Pith. sign in

REVIEW 4 major objections 6 minor 67 references

OpenForgeRL claims that a recording proxy plus remote per-rollout containers can turn any agent harness into trainable RL data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 07:03 UTC pith:AVUDWPYU

load-bearing objection A genuinely useful systems contribution with broad empirical validation, but the Eq. (1) credit-assignment assumption needs an ablation and the benchmark results need seeds before publication. the 4 major comments →

arxiv 2607.21557 v2 pith:AVUDWPYU submitted 2026-07-23 cs.AI cs.CL

OpenForgeRL: Train Harness-native Agents in Any Environment

Xiao Yu , Baolin Peng , Ruize Xu , Hao Zou , Qianhui Wu , Hao Cheng , Wenlin Yao , Nikhil Singh
show 2 more authors
Zhou Yu Jianfeng Gao
This is my paper
classification cs.AI cs.CL
keywords harness-based agentsreinforcement learningproxy trajectory reconstructioncontainerized rolloutsRL training infrastructuretool-use agentsGUI agentsagentic reliability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

OpenForgeRL is an attempt to make end-to-end training of harness-based agents practical for open research. The paper's central claim is that a lightweight proxy which intercepts and records every model call made by an agent harness, combined with a container orchestrator that runs each rollout in its own remote sandbox, converts the harness's stateful multi-process inference into ordinary state-action-reward samples that any standard RL codebase can consume. If correct, this removes the train-deploy mismatch that forces open efforts to reimplement simplified harnesses, and lets agents be optimized in the exact harness and environment they will be deployed in. The paper validates this on tool-use and multimodal GUI agents with only hundreds to a few thousand tasks: OpenForge-Claw reaches 31.7 on ClawEval and 33.7 on QwenClawBench, and OpenForge-GUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager, beating similar-size open baselines on nearly all benchmarks. It also reports that RL mainly improves agentic reliability—self-verification, tool coverage, multi-step plan completion—while error recovery remains weak, and that some harnesses are substantially harder to learn than others.

Core claim

The paper's core proposal is that a harness's inference process can be treated as a black-box MDP and made trainable by instrumentation rather than by reimplementation. A proxy sits in front of the model server, intercepts every generation request the harness issues, and records the resulting prompt-response pairs as trajectory steps (H(s_t), a_t). When a rollout ends, the terminal success reward is assigned to every recorded step with the discount gamma^(T-t), normally gamma=1.0. Because each rollout runs in its own remote container managed by an orchestrator, the trainer never needs to understand the harness's internal subagents, planners, or context management; it just receives standard (

What carries the argument

The load-bearing mechanism is the proxy-plus-orchestrator pair. The proxy wraps the harness's inference server and records every prompt-response pair the harness generates, including calls made by internal subagents and context managers, reconstructing each rollout as an unordered trajectory of (H(s_t), a_t) pairs. The orchestrator launches each rollout in its own remote container, so harness rollouts scale elastically and are fully decoupled from the training GPUs. Together they convert a stateful, multi-process harness into the flat (state, action, reward) sample format expected by standard RL training code, with terminal reward propagated to every step as gamma^(T-t) times r_T (gamma=1.0)

Load-bearing premise

The load-bearing premise is that every prompt-response pair the proxy records—including calls made by harness-internal subagents, planners, and context managers—is a genuine decision step that deserves the full terminal reward discounted with gamma=1.0; if those internal calls are not actually decision steps, RL credit assignment is training on noise.

