Pith. sign in

REVIEW 4 major objections 3 minor 55 references

This paper argues that autonomous repair of multi-agent failures can be formulated as a Monte Carlo Tree Search over rollback, guided-repair, and continuation actions, outperforming prior repair methods by 3–12 absolute percentage points at

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

2026-08-03 14:28 UTC pith:EBZHH65I

load-bearing objection A credible MCTS repair loop plus a genuinely reusable benchmark, but the judge-driven search is never calibrated to ground truth and the paper ships no release — send it to reviewers, not straight to accept. the 4 major comments →

arxiv 2607.29055 v1 pith:EBZHH65I submitted 2026-07-31 cs.LG cs.AIcs.MA

Autonomous Repair for Multi-Agent Systems via Monte-Carlo Tree Search

Hanxiao Lu , Tianyi Zhang This is my paper
classification cs.LG cs.AIcs.MA
keywords multi-agent systemsautonomous repairMonte Carlo Tree Searchpartial rolloutLLM-as-a-judgefailure taxonomyreplayable benchmark
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.

Multi-agent systems fail often, and repairing them currently requires humans to locate the mistake and steer the agents back. This paper proposes MARS, a framework that automates the repair loop by treating the search over possible interventions as a Monte Carlo Tree Search: roll the system back to a suspected error step, apply a fix suggestion, run a few more steps, and score the partial trajectory with an LLM judge guided by a taxonomy of 14 common failure modes. The paper also contributes StateMAS, a benchmark of 1,310 replayable failure trajectories across four agent architectures and four LLM backbones. On this benchmark, MARS is reported to improve post-repair pass rates over strong baselines by 3.0 to 12.1 absolute percentage points while spending tokens comparable to the cheapest baseline. If correct, this makes automated repair practical for multi-agent deployments without a central orchestrator.

Core claim

The central claim is that repair of failed multi-agent trajectories need not be a monolithic regeneration or a single linear edit; it can be explored as a tree of small interventions. MARS is the concrete embodiment: each node is a replayable system state, and three actions—rollback, guided repair, and continuation—generate children. Because full rollouts are expensive, MARS evaluates children after only a small number of steps, using an LLM judge that scores partial progress against a five-tier rubric and diagnoses failures via a 14-mode taxonomy of common multi-agent failure patterns. The paper reports that this partial-rollout search outperforms the previous automated repair approach by 6

What carries the argument

The load-bearing mechanism is the combination of three repair actions with partial rollout and taxonomy-guided evaluation: rollback restores a saved state at a suspected faulty step; guided repair injects a fix suggestion and resumes for a limited number of steps; continuation resumes without edits. To score a partial trajectory, an LLM judge uses a 14-mode failure taxonomy and a five-level rubric, producing both a scalar reward for MCTS backpropagation and diagnostic feedback that the next expansion step verifies before choosing an action. This redesigned MCTS loop is what allows the search to explore many interventions cheaply instead of committing to one full regeneration.

Load-bearing premise

The search is steered by an LLM judge's scores of partially rolled-out trajectories, so if that judge is noisy or biased toward its own repair suggestions, the reported pass-rate gains could reflect judge preference rather than actual repair quality.

What would settle it

Take a random subset of benchmark trajectories, run the repair search with the judge's scores, and separately label whether the final answer matches ground truth; if the judge's intermediate scores correlate weakly with final ground-truth success (e.g., Spearman < 0.3), the search signal is unreliable. Equivalently, replacing the judge model with a different LLM and observing a large drop in post-repair pass rate would show the gain is judge-dependent.

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

If this is right

  • Automated repair becomes feasible even when no ground-truth oracle or test suite is available at inference time, since the search runs on the LLM judge's rubric alone.
  • Repair is no longer limited to centralized orchestrators; the three-action space applies to sequential, concurrent, and decentralized architectures as well.
  • Because trajectories are stored as replayable states, repair outcomes can be measured counterfactually: a researcher can see exactly which intervention flips a failure to success.
  • The reported token numbers suggest that search-based repair can be nearly as cheap as a single linear repair attempt, since rollback discards erroneous suffixes instead of retaining full histories.
  • The benchmark gives the field a standard set of 1,310 failure trajectories with state snapshots, making future repair methods directly comparable.

