20%0%chrome64\.368\.674\.364\.364\.362\.954\.361\.444\.344\.30%0%all \(309\)30\.534\.532\.330\.630\.625\.419\.520\.920\.519\.517\.8%10\.4%*Token side \(per task, 5\-run means\):*input \(K\)313\.1337\.1342\.4226\.1219\.5287\.3316\.4310\.1232\.4231\.1——peak \(p95\)1138511544114877314724310966114411143772547231——tok / 1%acc \(K\)10\.610\.111\.07\.87\.711\.416\.314\.911\.512\.0——Table 2:The static results grid: accuracy \(%, per domain and overall\) and token cost across the action space \(GUI vs\. GUI\+MCP\) and the context policies, named as in Figure[1](https://arxiv.org/html/2608.03327#S1.F1)b \(win\.kk==sliding window over the lastkkscreenshots;\+\+drop==drop\-on\-success, Eq\.[2](https://arxiv.org/html/2608.03327#S3.E2)\); 5\-run means\. “op\.” marks the RL operating point and “ctx\_opt” the compression setting of Section[4\.2](https://arxiv.org/html/2608.03327#S4.SS2)\.†task\-level tool invocation, measured at the operating point\. Tok/1%acc counts input\+\+output\.
#### Retention rule \(the second knob\)\.

The retained frameo~i\\tilde\{o\}\_\{i\}realizes the screenshot\-retention decision:

o~i=\{πifdropis on and​succ​\_​mcp​\(ci−1\),oiotherwise,\\tilde\{o\}\_\{i\}\\;=\\;\\begin\{cases\}\\pi&\\text\{if \{drop\} is on and \}\\mathrm\{succ\\\_mcp\}\(c\_\{i\-1\}\),\\\\\[2\.0pt\] o\_\{i\}&\\text\{otherwise,\}\\end\{cases\}\(2\)whereπ\\piis a short text placeholder\. Heresucc​\_​mcp\\mathrm\{succ\\\_mcp\}denotes*execution\-level*success: the call parsed, dispatched, and returned without error\. Semantic success is a separate matter, and the distinction returns in Section[4\.2](https://arxiv.org/html/2608.03327#S4.SS2)\. Image\-history depthkkand this drop rule are the two context knobs we vary below\. Both act only on the visual channel and leave the text traceHtH\_\{t\}intact\.

### 3\.1The Sign Reversal

#### Overall result\.

With everything except the base model held fixed, MCP injection lifts the reasoning model and drops the non\-reasoning one \(Table[2](https://arxiv.org/html/2608.03327#S3.T2), all\-309 row; both deltas beyond2​SE2\\,\\mathrm\{SE\}\)\. The best single Thinking run reached 37\.9%; we report five\-run means throughout\.

#### Per\-domain decomposition\.

The same pattern holds per domain \(Table[2](https://arxiv.org/html/2608.03327#S3.T2)\)\. Among the six domains with nonzero adoption, tools help the reasoning model on four and hurt the non\-reasoning model on four\. We exclude the zero\-adoption block from this count, since with no tool calls its changes reflect only prompt perturbation and run\-to\-run variance\. Adoption alone is not enough, though\. The reasoning model’s two losses are its highest\-adoption domains\. Writer is the sharpest case: half of its tasks invoke a tool, yet invoked tasks succeed far less often than non\-invoked ones, a mix of difficulty self\-selection and mis\-parameterized calls \(Appendix E\)\. The non\-reasoning model is hurt most precisely where it nominally adopts\.

DiagnosticThinkingInstruct*Supply→\\toadoption:*Tool\-reachable \(supply\)230/309 \(79 vision\-only\)Adoption, task\-level17\.8% \(55/309\)10\.4% \(32/309\)Adoption, reachable \(230\)23\.9%13\.9%TIRreal\{\}\_\{\\text\{real\}\}\(MCP/steps\)2\.8%2\.0%*Failure modes:*Hallucinated MCP steps\+0\.0\+0\.0pp\+0\.4\+0\.4ppFalse\-success rate21\.7% \(67/309\)33\.0% \(102/309\)Hallucinated tool names097 \(2 tasks\)Table 3:Behavioral diagnostics at the operating point \(GUI\+MCP, window\-4; 5\-run means\)\. Instruct adopts fewer tools yet false\-terminates*more*, ruling out a “cleaner SFT prior” explanation\. TIR==MCP steps / total steps\. Token costs: Table[2](https://arxiv.org/html/2608.03327#S3.T2), token block\.
#### What differs between the two models\.

The reversal tracks how the two models handle tools \(Table[3](https://arxiv.org/html/2608.03327#S3.T3)\)\. The non\-reasoning model ignores the spreadsheet tools entirely, hallucinates tool names, and false\-terminates*more*often, all while producing66–7×7\\timesshorter outputs\. The hallucinations concentrate in two tasks, and fixing them would barely move the total, so they are a symptom and not the cause\. The pattern instead suggests that without an explicit deliberation trace\(Weiet al\.[2022](https://arxiv.org/html/2608.03327#bib.bib46)\), the model never takes the step of asking whether a tool should be used\. Correctly injected tools are then ignored, misnamed, or hidden behind premature success\. The two checkpoints differ in more than that trace, so we read this as association and not mechanism\.

### 3\.2The Adoption Gap

Even the reasoning model leaves most tools unused\. It invokes a tool on fewer than one task in five overall, and on fewer than one in four where a tool is actually reachable \(Table[3](https://arxiv.org/html/2608.03327#S3.T3)\)\. Removing the capability ceiling does not remove the behavior gap\.

The two losses separate cleanly \(Table[3](https://arxiv.org/html/2608.03327#S3.T3), supply block\)\. About a quarter of the tasks are tool\-unreachable: three apps expose no MCP tools at all, and some tasks in tool\-equipped apps have no tool that applies\. Everywhere else the tools are present and injected, yet mostly unused\. VLC is the extreme case, where nearly every task is tool\-reachable and neither model ever calls one\. Section[5\.2](https://arxiv.org/html/2608.03327#S5.SS2)targets this purely behavioral gap\.

## 4Context Level: What Does Hybrid Cost?

Section[3](https://arxiv.org/html/2608.03327#S3)asked whether the agent takes the tool route; this section asks what each route costs\. Screenshots dominate the token budget, so the window depthkkand the drop rule of Eq\.[2](https://arxiv.org/html/2608.03327#S3.E2)set the serving price of a hybrid agent\.