What would settle it

Take a harness whose internal planner emits many prompt-response pairs before a single successful action, train two identical RL runs on the same tasks—one rewarding every recorded pair with the terminal reward and one rewarding only the final action—and compare final benchmark performance; roughly equal results would support the proxy-equivalent MDP assumption, while a reliable gap in favor of the final-action-only run would falsify it.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Agents can be trained end-to-end in the exact harness and environment they are deployed in, eliminating the train-deploy mismatch caused by simplified training reimplementations.
  • The same infrastructure works across text-based tool-use, browser-use, and computer-use agents, since only the sandbox container changes when the harness or environment changes.
  • RL on top of SFT improves agentic reliability—self-verification, tool coverage, and multi-step plan completion—using only a few hundred RL tasks, and these gains transfer to harnesses not seen in training.
  • Harness choice becomes a measurable training variable: simpler harnesses are easier to learn, and training across multiple harnesses helps more than training on a single one.
  • Open teams can reproduce and scale this recipe with a standard RL codebase and a cloud provider, because each rollout is an independent container and there is no special coupling to the trainer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's Eq. (1) treats every recorded prompt-response pair as a decision step receiving the full terminal reward; if a harness's internal subagent or planner makes calls that are not true policy decisions, RL credit assignment would train on noise. A natural ablation is to reward only the outermost agent's actions and compare.
  • The observation that some harnesses are much harder to learn than others suggests co-adaptation: one could treat harness prompts, tool descriptions, and control flows as trainable parameters alongside the policy, rather than as fixed infrastructure.
  • Since error recovery remains weak after RL, a testable extension is to add dedicated recovery-oriented rewards or curriculum tasks that force the agent to retry after failed commands, which the paper's data pipeline could generate automatically.
  • The 'any harness x any environment' claim could be stress-tested by training on a harness whose internal control flow is stochastic or generates many subagent calls, and checking whether trajectory reconstruction still yields stable RL improvements.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. OpenForgeRL proposes an infrastructure for RL training of harness-based agents: a proxy intercepts all model calls made by an agent harness (e.g., OpenClaw, Codex, Kimi-Agent), records them as prompt-response pairs, and reconstructs them as RL trajectories; a Kubernetes orchestrator runs each rollout in a remote container. The paper validates this on text-based claw/tool-use agents (OpenForge-Claw, Qwen3-30B-A3B) and multimodal GUI agents (OpenForge-GUI, Qwen3-VL-8B), reporting improvements over similar-size open baselines on ClawEval, QwenClawBench, MCPAtlas, OSWorld-Verified, Online-Mind2Web, and WebVoyager. It also analyzes how harness choice and RL affect agent behavior, finding that simpler harnesses are easier to learn, RL improves self-verification and tool coverage, and error recovery remains weak. The central claim is that the proxy-based trajectory reconstruction makes any harness and environment trainable with standard RL codebases such as veRL.

Significance. If the empirical claims hold, OpenForgeRL is a practically valuable contribution: it addresses a real infrastructure gap by enabling end-to-end RL in the actual harnesses used at deployment, avoiding train-deploy mismatch, and it demonstrates results across a wider range of environments than prior work. The paper is strong in its engineering details: the appendices carefully document training setups, evaluation protocols, data synthesis, and even the exact action spaces, and the planned open-source release would allow others to reproduce and extend the work. The behavioral analysis (Section 5) is a useful addition, going beyond scores to examine harness-dependent learning difficulty and generalization. However, the validity of the RL signal under the proxy-equivalent MDP assumption is not established, and the empirical comparisons lack the statistical grounding needed to fully support the headline benchmark claims.