Where Pith is reading between the lines

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

  • A natural extension is to feed the diagnosis back into the agent's policy rather than only into the search: if the LLM judge's failure-mode labels are reliable, they could serve as training signal for a smaller model that predicts which action to take next.
  • The dependence on a single LLM-as-a-judge makes the reported gains partly a function of judge quality; swapping the judge model or using a verifier (e.g., unit tests on tool outputs) would test how much of the improvement comes from the search structure itself.
  • The same partial-rollout MCTS pattern could transfer to other stateful LLM systems—data pipelines, web agents, or code-generation toolchains—wherever trajectories can be checkpointed and resumed.
  • The 3.0–12.1% improvement is measured on post-repair pass rate; in deployment, the practical gain depends on how often a failure is worth repairing at all, which the paper does not model.

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 / 3 minor

Summary. The paper proposes MARS, a Monte-Carlo Tree Search framework for automated repair of multi-agent system (MAS) trajectories. MARS introduces three repair actions (Rollback, GuidedRepair, Continuation), uses partial rollouts of L steps to limit token cost, and employs an LLM judge with a 14-mode MAST-derived taxonomy to score partially executed trajectories and produce diagnostic feedback. The paper also introduces StateMAS, a benchmark of 1,310 replayable failure trajectories across four agent architectures and four LLM backbones, built on GAIA and AssistantBench. Experiments report that MARS outperforms DoVer, Reflexion, and ReAct baselines by 3.0–12.1 percentage points in post-repair pass rate across settings, with supposedly comparable token consumption. An ablation study attributes gains to tree search, taxonomy-guided evaluation, and partial rollouts.

Significance. If the results hold, MARS would be a meaningful advance: it is the first MAS repair method that goes beyond centralized orchestrator edits, and its partial-rollout design is a plausible way to make search-based repair token-efficient. StateMAS is also a valuable resource, since existing failure-attribution benchmarks are not replayable and DoVer does not release replayable states. The paper is broad in scope: 1,310 trajectories, 16 configurations, ablations, sensitivity analysis, and stochasticity runs. However, the central claims rest on an unvalidated LLM-judge reward signal and on token-cost statements that do not fully match the reported table. The benchmark and method are promising, but the evidence as presented does not yet establish the headline quantitative claims with the required rigor.

major comments (4)
  1. [Method: Taxonomy-Augmented Evaluation and Diagnosis; Algorithm 1] MARS's selection, expansion, and final answer all optimize the scalar reward r produced by an LLM judge (Algorithm 1, lines 12–22, 14–15). The paper never calibrates this judge against ground-truth pass/fail outcomes, and the judge uses the same LLM backbone as the repaired agents (Implementation Details). Since the search objective is r, the reported pass-rate gains are only evidence that MARS works if r is shown to correlate with true task success. Please add (i) a calibration or agreement analysis of judge scores vs. oracle outcomes, and (ii) an ablation using an oracle/ground-truth reward on a subset, or selecting by final-answer validity instead of r. Without this, part of the gain could be an artifact of optimizing a proxy rather than genuine repair quality.
  2. [Token Cost Analysis; Table 4] The abstract and main text make token-cost claims that are not supported by Table 4. The abstract says MARS has 'only 5.6% overhead' over the most economical baseline; the text says MARS 'reduces token consumption by 1.6% to 54.7%' vs Reflexion and incurs a 'relative increase ... from 14.0% to 38.7%' vs ReAct. But in Table 4, MARS vs Reflexion on Centralized/Qwen3.5-9B is 2,451,982 vs 1,916,430 tokens, a 27.9% increase, and MARS vs ReAct on Decentralized/Nemotron-3 is 620,946 vs 527,884, a 17.6% increase. Please recompute the percentages or clarify the aggregation behind each stated figure. The current text overstates the token-efficiency advantage.
  3. [Main Results; Tables 2-3 and Stochasticity Analysis] No confidence intervals or significance tests are reported. GAIA-Free has only 66 tasks per configuration and AssistantBench has 33, so a 3.0% improvement corresponds to roughly two tasks in GAIA-Free and a 6.1% improvement to two tasks in AssistantBench. The stochasticity analysis (Table 6) itself shows a 3.0 percentage-point spread for centralized and concurrent runs, equal to the smallest claimed improvement in the paper. Please report bootstrap confidence intervals or per-task significance tests (e.g., McNemar) for the headline comparisons, and state the exact number of tasks and failed trajectories underlying each percentage.
  4. [Algorithm 1; Implementation Details] Algorithm 1 returns the node with maximum r (lines 14–15) as best_traj. Because actions generate only L=4 partial-rollout steps, best_traj can be an incomplete trajectory. The pass-rate metric is externally evaluated on ground truth, but the paper does not specify how a partial best_traj is completed or counted. This is load-bearing for the experimental metric. Please state the completion procedure explicitly (e.g., always run the selected trajectory to termination after search, or only return nodes with r=1) and align the pseudocode and metric definition with that procedure.
minor comments (3)
  1. [Problem Formulation] Typo: 'Each state st ∈ S is a a replayable snapshot' should read 'is a replayable snapshot.'
  2. [Experiments: Setup] The hyperparameters K=3, L=4, C=1.41, and N_iter=75 are tuned on a single configuration (GAIA-Free, Qwen3-30B-A3B, centralized). Please state explicitly whether these values were frozen across all other configurations or re-tuned per setting; the sensitivity analysis alone does not establish generalization.
  3. [Benchmark Availability] As StateMAS is a proposed benchmark, the paper should include an artifact availability statement with a release URL, license, and instructions for replaying the trajectories. This is especially important because deterministic replay is a central claimed feature.