### 4\.1Accuracy–Token Frontier and Operating Points

Image\-history depth:*window\-4*keeps the last four screenshots,*window\-2*the last two\. Post\-tool retention: by default the next screenshot is kept, while*drop*\(drop\-on\-success, Eq\.[2](https://arxiv.org/html/2608.03327#S3.E2)\) replaces it with a text placeholder\. This gives five operating points: GUI\-only; window\-4, our accuracy\-oriented hybrid baseline; window\-2; window\-4\+\+drop; and window\-2\+\+drop, the token\-efficient “ctx\_opt” setting\.

The two knobs do different jobs \(Table[2](https://arxiv.org/html/2608.03327#S3.T2), token block\)\. Window depth is the main token lever\. Window\-2 cuts cumulative input by about a third and peak context by nearly 40%, but it also causes the only accuracy loss beyond2​SE2\\,\\mathrm\{SE\}\. The drop rule is nearly free in accuracy at both depths, because it removes a frame whose tool result is already in text\. On its own it is not a token lever, however\. At window\-4 the slightly longer completions offset the per\-frame saving, so its benefit appears only alongside the shorter window\. Together, window\-2\+\+drop is the token\-efficiency knee\.

We keepwindow\-4\(no drop\) as the RL operating point andwindow\-2\+\+dropas the compression studied in Section[4\.2](https://arxiv.org/html/2608.03327#S4.SS2)\.

### 4\.2Inference\-Only Compression

Applied only at inference, the token\-efficient setting of Section[4\.1](https://arxiv.org/html/2608.03327#S4.SS1)\(window\-2\+\+drop, “ctx\_opt”\) costs−3\.9\-3\.9pp \(±1\.0\\pm 1\.0\)\. Paired per\-task analysis points to*diffuse degradation*rather than lost capability\. Only 3/309 tasks flip hard under compression, while a pre\-registered*degraded subset*D13 \(13 tasks, with a 12\.8pp rich–lean gap under matched greedy anchors\) concentrates the effect\. Our hypothesis is therefore mis\-adaptation, not incapability, and not simply a horizon\-budget issue\(Kanget al\.[2025](https://arxiv.org/html/2608.03327#bib.bib49)\)\. The policy is asked to act on an observation distribution it never saw during training\. If that is right, the loss should be recoverable by making rollouts, evaluation, and deployment share one observation policy, which Section[5\.3](https://arxiv.org/html/2608.03327#S5.SS3)tests\.

#### Scope\.

succ​\_​mcp\\mathrm\{succ\\\_mcp\}is execution\-level, so a call can succeed mechanically yet fail semantically\. A find\-and\-replace may return success with zero replacements, and the drop rule then discards the only visual evidence of that failure\. Section[5\.2](https://arxiv.org/html/2608.03327#S5.SS2)shows this is common on parameter\-heavy tools\. Part of the residual drop cost is therefore lost error\-correction signal, not merely a dropped duplicate\.

## 5Multi\-Turn RL: Steering Tool Use and Matching the Observation Rule

RL lets us probe both levels directly\. At the action level, can an outcome\-level signal make the policy use tools it already has? At the context level, can matched training recover the compression penalty? All runs use the verified\-clean pipeline of Appendix A and are on\-policy by construction \(per\-windowclip\_frac=0=0, ratio=1\.0=1\.0\)\.

### 5\.1Setup

#### Multi\-turn GRPO\.

For each taskxxwe roll outG=8G\{=\}8trajectories at temperature1\.01\.0across 96 parallel environments\. All 74 curated training tasks \(below\) are rolled out at every step, with horizonTmax=max\_steps=50T\_\{\\max\}\{=\}\\texttt\{max\\\_steps\}\{=\}50\. Evaluation uses greedy decoding\. A trajectoryτ\\tauwithTTsteps and terminal outcomesucc​\(τ\)∈\{\+1,−1\}\\mathrm\{succ\}\(\\tau\)\\in\\\{\+1,\-1\\\}receives the return

R​\(τ\)=succ​\(τ\)−λlen​TTmax−λcap​1​\[T≥Tmax\],R\(\\tau\)\\;=\\;\\mathrm\{succ\}\(\\tau\)\\;\-\\;\\lambda\_\{\\text\{len\}\}\\frac\{T\}\{T\_\{\\max\}\}\\;\-\\;\\lambda\_\{\\text\{cap\}\}\\,\\mathbb\{1\}\\\!\\left\[T\\geq T\_\{\\max\}\\right\],\(3\)with tie\-breaker coefficientsλlen=0\.05\\lambda\_\{\\text\{len\}\}\{=\}0\.05andλcap=0\.2\\lambda\_\{\\text\{cap\}\}\{=\}0\.2; the±1\\pm 1outcome term dominates\. Returns arezz\-scored within the group \(μx,σx\\mu\_\{x\},\\sigma\_\{x\}\) and broadcast uniformly to steps\. This is a Dr\.GRPO\-style length debiasing\(Liuet al\.[2025c](https://arxiv.org/html/2608.03327#bib.bib35)\)that keeps long trajectories from dominating the gradient\. Each step then becomes one training sample, with promptCtC\_\{t\}\(Eq\.[1](https://arxiv.org/html/2608.03327#S3.E1)\) and responseata\_\{t\}:

A^t=1T⋅R​\(τ\)−μxσx\+λmcp​bt\.\\hat\{A\}\_\{t\}\\;=\\;\\frac\{1\}\{T\}\\cdot\\frac\{R\(\\tau\)\-\\mu\_\{x\}\}\{\\sigma\_\{x\}\}\\;\+\\;\\lambda\_\{\\text\{mcp\}\}\\,b\_\{t\}\.\(4\)

#### A dense tool bonus that survives normalization\.

The bonusbt∈\{0,1\}b\_\{t\}\\in\\\{0,1\\\}\(used in Section[5\.2](https://arxiv.org/html/2608.03327#S5.SS2)\) is added*after*normalization\. Placed insideR​\(τ\)R\(\\tau\)it would be diluted to∼10−4\{\\sim\}10^\{\-4\}by trajectory averaging,zz\-scoring, and the1/T1/Tbroadcast, which is empirically a dead signal\. It fires only for an execution\-successful, non\-read\-only call whose \(tool, arguments\) key has not appeared earlier in the trajectory\. Firing once per key per trajectory, and not gating on task success, prevents reward farming by repeated or side\-effect\-free calls\. Withλmcp=0\.1\\lambda\_\{\\text\{mcp\}\}\{=\}0\.1the bonus is louder than the primary signal at the steps where it fires, since the outcome\-derived part ofA^t\\hat\{A\}\_\{t\}has magnitude∼0\.07\{\\sim\}0\.07after normalization\.