major comments (4)
  1. [§3.2, Eq. (1)] The proxy-equivalent MDP assumption is load-bearing but unvalidated. Eq. (1) assigns every recorded prompt-response pair, including harness-internal subagent/planner/context-manager calls, a reward r_t = γ^(T−t)·r_T with γ=1.0. This means the policy gradient trains on all internal calls as if they were environment-level decisions, giving full terminal credit to non-decision steps. The paper itself acknowledges in §3.2 that partial rollouts can inject misleading signal and defers better credit assignment to future work, yet this same broadcast reward is used in every reported RL run. Consequently, the benchmark gains in Tables 2–4 cannot distinguish genuine policy improvement from artifacts such as length bias or credit attributed to internal calls that did not causally produce the outcome. I request an ablation that isolates this assumption—for example, training with only terminal-step r
  2. [Tables 2 and 3] All reported benchmark numbers are single runs with no error bars, multiple seeds, or significance tests. This is particularly concerning in Table 2, where the gain of OpenForge-Claw(SFT+RL) over the SFT checkpoint on ClawEval pass3 is 31.7 vs. 21.7, but the difference over the strong baseline Qwen3-Coder-30B-A3B-Instruct (30.4) is small; without variance estimates the headline improvements may be within noise. Table 3 also compares models under different step caps (#Steps column): OpenForge-GUI is evaluated with 30 steps while several baselines use 50 or 100 steps. Because GUI success rates are sensitive to step budget, this is a confound. The MCPAtlas result is additionally on a reduced 89-task subset, which the paper transparently states, but it makes the '28.1' number not directly comparable to other MCPAtlas leaderboard values. I ask for error bars or multiple seeds on the main resu
  3. [§B.2] The SFT data filtering uses GPT-5.4 as a judge without the test-and-refine environment verifier, while RL tasks use the full verifier. The paper states this is affordable because SFT success signals are used only for one-time filtering, but the SFT trajectories are the initialization for RL, so label noise in SFT can propagate. The manuscript does not assess the judge's accuracy against the verifier, nor the effect of SFT label noise on the final RL policy. I request a small empirical check—e.g., judge agreement on a sample of RL-verified tasks—or a discussion of why SFT label noise is benign for the downstream RL result.
  4. [§5.2, Table 5] The generalization-to-unseen-harness analysis reports improvements of +3.3 and +4.6 on OpenClaw and Codex for a ZeroClaw-only model. These are single-run deltas over the base model and are small relative to the variance typical of these benchmarks. The paper interprets them as evidence of cross-harness transfer, but without repeated evaluations or confidence intervals the claim is not firmly supported. The same applies to the comparison between ZeroClaw-only and multi-harness training (46.0 vs. 48.5), which is a single trial. Please provide multiple evaluation runs or otherwise quantify uncertainty for the key numbers in Table 5.
minor comments (6)
  1. [§3.1] The trajectory τ is described as an 'unordered collection' but is then written as an ordered sequence. Please correct to 'ordered sequence' or 'list'.
  2. [§3.2, Eq. (1)] Notation is inconsistent: the text defines prompt-response pairs as (H(s_t), a_t) but Eq. (1) uses (s_H, a) without the time index. Align the notation.
  3. [§B.2, §B.3] Typos: 'suitbale' and 'afforadable' (§B.2), 'distrations' (§B.3), 'anf' (§B.3). Also, the phrase 'reserve only the last turn for more than 3 consecutive identical actions' is ambiguous—presumably it means remove repeated actions except the last, but the writing should be clarified.
  4. [Figure 1] The right panel labels 'OpenForge-RL' in the legend, but the corresponding model is called 'OpenForge-GUI' elsewhere. Also, the figure is dense and the reader's eye is drawn to the left panel; consider splitting or enlarging.
  5. [Table 3 caption] The header 'OnlineMind2Web' should be 'Online-Mind2Web' for consistency with the text.
  6. [§4.1, Table 2] The metric names 'pass3' and 'pass@3' are used side by side; clarify whether these are the same or different metrics (e.g., 'pass at 3 attempts' vs. 'pass@3 sampling').

Circularity Check

0 steps flagged

No significant circularity: the paper's central claims are benchmark outcomes measured on external evaluators, and no derivation step reduces to its own inputs or to a load-bearing self-citation.

full rationale

OpenForgeRL is an empirical systems paper; its headline results are pass rates on external benchmarks (ClawEval, QwenClawBench, MCPAtlas, OSWorld-Verified, Online-Mind2Web, WebVoyager) using official evaluation protocols. Equation (1), r_t = gamma^(T-t)*r_T with gamma=1.0, is a training credit-assignment choice, not a fitted constant renamed as a prediction: the terminal reward comes from task verifiers or LLM judges, and the reported benchmark gains are not entailed by this reward rule alone. The paper's self-citations to Orchard, OpenWebRL, and Synthetic-Computer-Use supply infrastructure, data pipelines, and asset pools; they are cited as code/data sources, not as uniqueness theorems or as the origin of the benchmark numbers. The paper itself flags the partial-rollout credit-assignment limitation in Section 3.2 ('Designing better credit assignment ... left to future work'), which is a correctness/robustness concern rather than a circular reduction. No step defines X in terms of Y, fits a parameter and then predicts a closely related quantity, or imports a forced choice through an unverified self-citation chain. The evaluation is self-contained against external benchmarks, so there is no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

This is an empirical systems paper, not a derivation: no fitted constants in a mathematical proof and no invented physical/theoretical entities. The free parameters are hand-set training/evaluation choices; the axioms are the RL background, the unvalidated proxy-level credit-assignment assumption, verifier/judge correctness, benchmark disjointness, and reliance on the self-cited Orchard orchestrator.