Circularity Check

0 steps flagged

No significant circularity; the reported repair gains are grounded in external benchmark pass rates and the taxonomy is inherited from prior external work.

full rationale

The paper's central claim is empirically grounded rather than derived from its own inputs. MARS's search reward is an LLM-judge score produced by the taxonomy-augmented evaluation prompt, and Algorithm 1 returns the highest-scoring trajectory, but the reported post-repair pass rate is measured against benchmark ground truth: the paper states, 'We label an execution as successful if the solution produced by the MAS matches the benchmark ground truth and failed otherwise.' Thus the final evaluation metric is external to the judge reward. The 14 failure patterns are explicitly taken from the prior MAST taxonomy ('We curated 14 failure patterns based on the MAST taxonomy, which was rigorously developed by a comprehensive analysis of 150 MAS failure trajectories (Cemri et al. 2026)'), not fit to StateMAS results. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported from the authors' own prior work, and no definitional equivalence makes the improvement force itself. The limitation section is honest about coverage and framework dependence. The uncalibrated-judge concern raised by the reader is a validity or reward-hacking risk, not a circularity step: the search optimizes a proxy, but the reported claim is not that proxy. The absence of judge calibration could affect reliability but does not make any equation reduce to its own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

All central results depend on the six hand-tuned hyperparameters above plus an uncalibrated LLM-judge reward; the taxonomy is borrowed, not derived. No new entities.

free parameters (6)
  • Branching factor K = 3
    Set from sensitivity analysis (K=2 gives 24.2%, K=3 gives 30.3% pass rate).
  • Partial-rollout length L = 4
    Set from sensitivity (L=2 gives 22.7%, L=4 gives 30.3%).
  • UCT exploration weight C = 1.41
    Standard value from prior MCTS work; no domain-specific justification.
  • Max search depth D_max = 8
    Chosen by hand; no ablation reported.
  • Iteration budget N_iter = 75
    Chosen via sensitivity (60→27.3%, 75→30.3%, 90→31.8%).
  • Judge/selector temperatures = 0.2 / 0.7
    Hand-chosen; no ablation.
axioms (6)
  • domain assumption Access to an initial failed trajectory τ_fail
    Problem Formulation defines repair starting from a known failed run; if such a trace is not available, MARS cannot start.
  • domain assumption No oracle is available during search
    Problem Formulation: 'we do not assume access to an oracle' — the method must rely on judge rewards.
  • ad hoc to paper LLM judge reliably scores partial trajectories
    MCTS rewards and action selection come from judge scores/failure-mode diagnoses; no calibration experiment is reported.
  • domain assumption Temperature 0 makes LLM replay deterministic
    Benchmark section asserts deterministic replay; no evidence for closed-source GPT-5.4-mini or for parallel/batched inference.
  • domain assumption MAST 14-mode taxonomy covers relevant failures
    The evaluation prompt relies on these modes taken from prior work; omissions would misdirect diagnosis.
  • domain assumption MAF state snapshots fully capture execution state
    StateMAS assumes serialized states can restore and resume execution exactly; not independently verified.

pith-pipeline@v1.3.0-daily-deepseek · 16178 in / 12720 out tokens · 110847 ms · 2026-08-03T14:28:30.132791+00:00 · methodology