#### Optimization\.

We use GRPO with a KL penalty to the*rollout\-time*policy \(β=0\.02\\beta\{=\}0\.02; hyperparameters in Appendix F\)\. Anchoring to the base model instead cancels the bonus, because the per\-step drift it induces is smaller than the pull back to base\. Groups with mean success outside\(0\.05,0\.95\)\(0\.05,0\.95\)are dropped, so every kept group mixes successes and failures\. Each configuration is trained once; reported accuracies are greedy probes repeated three times\.

#### Data split and gradient\-band curation\.

The eight trained apps \(calc, writer, impress, vs\_code, os, gimp, vlc, thunderbird\) contribute a 172\-task training pool and 48 held\-out tasks; chrome and multi\_apps \(89 tasks\) are never trained on, a true out\-of\-distribution bucket\. From this pool we curate the actual training set: a*gradient band*of 74 tasks with empirical pass ratep∈\(0\.1,0\.9\)p\\in\(0\.1,0\.9\)under the sameG=8G\{=\}8temperature\-1\.0 rollouts\. Groups outside the band have zero within\-group variance and contribute zero gradient, so pool tasks outside the band are never rolled out\. The band is profiled once before launch and held fixed for all runs reported here, which leaves 235 tasks never trained on\.

#### Outcome\-only RL\.

Withλmcp=0\\lambda\_\{\\text\{mcp\}\}\{=\}0, no swept configuration moves held\-out or out\-of\-distribution accuracy\. In\-distribution accuracy does rise \(Section[5\.3](https://arxiv.org/html/2608.03327#S5.SS3)\); nothing transfers\. The sweep covers learning rate, KL strength and anchor, normalization scheme, task density, and horizon \(Appendix F\)\. The signal is simply too sparse: a long\-horizon task yields a single±1\\pm 1over up to 50 steps, and behaviors the base policy never samples, above all successful calls to the harder tools, receive no gradient under R1\-zero\-style RL\. Any movement reported below therefore comes from the dense bonus \(Section[5\.2](https://arxiv.org/html/2608.03327#S5.SS2)\) or from training on the gradient band itself\. A matched*rich\-observation*control \(identical recipe, compression off, evaluated at the same 30\-step checkpoint\) settles the attribution in Section[5\.3](https://arxiv.org/html/2608.03327#S5.SS3)\.

### 5\.2Result 1: A Dense Tool Bonus Moves Adoption, Not Competence

#### The tool decision is fully steerable\.

With the dense bonus, applied on a 24\-task subset of the gradient band, spreadsheet adoption rises from 0\.03 to 0\.33 within 23 training steps\. It transfers to greedy decoding \(0\.02→\\to0\.29\), so the change is learned policy and not sampling noise, and step\-level usage rises 4\.7×\\times\. One post\-normalization reward term is enough to change the decision that Section[3\.2](https://arxiv.org/html/2608.03327#S3.SS2)showed models rarely make\.

#### Held\-out accuracy localizes the bottleneck\.

Across the 48 held\-out tasks the run produces zero sustained fail→\\topass flips, and on the seven held\-out spreadsheet tool tasks accuracy stays at the base level throughout \(Figure[2](https://arxiv.org/html/2608.03327#S5.F2)\)\. The probe therefore separates the tool*decision*, which RL controls, from tool\-call*competence*, which it leaves untouched\.

![Refer to caption](https://arxiv.org/html/2608.03327v1/x2.png)Figure 2:Adoption–competence decoupling: the dense tool bonus drives adoption into the deterministic policy \(top\) while held\-out accuracy stays at the base level \(bottom\); dashed connector: visual guide, not per\-step data\.
#### Diagnosis: calls execute but fail semantically\.

Tool calls execute reliably \(98–100% API success\), but on parameter\-heavy tools the*semantic*success rate is zero: 0/23 for regex find\-and\-replace and 0/16 for format conversion\. The server reportssuccess:trueon zero\-effect calls, such as a regex that matches nothing, which feeds a false\-success prior\. This is the same execution\-versus\-semantics gap that limits the drop rule \(Section[4\.2](https://arxiv.org/html/2608.03327#S4.SS2)\)\. Within\-task matched comparisons show no win\-rate difference with tools \(−3\.1\-3\.1pp, n\.s\.\)\. Three mechanisms give the same result: an outcome\-independent bonus \(RL\), positive\-advantage cloning \(RL\), and tool documentation injected at inference time \(prompting\) all raise adoption substantially \(hint\-targeted calls rise3×3\\times\), and none changes accuracy\. The bottleneck is tool\-call semantics, which makes it a data problem and not a reward\-design problem\.

### 5\.3Result 2: Keeping the Token Savings Without the Accuracy Loss

![Refer to caption](https://arxiv.org/html/2608.03327v1/x3.png)Figure 3:Final deployment plane: RL checkpoints under both observation policies \(greedy×\\times3 anchors; point labels are training steps; star==the step\-40 deployment pick\) on\(a\)the full suite and\(b\)the 235 tasks outside the RL training set\. The step\-50 compressed point is omitted \(token telemetry unavailable\)\.![Refer to caption](https://arxiv.org/html/2608.03327v1/x4.png)Figure 4:Matched\-training recovery\.\(a\)Training reward against the rich\-observation control \(paused at step 30 after probing\)\.\(b\)D13 under both observation policies \(shaded: the rich–lean gap\)\.#### Design\.

We reuse the recipe of Section[5\.1](https://arxiv.org/html/2608.03327#S5.SS1)withλmcp=0\\lambda\_\{\\text\{mcp\}\}\{=\}0\. The*only*change is the observation policy\. The compression rule of Section[4\.2](https://arxiv.org/html/2608.03327#S4.SS2)is now active in both rollout and evaluation, logged and replayed so that train and inference share it exactly\. Two judgment criteria were pre\-registered before launch: the degraded subset D13 \(the 13 tasks where inference\-only compression concentrates its loss\), and a difference\-in\-differences \(DiD\) criterion requiring the compressed\-side gain to exceed the rich\-side gain by≥\\geq15pp, which subtracts memorization common to both sides\.