free parameters (4)
  • RL hyperparameters (LR, KL, entropy, batch/group, timeout) = Claw: lr 1e-6, kl 0.001, entropy 0, batch 8, group 8, timeout 900s; Computer: 5e-7/0.01/0, 8/8, 600s; Browser: 1e-6/0.0/
    Hand-selected per domain; not fitted to evaluation benchmarks but influence training dynamics and final scores.
  • Trajectory reward decay γ = 1.0
    Eq. (1) assigns r_t = γ^(T−t) r_T; γ=1.0 is a modeling choice about credit assignment, not benchmark-fit.
  • Browser task dedup threshold = 0.55 (embedding similarity)
    Greedy similarity threshold chosen to preserve diversity; arbitrary and affects training pool composition.
  • Evaluation step/time caps = 30 max steps for OSWorld/Mind2Web/WebVoyager; 20 max RL browser turns; 600-900s rollout timeouts
    Hand-set caps trade coverage vs. cost; differing step budgets across baselines confound comparisons.
axioms (5)
  • standard math MDP formulation and GRPO advantage estimation are assumed correct.
    Used without derivation in Sections 3.1-3.2; standard in RL literature, following Feng et al. 2025 and Guo et al. 2025.
  • ad hoc to paper Proxy-recorded io-pairs (H(s_t), a_t) form a valid MDP trajectory; all harness-internal model calls are legitimate training steps and can receive r_t = γ^(T−t) r_T.
    Section 3.2, Eq. (1); load-bearing, unvalidated.
  • domain assumption Synthetic verifiers and LLM judges score task success correctly; SFT tasks are filtered by GPT-5.4 judge without test-and-refine.
    Sections 3.3, B.2, B.3; if verifier/judge is wrong, SFT/RL data contain false positives/negatives.
  • domain assumption Training and evaluation task distributions do not overlap; Claw tasks seeded from ZClawBench/ClawHub are disjoint from ClawEval/QwenClawBench.
    Section 4.1, B.2; overlap risk is not quantified.
  • domain assumption Orchard's Kubernetes orchestration works as described and is correctly inherited.
    Section 3.2; central to rollout scaling but not independently tested in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 24050 in / 15162 out tokens · 139431 ms · 2026-08-01T07:03:55.042138+00:00 · methodology

0 comments
read the original abstract

Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks cannot natively express stateful, multi-process harness inference. To address this, we present OpenForgeRL, an open-source framework for training harness-based agents end-to-end in diverse environments. OpenForgeRL achieves this with a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase (e.g., veRL), and a Kubernetes orchestrator that runs each rollout in its own remote container, together enabling training on any harness in any environment at scale. By decoupling training and inference, OpenForgeRL allows researchers to easily train, study, and improve agents directly in the real harnesses and environments they are deployed with. We validate our framework across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents. Using only hundreds to a few thousand tasks, OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. Both outperform open baselines of similar size on nearly all benchmarks, and in the GUI setting match or surpass models several times larger. Beyond benchmarks, we analyze how harness choice (e.g., ZeroClaw, OpenClaw, Codex) and RL shape agent behavior. We find that some harnesses are substantially harder to learn than others, and that RL improves agentic reliability, such as self-verification, tool coverage, and completing multi-step plans, though critical abilities such as error recovery remain weak.

Figures

Figures reproduced from arXiv: 2607.21557 by Baolin Peng, Hao Cheng, Hao Zou, Jianfeng Gao, Nikhil Singh, Qianhui Wu, Ruize Xu, Wenlin Yao, Xiao Yu, Zhou Yu.

Figure 1
Figure 1. Figure 1: Left: OPENFORGE RL builds on Orchard Env (Peng et al., 2026) and connects any harness × any environment to standard RL codebases such as veRL, with no train–deploy mismatch. Right: OPENFORGE-trained models evaluated with six harnesses across six Claw and GUI environments. ∗Equal Advisory Contribution; † Project Lead 1 arXiv:2607.21557v2 [cs.AI] 24 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of OPENFORGE RL. An orchestrator spawns remote sandboxes in which an LLM/VLM interacts with its environment through a harness. A proxy intercepts the harness’s LLM calls, routes them to the RL framework’s inference engines, and records the exchanged io-pairs as training trajectories. Supporting a new harness or environment only modifies the sandbox. over generation throughout a rollout; harnesses … view at source ↗
Figure 4
Figure 4. Figure 4: Distribution (%) of tasks used in the Claw, Computer-Use, and Browser-Use domains (top [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our data/task synthesis pipeline. This work aims to train harness-based agents end-to￾end across diverse environments beyond coding, such as browser- and computer-use. Unlike coding (Badertdinov et al., 2026), these domains offer far fewer training tasks, harnesses, and RL-ready environments. To help exper￾iment our framework in diverse environments, we there￾fore also built a simple pipeline t… view at source ↗
Figure 5
Figure 5. Figure 5: SFT vs. SFT+RL behavior on ClawEval. RL shifts calls from the generic [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 4 linked inside Pith

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    2026 , eprint =

    MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers , author =. 2026 , eprint =

    2026

  5. [5]

    2024 , url=

    John Yang and Carlos E Jimenez and Alexander Wettig and Kilian Lieret and Shunyu Yao and Karthik R Narasimhan and Ofir Press , booktitle=. 2024 , url=

    2024

  6. [6]

    2025 , note=

    Evolving SkyRL into a Highly-Modular RL Framework , author=. 2025 , note=

    2025

  7. [7]

    arXiv preprint arXiv:2605.24220 , year=

    Polar: Agentic RL on Any Harness at Scale , author=. arXiv preprint arXiv:2605.24220 , year=

    Pith/arXiv arXiv

  8. [8]

    2024 , journal =

    HybridFlow: A Flexible and Efficient RLHF Framework , author =. 2024 , journal =

    2024

  9. [9]

    2025 , howpublished =

    Zilin Zhu and Chengxing Xie and Xin Lv and slime Contributors , title =. 2025 , howpublished =

    2025

  10. [10]

    arXiv preprint arXiv:2405.11143 , year=

    OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework , author=. arXiv preprint arXiv:2405.11143 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2505.10978 , year=

    Group-in-Group Policy Optimization for LLM Agent Training , author=. arXiv preprint arXiv:2505.10978 , year=

    Pith/arXiv arXiv

  12. [12]

    2025 , eprint=

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author=. 2025 , eprint=

    2025

  13. [13]

    2025 , eprint=

    AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems , author=. 2025 , eprint=

    2025

  14. [14]

    2026 , eprint=

    OpenComputer: Verifiable Software Worlds for Computer-Use Agents , author=. 2026 , eprint=

    2026

  15. [15]

    2026 , eprint=

    MolmoWeb: Open Visual Web Agent and Open Data for the Open Web , author=. 2026 , eprint=

    2026

  16. [16]

    Proposer-Agent-Evaluator (

    Zhou, Yifei and Yang, Qianlan and Lin, Kaixiang and Bai, Min and Zhou, Xiong and Wang, Yu-Xiong and Levione, Sergey and Li, Erran , booktitle =. Proposer-Agent-Evaluator (. 2025 , url=

    2025

  17. [17]

    2026 , eprint=

    WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks , author=. 2026 , eprint=

    2026

  18. [18]

    2026 , eprint=

    OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents , author=. 2026 , eprint=

    2026

  19. [19]

    2026 , eprint=

    Orchard: An Open-Source Agentic Modeling Framework , author=. 2026 , eprint=

    2026

  20. [20]

    2024 , eprint=

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering , author=. 2024 , eprint=

    2024

  21. [21]

    2026 , eprint=

    Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces , author=. 2026 , eprint=

    2026

  22. [22]

    Forty-second International Conference on Machine Learning , year=

    The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models , author=. Forty-second International Conference on Machine Learning , year=

  23. [23]

    2024 , eprint=

    WebArena: A Realistic Web Environment for Building Autonomous Agents , author=. 2024 , eprint=

    2024

  24. [24]

    2025 , eprint=

    An Illusion of Progress? Assessing the Current State of Web Agents , author=. 2025 , eprint=

    2025

  25. [25]

    2024 , eprint=

    OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments , author=. 2024 , eprint=

    2024

  26. [26]

    2024 , eprint=

    SWE-bench: Can Language Models Resolve Real-World GitHub Issues? , author=. 2024 , eprint=

    2024

  27. [27]

    OpenAI , year=. Hello

  28. [28]

    Introducing

    OpenAI , year=. Introducing

  29. [29]

    Overview -

    Anthropic , year=. Overview -

  30. [30]

    SkyRL-v0: Train Real-World Long-Horizon Agents via Reinforcement Learning , author =

  31. [31]

    2025 , eprint=

    SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution , author=. 2025 , eprint=