#### Result\.

The training reward rises 0\.52→\\to0\.667 \(peak, step 41\)\. What matters is not the rise itself, since in\-distribution gains can contain memorization, but what it is made of\. Checkpoints at steps 20–50 are probed on all 309 tasks under*both*observation policies against same\-period base anchors \(greedy×\\times3, each within the operating point’s 5\-run spread\):

- •Deployment headline \(Figure[3](https://arxiv.org/html/2608.03327#S5.F3)\)\.Matched training removes the price of compression\. The step\-40 compressed checkpoint runs at 53% of the input cost, with−37\-37% peak context, and loses nothing outside the training set \(\+0\.8\+0\.8pp, n\.s\., on the 235 non\-training tasks\): equal accuracy at half the cost\. On the full suite it reaches 37\.8% against 33\.0% for the uncompressed base operating point,\+4\.8\+4\.8pp \(step 50 holds\+3\.9\+3\.9pp\), but∼4\.1\{\\sim\}4\.1pp of that margin comes from the 74 training tasks \(55\.8%→\\,\\to\\,72\.9%\), so we read the full\-suite number as in\-distribution\. The optimum is also an*early*checkpoint, not the endpoint\.
- •Degraded subset D13\.The rich–lean gap collapses to 0 at step 30 and stays closed within noise \(3\.83\.8–5\.15\.1pp; Figure[4](https://arxiv.org/html/2608.03327#S5.F4)b\), with both sides ending far above base \(compressed 51\.3%→\\,\\to\\,82\.1%, rich 64\.1%→\\,\\to\\,87\.2%\)\. Of the two pre\-registered criteria, the*direction*holds and the*magnitude*does not\. The DiD turns positive at step 30 \(\+12\.8\+12\.8pp\) and stays positive thereafter \(\+7\.7\+7\.7–\+9\.0\+9\.0pp,±∼11\\pm\{\\sim\}11pp,n=13n\{=\}13\), but the\+15\+15pp bar is met at no checkpoint\. That bar was calibrated against single\-rep anchors which, we now know, overstated the base gap; we record it as failed rather than restate it at a revised threshold\.
- •Late training favors the rich side\.After the compressed side adapts at step 30, further training helps mainly under rich observation \(on training tasks,\+10\.9\+10\.9pp from step 30 to 50 versus flat compressed\), hence the*early*step\-40 optimum\.

#### Scope and attribution\.

The gains above are in\-distribution\. Held\-out accuracy under compression is unchanged between the base and the final checkpoint\. The rich\-observation control settles where the training\-side rise comes from\. At the same 30\-step checkpoint it gains nearly twice as much on the training band \(\+20\.5\+20\.5pp vs\.\+11\.5\+11\.5pp; on D13,\+12\.8\+12\.8pp vs\.\+5\.1\+5\.1pp\)\. The rise is therefore a property of the optimization recipe and not of consistency training, and compression is if anything a training\-time handicap\. Both the deployment result and the D13 dynamics are measured within the compressed run, so neither depends on the control\. The control’s faster learning is also not free: it reaches 41\.0% at step 40, but emits 42\.6K output tokens per task against the compressed run’s 10\.7K and keeps the full uncompressed input cost, about1\.6×1\.6\\timesthe compressed total\. The two runs therefore sit at non\-comparable points of the cost–accuracy frontier\. Near\-lossless is also not lossless: spreadsheet and slide tasks can still need visual confirmation\.

## 6Conclusion

Hybrid GUI–MCP agents face the same question twice:*screenshots or tools?*At the action level, tools help only when the model decides to use and integrate them\. The same injection helps the reasoning model and hurts the non\-reasoning one, RL steers the decision freely, and tool\-call semantics remains the open problem\. At the context level, dropping redundant post\-tool screenshots pays off only when training and inference share one observation rule; matched retraining then halves input cost with no out\-of\-distribution accuracy loss\. Both levels have one shape: a cheaper route exists and nothing in training teaches the policy to take it\. Tool availability is not enough, and current hybrid agents leave many usable tools on the table\.

#### Limitations and future work\.

The sign reversal is bounded: within one backbone, the sign of MCP injection tracks tool\-decision behavior, but we do not show that reasoning capability sets that sign in general\. The two checkpoints differ in more than their reasoning trace, and a within\-model thinking toggle would settle the mechanism\. The RL probe identifies the competence bottleneck but does not close it\. Supervised injection of verified tool trajectories from stronger teachers is the next step, with turn\- and segment\-level credit assignment\(Weiet al\.[2025](https://arxiv.org/html/2608.03327#bib.bib40); Guoet al\.[2025b](https://arxiv.org/html/2608.03327#bib.bib41); Liet al\.[2026](https://arxiv.org/html/2608.03327#bib.bib42)\)complementary\. The compression\-recovery gain does not transfer to held\-out tasks, and the in\-distribution rise is a recipe effect \(Section[5\.3](https://arxiv.org/html/2608.03327#S5.SS3)\)\. The context rule is fixed rather than learned\(Liuet al\.[2025a](https://arxiv.org/html/2608.03327#bib.bib50)\), and cross\-OS scaling\(Liuet al\.[2025b](https://arxiv.org/html/2608.03327#bib.bib20)\)is left to future work\.

## References

- OSWorld\-Human: benchmarking the efficiency of computer\-use agents\.arXiv preprint arXiv:2506\.16042\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Agashe, K\. Wong, V\. Tu, J\. Yang, A\. Li, and X\. E\. Wang \(2025\)Agent S2: a compositional generalist\-specialist framework for computer use agents\.arXiv preprint arXiv:2504\.00906\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- Anthropic \(2024a\)Introducing computer use, a new Claude 3\.5 Sonnet, and Claude 3\.5 Haiku\.Note:https://www\.anthropic\.com/news/3\-5\-models\-and\-computer\-useAccessed July 2026Cited by:[§1](https://arxiv.org/html/2608.03327#S1.p1.1)\.
- Anthropic \(2024b\)Introducing the Model Context Protocol\.Note:https://www\.anthropic\.com/news/model\-context\-protocolSpecification athttps://modelcontextprotocol\.io; accessed July 2026Cited by:[§1](https://arxiv.org/html/2608.03327#S1.p1.1)\.
- H\. Bai, Y\. Zhou, M\. Cemri, J\. Pan, A\. Suhr, S\. Levine, and A\. Kumar \(2024\)DigiRL: training in\-the\-wild device\-control agents with autonomous reinforcement learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2406\.11896Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen,et al\.\(2025\)Qwen3\-VL technical report\.arXiv preprint arXiv:2511\.21631\.Cited by:[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px1.p1.2)\.
- F\. Brand and G\. Burnham \(2025\)What does OSWorld tell us about AI’s ability to use computers?\.Note:Epoch AI Blog,https://epoch\.ai/blog/what\-does\-osworld\-tell\-us\-about\-ais\-ability\-to\-use\-computersPublished October 30, 2025; accessed July 2026Cited by:[§1](https://arxiv.org/html/2608.03327#S1.SS0.SSS0.Px2.p2.1)\.
- X\. Deng, Y\. Gu, B\. Zheng, S\. Chen, S\. Stevens, B\. Wang, H\. Sun, and Y\. Su \(2023\)Mind2Web: towards a generalist agent for the web\.InAdvances in Neural Information Processing Systems \(NeurIPS\) Datasets and Benchmarks Track,Note:arXiv:2306\.06070Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- L\. Feng, Z\. Xue, T\. Liu, and B\. An \(2025\)Group\-in\-group policy optimization for LLM agent training\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2505\.10978Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- W\. Fu, J\. Gao, X\. Shen, C\. Zhu, Z\. Mei,et al\.\(2025\)AReaL: a large\-scale asynchronous reinforcement learning system for language reasoning\.arXiv preprint arXiv:2505\.24298\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- B\. Gou, R\. Wang, B\. Zheng, Y\. Xie, C\. Chang, Y\. Shu, H\. Sun, and Y\. Su \(2025\)Navigating the digital world as humans do: universal visual grounding for GUI agents\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2410\.05243Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu,et al\.\(2025a\)DeepSeek\-R1 incentivizes reasoning in LLMs through reinforcement learning\.Nature645,pp\. 633–638\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- Y\. Guo, L\. Xu, J\. Liu, D\. Ye, and S\. Qiu \(2025b\)Segment policy optimization: effective segment\-level credit assignment in RL for large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2505\.23564Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1),[§6](https://arxiv.org/html/2608.03327#S6.SS0.SSS0.Px1.p1.1)\.
- W\. Hong, W\. Wang, Q\. Lv, J\. Xu, W\. Yu, J\. Ji, Y\. Wang, Z\. Wang,et al\.\(2024\)CogAgent: a visual language model for GUI agents\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Note:arXiv:2312\.08914Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- X\. Hu, X\. Zhang, H\. Xu, K\. Qiao, J\. Yang, X\. Huang, J\. Shao, M\. Yan, and J\. Ye \(2026\)ToolCUA: towards optimal GUI\-tool path orchestration for computer use agents\.arXiv preprint arXiv:2605\.12481\.Cited by:[§1](https://arxiv.org/html/2608.03327#S1.p3.1),[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.3.1)\.
- H\. Jia, J\. Liao, X\. Zhang, H\. Xu, T\. Xie, C\. Jiang, M\. Yan, S\. Liu,et al\.\(2026\)OSWorld\-MCP: benchmarking MCP tool invocation in computer\-use agents\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2510\.24563Cited by:[§1](https://arxiv.org/html/2608.03327#S1.SS0.SSS0.Px2.p2.1),[§1](https://arxiv.org/html/2608.03327#S1.p1.1),[§1](https://arxiv.org/html/2608.03327#S1.p2.2),[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.2.1),[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px1.p1.2)\.
- H\. Jiang, Q\. Wu, C\. Lin, Y\. Yang, and L\. Qiu \(2023\)LLMLingua: compressing prompts for accelerated inference of large language models\.InProceedings of the Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Note:arXiv:2310\.05736Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px4.p1.1)\.
- M\. Kang, W\. Chen, D\. Han, H\. A\. Inan, L\. Wutschitz, Y\. Chen, R\. Sim, and S\. Rajmohan \(2025\)ACON: optimizing context compression for long\-horizon LLM agents\.arXiv preprint arXiv:2510\.00615\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px4.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.6.1),[§4\.2](https://arxiv.org/html/2608.03327#S4.SS2.p1.3)\.
- H\. Lai, X\. Liu, Y\. Zhao, H\. Xu, H\. Zhang, B\. Jing, Y\. Ren, S\. Yao, Y\. Dong, and J\. Tang \(2025\)ComputerRL: scaling end\-to\-end online reinforcement learning for computer use agents\.arXiv preprint arXiv:2508\.14040\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.5.1)\.
- J\. Li, P\. Zhou, R\. Meng, M\. P\. Vadera, L\. Li, and Y\. Li \(2026\)Turn\-PPO: turn\-level advantage estimation with PPO for improved multi\-turn RL in agentic LLMs\.InProceedings of the Conference of the European Chapter of the Association for Computational Linguistics \(EACL\),Note:arXiv:2512\.17008Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1),[§6](https://arxiv.org/html/2608.03327#S6.SS0.SSS0.Px1.p1.1)\.
- S\. Liu, J\. Yang, B\. Jiang, Y\. Li, J\. Guo, X\. Liu, and B\. Dai \(2025a\)Context as a tool: context management for long\-horizon SWE\-agents\.arXiv preprint arXiv:2512\.22087\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px4.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.7.1),[§6](https://arxiv.org/html/2608.03327#S6.SS0.SSS0.Px1.p1.1)\.
- Z\. Liu, J\. Xie, Z\. Ding, Z\. Li, B\. Yang, Z\. Wu, X\. Wang, Q\. Sun,et al\.\(2025b\)ScaleCUA: scaling open\-source computer use agents with cross\-platform data\.arXiv preprint arXiv:2509\.15221\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2608.03327#S6.SS0.SSS0.Px1.p1.1)\.
- Z\. Liu, C\. Chen, W\. Li, P\. Qi, T\. Pang, C\. Du, W\. S\. Lee, and M\. Lin \(2025c\)Understanding R1\-zero\-like training: a critical perspective\.arXiv preprint arXiv:2503\.20783\.Note:Dr\.GRPOCited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1),[§5\.1](https://arxiv.org/html/2608.03327#S5.SS1.SSS0.Px1.p1.14)\.
- Z\. Lu, Y\. Chai, Y\. Guo, X\. Yin, L\. Liu, H\. Wang, H\. Xiao, S\. Ren,et al\.\(2025\)UI\-R1: enhancing efficient action prediction of GUI agents by reinforcement learning\.arXiv preprint arXiv:2503\.21620\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- R\. Luo, L\. Wang, W\. He, L\. Chen, J\. Li, and X\. Xia \(2025\)GUI\-R1: a generalist R1\-style vision\-language action model for GUI agents\.arXiv preprint arXiv:2504\.10458\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- OpenAI \(2025\)Computer\-using agent: a universal interface for AI to interact with the digital world\.Note:https://openai\.com/index/computer\-using\-agent/Accessed July 2026Cited by:[§1](https://arxiv.org/html/2608.03327#S1.p1.1)\.