    2025

  32. [32]

    arXiv preprint arXiv:2603.10165 , year=

    OpenClaw-RL: Train Any Agent Simply by Talking , author=. arXiv preprint arXiv:2603.10165 , year=

    Pith/arXiv arXiv

  33. [33]

    2026 , eprint=

    Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents , author=. 2026 , eprint=

    2026

  34. [34]

    2024 , eprint=

    Agentless: Demystifying LLM-based Software Engineering Agents , author=. 2024 , eprint=

    2024

  35. [35]

    2025 , eprint=

    OpenHands: An Open Platform for AI Software Developers as Generalist Agents , author=. 2025 , eprint=

    2025

  36. [36]

    2024 , eprint=

    CogAgent: A Visual Language Model for GUI Agents , author=. 2024 , eprint=

    2024

  37. [37]

    2025 , eprint=

    UI-TARS: Pioneering Automated GUI Interaction with Native Agents , author=. 2025 , eprint=

    2025

  38. [38]

    2023 , eprint=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. 2023 , eprint=

    2023

  39. [39]

    2023 , eprint=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. 2023 , eprint=

    2023

  40. [40]

    2023 , eprint=

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation , author=. 2023 , eprint=

    2023

  41. [41]

    2024 , eprint=

    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework , author=. 2024 , eprint=

    2024

  42. [42]

    2025 , eprint=

    MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use , author=. 2025 , eprint=

    2025

  43. [43]

    2024 , eprint=

    DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning , author=. 2024 , eprint=

    2024

  44. [44]

    2025 , eprint=

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning , author=. 2025 , eprint=

    2025

  45. [45]

    2025 , eprint=

    WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning , author=. 2025 , eprint=

    2025

  46. [46]

    2026 , eprint=

    AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning , author=. 2026 , eprint=

    2026

  47. [47]

    2025 , eprint=

    Dyna-Mind: Learning to Simulate from Experience for Better AI Agents , author=. 2025 , eprint=

    2025

  48. [48]

    2024 , eprint=

    Agent S: An Open Agentic Framework that Uses Computers Like a Human , author=. 2024 , eprint=

    2024

  49. [49]

    2023 , eprint=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. 2023 , eprint=

    2023

  50. [50]

    2026 , eprint=

    SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale , author=. 2026 , eprint=

    2026

  51. [51]

    2026 , howpublished =

    ZClawBench , author=. 2026 , howpublished =

    2026

  52. [52]

    2026 , eprint=

    AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents , author=. 2026 , eprint=

    2026

  53. [53]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  54. [54]

    2026 , howpublished =

    MiniMax-M2.5 , author=. 2026 , howpublished =

    2026

  55. [55]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  56. [56]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Ziyi and Liu, Aixin and Xue, Bing and Wang, Bingxuan and Wu, Bochao and Feng, Bei ...

  57. [57]

    2026 , eprint=

    Synthetic Computers at Scale for Long-Horizon Productivity Simulation , author=. 2026 , eprint=

    2026

  58. [58]

    2025 , eprint=

    OpenCUA: Open Foundations for Computer-Use Agents , author=. 2025 , eprint=

    2025

  59. [59]

    Stealth Browsers for AI Agents , year =

  60. [60]

    2026 , eprint=

    Kimi K2.5: Visual Agentic Intelligence , author=. 2026 , eprint=

    2026

  61. [61]

    2024 , eprint=

    WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models , author=. 2024 , eprint=

    2024

  62. [62]

    2025 , eprint=

    AgentTrek: Agent Trajectory Synthesis via Guiding Replay with Web Tutorials , author=. 2025 , eprint=

    2025

  63. [63]

    2024 , eprint=

    Executable Code Actions Elicit Better LLM Agents , author=. 2024 , eprint=

    2024

  64. [64]

    arXiv:2511.19663 , year=

    Fara-7B: An Efficient Agentic Model for Computer Use , author=. arXiv:2511.19663 , year=

    arXiv

  65. [65]

    2026 , howpublished =

    Introducing Claude 4.6 , author=. 2026 , howpublished =

    2026

  66. [66]

    2025 , eprint=

    InSTA: Towards Internet-Scale Training For Agents , author=. 2025 , eprint=

    2025

  67. [67]

    2025 , eprint=

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. 2025 , eprint=

    2025