- C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez \(2023\)MemGPT: towards LLMs as operating systems\.arXiv preprint arXiv:2310\.08560\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px4.p1.1)\.
- S\. G\. Patil, T\. Zhang, X\. Wang, and J\. E\. Gonzalez \(2024\)Gorilla: large language model connected with massive APIs\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2305\.15334Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Qi, X\. Liu, I\. L\. Iong, H\. Lai, X\. Sun, W\. Zhao, Y\. Yang, X\. Yang,et al\.\(2025\)WebRL: training LLM web agents via self\-evolving online curriculum reinforcement learning\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2411\.02337Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- Y\. Qin, S\. Liang, Y\. Ye, K\. Zhu, L\. Yan, Y\. Lu, Y\. Lin, X\. Cong,et al\.\(2024\)ToolLLM: facilitating large language models to master 16000\+ real\-world APIs\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2307\.16789Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Qin, Y\. Ye, J\. Fang, H\. Wang, S\. Liang, S\. Tian, J\. Zhang, J\. Li,et al\.\(2025\)UI\-TARS: pioneering automated GUI interaction with native agents\.arXiv preprint arXiv:2501\.12326\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px3.p1.10)\.
- C\. Rawles, S\. Clinckemaillie, Y\. Chang, J\. Waltz, G\. Lau, M\. Fair, A\. Li, W\. E\. Bishop,et al\.\(2025\)AndroidWorld: a dynamic benchmarking environment for autonomous agents\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2405\.14573Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- S\. E\. Robertson and H\. Zaragoza \(2009\)The probabilistic relevance framework: BM25 and beyond\.Foundations and Trends in Information Retrieval3\(4\),pp\. 333–389\.Cited by:[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px1.p1.2)\.
- T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom \(2023\)Toolformer: language models can teach themselves to use tools\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2302\.04761Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px2.p1.1)\.
- J\. Schulman, F\. Wolski, P\. Dhariwal, A\. Radford, and O\. Klimov \(2017\)Proximal policy optimization algorithms\.arXiv preprint arXiv:1707\.06347\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song,et al\.\(2024\)DeepSeekMath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Note:GRPOCited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2303\.11366Cited by:[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px3.p1.19)\.
- W\. Sun, M\. Lu, Z\. Ling, K\. Liu, X\. Yao, Y\. Yang, and J\. Chen \(2025\)Scaling long\-horizon LLM agent via context\-folding\.arXiv preprint arXiv:2510\.11967\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px4.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.8.1)\.
- H\. Wang, H\. Zou, H\. Song, J\. Feng, J\. Fang, J\. Lu, L\. Liu, Q\. Luo,et al\.\(2025a\)UI\-TARS\-2 technical report: advancing GUI agent with multi\-turn reinforcement learning\.arXiv preprint arXiv:2509\.02544\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- S\. Wang, W\. Liu, J\. Chen, Y\. Zhou, W\. Gan, X\. Zeng, Y\. Che, S\. Yu,et al\.\(2024\)GUI agents with foundation models: a comprehensive survey\.arXiv preprint arXiv:2411\.04890\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- X\. Wang, B\. Wang, D\. Lu, J\. Yang, T\. Xie, J\. Wang, J\. Deng, X\. Guo,et al\.\(2025b\)OpenCUA: open foundations for computer\-use agents\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2508\.09123Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- Z\. Wang, W\. Chen, L\. Yang, S\. Zhou, S\. Zhao, H\. Zhan, J\. Jin, L\. Li,et al\.\(2025c\)MP\-GUI: modality perception with MLLMs for GUI understanding\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Note:arXiv:2503\.14021Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. H\. Chi, Q\. V\. Le, and D\. Zhou \(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2201\.11903Cited by:[§3\.1](https://arxiv.org/html/2608.03327#S3.SS1.SSS0.Px3.p1.2)\.
- Q\. Wei, S\. Zeng, C\. Li, W\. Brown, O\. Frunza,et al\.\(2025\)Reinforcing multi\-turn reasoning in LLM agents via turn\-level reward design\.arXiv preprint arXiv:2505\.11821\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1),[§6](https://arxiv.org/html/2608.03327#S6.SS0.SSS0.Px1.p1.1)\.
- Z\. Wu, Z\. Wu, F\. Xu, Y\. Wang, Q\. Sun, C\. Jia, K\. Cheng, Z\. Ding,et al\.\(2025\)OS\-ATLAS: a foundation action model for generalist GUI agents\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2410\.23218Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- T\. Xie, D\. Zhang, J\. Chen, X\. Li, S\. Zhao, R\. Cao,et al\.\(2024\)OSWorld: benchmarking multimodal agents for open\-ended tasks in real computer environments\.InAdvances in Neural Information Processing Systems \(NeurIPS\) Datasets and Benchmarks Track,Note:arXiv:2404\.07972Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px1.p1.2),[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px3.p1.10)\.
- Y\. Xu, Z\. Wang, J\. Wang, D\. Lu, T\. Xie, A\. Saha, D\. Sahoo, T\. Yu, and C\. Xiong \(2025\)Aguvis: unified pure vision agents for autonomous GUI interaction\.InProceedings of the 42nd International Conference on Machine Learning \(ICML\),Note:arXiv:2412\.04454Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Yan, S\. Wang, J\. Du, Y\. Yang, Y\. Shan, Q\. Qiu, X\. Jia, X\. Wang,et al\.\(2025\)MCPWorld: a unified benchmarking testbed for API, GUI, and hybrid computer use agents\.arXiv preprint arXiv:2506\.07672\.Cited by:[§1](https://arxiv.org/html/2608.03327#S1.p1.1),[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Yang, Z\. Yang, Z\. Dou, A\. Nguyen, K\. You, O\. Attia, A\. Szot, M\. Feng,et al\.\(2025\)UltraCUA: a foundation model for computer use agents with hybrid action\.arXiv preprint arXiv:2510\.17790\.Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2608.03327#S2.T1.1.4.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao \(2023\)ReAct: synergizing reasoning and acting in language models\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2210\.03629Cited by:[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2608.03327#S3.SS0.SSS0.Px3.p1.19)\.
- Q\. Yu, Z\. Zhang, R\. Zhu, Y\. Yuan, X\. Zuo, Y\. Yue, W\. Dai, T\. Fan,et al\.\(2025\)DAPO: an open\-source LLM reinforcement learning system at scale\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2503\.14476Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px3.p1.1)\.
- S\. Zhou, F\. F\. Xu, H\. Zhu, X\. Zhou, R\. Lo, A\. Sridhar, X\. Cheng, T\. Ou,et al\.\(2024\)WebArena: a realistic web environment for building autonomous agents\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2307\.13854Cited by:[§2](https://arxiv.org/html/2608.03327#S2.SS0.SSS0.Px1.p1.1)\.

## Appendix AHarness Correctness

The reasoning\-gating claim presumes a*correct*harness, which we separate from the tool\-surfacing design that the claim is about\. Correctness has three requirements\. \(1\) The action parse format must match the model’s nativecomputer\_usecalling pattern\. \(2\) The coordinate convention must match the model: Qwen3\-VL emits*relative*coordinates on a10001000\-grid and requires a resize that absolute\-pixel models do not\. \(3\) The VM–agent feedback loop \(action execution→\\toobservation/error return\) must be closed and deterministic\. If any of these is wrong, observed failures stem from format rather than capability and the GUI\-vs\-MCP comparison is unfair\.

#### Verification\.

We check these requirements directly rather than by inspection\. Training is on\-policy by construction: per\-windowclip\_frac=0=0and ratio=1\.0=1\.0\. A saved checkpoint is byte\-identical to the base model except for the weights, so no configuration drift enters across resumes\. Every RL run reported in the paper uses this verified pipeline\.

## Appendix BContext\-Construction Strategy Menu

Table[4](https://arxiv.org/html/2608.03327#A2.T4)lists the full set of context\-construction strategies as choices over image\-history depthkkand the screenshot retention rule in Eq\. \(1\) of the main paper\. The first three rows correspond to the five harness configurations evaluated in Section 4\.1 of the main paper; the lower rows are natural extensions along the same axes\.

StrategyConstructionCost / trade\-offfullall screenshots\+\+all responsesmaximal info; context explodesbaseline \(default\)window of lastk=4k\{=\}4screenshots; older steps as textstandard bounded memoryskip\_on\_mcp\_successwithin the window, replace the screenshot after a successful MCP call with a text placeholdernear\-lossless; trims costliest frames*Extensions along the same axes:*windowk=1k\{=\}1keep only the current screenshotcheapest; loses spatial memorystate\-change skipdrop a frame byte\-identical to the previous onemore general than MCP\-gated skipdownscale / a11y treeshrink old frames or swap for accessibility textfurther token cutsTable 4:Context\-construction strategy menu\. Image\-history depthkkand retention rule \(whether to keep or replace the screenshot with a placeholder\) are the two axes\. The first three strategies correspond to the five configurations in Section 4\.1 of the main paper\.
## Appendix CConcrete Context Template

The message array instantiating Eq\. \(1\) of the main paper interleaves onesystemmessage with alternatinguser/assistantturns; within the depth\-kkwindow eachuserturn carries an optional relayed tool resultρ​\(ri\)\\rho\(r\_\{i\}\)and the retained frameo~i\\tilde\{o\}\_\{i\}, and the*current*turn additionally appends the instructionIIand the full action traceHtH\_\{t\}\. GUI and MCP calls share one<tool\_call\>schema \(the “unified action space”\)\.

```
system   : # Tools <tools>{computer_use}
         {libreoffice_calc.set_cell_value}
         ...BM25 top-18 signatures...
         </tools> + format rules
user     : [Tool result: ...]?
         + screenshot_i  <- windowed step i
assistant: (<think>...</think>)? Action: ...
         <tool_call>{...}</tool_call>
```

```
...        (last k steps kept)
user     : [Tool result: ...]?
         + current screenshot
         + Instruction: <I>
         + Previous actions: Step 1..t-1
           (H_t)
```

#### Worked example \(calc, “set A1 to 100”\)\.

*Step 1 user:*screenshot of the empty sheet,Instruction: set A1 to 100,Previous actions: None\.*Step 1 assistant*\(Thinking\):<think\>an MCP set\-value is faster than click\+type</think\> Action: set A1 via MCP\. <tool\_call\> \{"name": "libreoffice\_calc\.set\_cell\_value", "arguments": \{"cell": "A1", "value": 100\}\} </tool\_call\>\.*Step 2 user:*Tool result: \{"ok": true, "cell": "A1", "value": 100\}\+\+updated screenshot\+\+Previous actions: Step 1: set A1 via MCP; here thedroprule \(Eq\. \(2\) of the main paper\) would replace the screenshot withπ\\pisincec1c\_\{1\}was a successful MCP call\.*Step 2 assistant:*terminate/success\. Per\-step outputs measure∼1500\{\\sim\}1500characters for Thinking \(with the<think\>trace\) versus∼217\{\\sim\}217for Instruct, matching the66–7×7\\timestoken gap reported in Section 3\.1 of the main paper\.

## Appendix DAccuracy–Token Trade\-off \(Full Table\)

Table[5](https://arxiv.org/html/2608.03327#A4.T5)gives the exact per\-configuration numbers behind the context operating points in the token block of Table 2 of the main paper \(reasoning model, mean over 5 runs\); Figure[5](https://arxiv.org/html/2608.03327#A4.F5)plots the full accuracy–cost breakdown for both models\.

Config \(Thinking\)AccCumul\. input \(K\)Peak input p95Tok / 1%Acc \(K\)GUI only \(window\-4\)30\.5313\.11138510\.6window\-4 \(op\. point\)34\.5337\.11154410\.1window\-230\.6226\.173147\.8window\-4\+\+drop32\.3342\.41148711\.0window\-2\+\+drop \(ctx\_opt\)30\.6219\.572437\.7Table 5:Accuracy–token trade\-off for the reasoning model \(mean over 5 runs; peak is p95 over trajectories; Tok/1%Acc counts input\+\+output\)\. The RL operating point is window\-4 without drop; the context\-level compression is window\-2\+\+drop\.![Refer to caption](https://arxiv.org/html/2608.03327v1/x5.png)Figure 5:Accuracy vs\. cost across the five context configurations \(5\-run means±\\pmstd; shape = configuration, color = model\)\.\(a\)cumulative input tokens: the*\+MCP*arrows show the sign reversal; window\-4 \(no drop\) is the RL operating point, window\-2\+\+drop the token\-efficient knee\.\(b\)output tokens \(66–7×7\\timesreasoning premium\);\(c\)completion steps;\(d\)peak input \(p95\)\.
## Appendix EAdditional Results and Details

#### Per\-domain mechanism notes \(Section 3\.1 of the main paper\)\.

On writer, tasks where a tool is invoked succeed42%42\\%of the time versus82%82\\%without a call — a combination of difficulty self\-selection \(the model reaches for tools on harder tasks\) and mis\-parameterized calls to writer’s more complex tools — whereas calc and impress show almost no invoked/non\-invoked success\-rate difference\. In multi\_apps, tasks are often tool\-unreachable because early steps run in an application whose tools do not apply\. Gimp, Thunderbird, and Chrome expose no MCP tools at all; VLC exposes 12 native tools with 16/17 tasks tool\-reachable, yet neither model ever calls one\.

## Appendix FRL Training Hyperparameters

#### Hardware and infrastructure\.

All training runs use 1 node with 8×\\timesA100 \(80 GB\) GPUs under DeepSpeed ZeRO\-3 in bfloat16 mixed precision\. The policy model is Qwen3\-VL\-8B\-Thinking\. Environments run in 96 parallel Docker VMs \(Ubuntu\-MCP\.qcow2\) mounted in/dev/shm\. Experiment tracking via Weights & Biases \(projecthybrid\-routing\-rl\)\.

#### Rollout\.

Each training step rolls outG=8G\{=\}8trajectories per task across 96 parallel environments; with full coverage, all 74 gradient\-band tasks \(empirical pass ratep∈\(0\.1,0\.9\)p\\in\(0\.1,0\.9\)under temperature\-1\.0 sampling withk=8k\{=\}8; re\-profiled as the policy improves\) are rolled out every step \(74×8=59274\{\\times\}8\{=\}592trajectories per training step\)\. The dense\-bonus run of Section 5\.2 of the main paper uses a 24\-task fast\-iteration subset of the band\. Rollouts use temperature1\.01\.0,top\_p=1\.0=1\.0,max\_steps=50=50,max\_tokens=2048=2048per step, and image\-history depthk=4k\{=\}4\. The MCP tool retriever is BM25 top\-18, with fallback to GUI if no tool is called\. Evaluation uses greedy decoding \(temperature0\.00\.0\) under the same observation policy as rollout; greedy probes are repeated \(3 repetitions\) and judged against pre\-registered criteria\. For the matched\-training run of Section 5\.3 of the main paper, the compression rule \(drop\+\+k=2k\{=\}2\) is active in both rollout and evaluation, with each image decision logged and replayed in training; all other runs usek=4k\{=\}4without skip\.

#### Reward function\.

The per\-trajectory return is Eq\. \(3\) of the main paper: the±1\\pm 1outcome term plus the length and step\-cap tie\-breakers\. The tool bonusλmcp\\lambda\_\{\\text\{mcp\}\}is*not*part ofR​\(τ\)R\(\\tau\): it is added to the per\-step advantage after group normalization \(Eq\. \(4\) of the main paper\) and fires only for execution\-successful calls of non\-read\-only MCP tools whose \(tool, arguments\) key appears for the first time in the trajectory; it is not gated on task outcome\. Folding the bonus intoR​\(τ\)R\(\\tau\)instead is a dead signal: trajectory averaging,zz\-scoring, and the1/T1/Tbroadcast dilute it to∼10−4\{\\sim\}10^\{\-4\}\. All coefficients are listed in Table[6](https://arxiv.org/html/2608.03327#A6.T6)\.

#### Outcome\-only sweep\.

The sweep referenced in Section 5\.1 of the main paper covers learning rate \(10−610^\{\-6\}–10−510^\{\-5\}\), KL strength and anchor \(base vs\. rollout\-time\), normalization scheme \(group\- vs\. step\-level\), task density, and horizon\. No configuration moves held\-out or greedy accuracy, with zero sustained fail→\\topass flips; training\-side sampled curves separate only by KL anchor\. The one exception is behavioral rather than accuracy\-driven: a rollout\-time anchor combined with the dense bonus produces adoption drift that transfers to greedy decoding with no accuracy movement \(Section 5\.2 of the main paper\)\. The advantage standard deviation is invariant at0\.0680\.068–0\.0760\.076throughout\.

GroupParameterValueRewardλlen\\lambda\_\{\\text\{len\}\}\(length penalty coef\)0\.05TmaxT\_\{\\max\}\(length penalty denom\.\)50λcap\\lambda\_\{\\text\{cap\}\}\(step\-cap penalty\)0\.20λmcp\\lambda\_\{\\text\{mcp\}\}\(post\-norm\. step bonus\)0\.10GRPOGroup sizeGG8Clipϵ\\epsilon0\.20KL coefficientβ\\beta0\.02KL referencerollout\-time policyKL estimatork3k\_\{3\}Acc\. filter\(a​c​clo,a​c​chi\)\(acc\_\{\\text\{lo\}\},\\ acc\_\{\\text\{hi\}\}\)\(0\.05,0\.95\)\(0\.05,\\ 0\.95\)OptimizerAlgorithmAdamWLearning rate5×10−65\{\\times\}10^\{\-6\}\(β1,β2\)\(\\beta\_\{1\},\\beta\_\{2\}\)\(0\.9,0\.999\)\(0\.9,\\ 0\.999\)Weight decay0LR schedulecosine, no warmupMax grad norm1\.0TrainingPolicy updates per step4Max prompt length16 384 tokensMax gen length8 192 tokensMixed precisionbfloat16Random seed10086Table 6:RL hyperparameters\. All reward shaping coefficients not listed are 0\. The dense\-bonus run \(Section 5\.2 of the main paper\) usesλmcp=0\.1\\lambda\_\{\\text\{mcp\}\}\{=\}0\.1; the outcome\-only control and the matched\-training run \(Section 5\.3\) useλmcp=0\\lambda\_\{\\text\{mcp\}\}\{=\}0\. Anchoring the KL to the base model instead of the rollout\-time policy is the ablation discussed in Section 5\.1 of the main paper\.