0 comments
read the original abstract

Multi-agent systems (MAS) are increasingly deployed to solve complex tasks. In case of incorrect or unsatisfactory outputs, users have to manually locate agent mistakes by inspecting agent trajectories (i.e., {\em failure attribution}) and provide feedback to refine the outputs (i.e., {\em repair}). Despite some recent work in MAS failure attribution, automated mechanisms to recover from such mistakes remain largely unexplored. To bridge this gap, we propose MARS, a search-based framework that formulates MAS repair as a Monte Carlo Tree Search (MCTS) process and navigates the vast space of potential repairs via diagnosis-guided expansion with taxonomy-augmented evaluation. Unlike standard MCTS, which evaluates a complete simulation via full rollout, MARS evaluates the agent trajectory using partial rollout to reduce token consumption. Furthermore, we introduce StateMAS, a large-scale MAS repair benchmark with 1,310 replayable multi-agent failure trajectories spanning four types of agent architectures and four LLM backbones. Experiments on StateMAS demonstrate that MARS consistently outperforms state-of-the-art methods, achieving an absolute improvement from 3.0\% to 12.1\% across all settings, while maintaining a comparable token consumption cost. The ablation study further confirms that taxonomy-augmented evaluation and diagnosis-guided expansion are critical to achieving these performance gains.

Figures

Figures reproduced from arXiv: 2607.29055 by Hanxiao Lu, Tianyi Zhang.

Figure 1
Figure 1. Figure 1: The overview framework of MARS [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Sensitivity results with centralized orchestration [PITH_FULL_IMAGE:figures/full_fig_p009_2.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

55 extracted references · 10 linked inside Pith

  1. [1]

    Proceedings of Machine Learning Research , volume=

    Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems , author=. Proceedings of Machine Learning Research , volume=. 2025 , publisher=

    2025

  2. [2]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

    Why Do Multi-Agent LLM Systems Fail? , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

  3. [3]

    Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , pages=

    Interactive debugging and steering of multi-agent ai systems , author=. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , pages=

    2025

  4. [4]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Chatdev: Communicative agents for software development , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  5. [5]

    arXiv preprint arXiv:2409.16299 , year=

    Hyperagent: Generalist software engineering agents to solve coding tasks at scale , author=. arXiv preprint arXiv:2409.16299 , year=

    Pith/arXiv arXiv

  6. [6]

    The Twelfth International Conference on Learning Representations , year=

    MetaGPT: Meta programming for a multi-agent collaborative framework , author=. The Twelfth International Conference on Learning Representations , year=

  7. [7]

    First Conference on Language Modeling , year=

    Autogen: Enabling next-gen LLM applications via multi-agent conversations , author=. First Conference on Language Modeling , year=

  8. [8]

    The Twelfth International Conference on Learning Representations , year=

    Gaia: a benchmark for general ai assistants , author=. The Twelfth International Conference on Learning Representations , year=

  9. [9]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Assistantbench: Can web agents solve realistic and time-consuming tasks? , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Camel: Communicative agents for" mind" exploration of large language model society , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    arXiv preprint arXiv:2410.20285 , year=

    Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement , author=. arXiv preprint arXiv:2410.20285 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2310.04406 , year=

    Language agent tree search unifies reasoning acting and planning in language models , author=. arXiv preprint arXiv:2310.04406 , year=

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Rethinkmcts: Refining erroneous thoughts in monte carlo tree search for code generation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Generating code world models with large language models guided by monte carlo tree search , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Reasoning with language model is planning with world model , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  16. [16]

    The Eleventh International Conference on Learning Representations , year=

    Planning with Large Language Models for Code Generation , author=. The Eleventh International Conference on Learning Representations , year=

  17. [17]

    International Conference on Machine Learning , pages=

    Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  18. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Webpilot: A versatile and autonomous multi-agent system for web task execution with strategic exploration , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  19. [19]

    arXiv preprint arXiv:2407.01476 , year=

    Tree search for language model agents , author=. arXiv preprint arXiv:2407.01476 , year=

    arXiv

  20. [20]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    From generation to judgment: Opportunities and challenges of llm-as-a-judge , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  21. [21]

    The Twelfth International Conference on Learning Representations , year=

    Prometheus: Inducing fine-grained evaluation capability in language models , author=. The Twelfth International Conference on Learning Representations , year=

  22. [22]

    International conference on computers and games , pages=

    Efficient selectivity and backup operators in Monte-Carlo tree search , author=. International conference on computers and games , pages=. 2006 , organization=

    2006

  23. [23]

    2026 , url =

    Microsoft Agent Framework , author =. 2026 , url =

    2026

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Self-refine: Iterative refinement with self-feedback , author=. Advances in Neural Information Processing Systems , volume=

  25. [25]

    2023 , eprint=

    Scaling Relationship on Learning Mathematical Reasoning with Large Language Models , author=. 2023 , eprint=

    2023

  26. [26]

    2026 , note =

    Microsoft , title =. 2026 , note =

    2026

  27. [27]

    GitHub repository , howpublished =

    OpenAI , title =. GitHub repository , howpublished =. 2024 , publisher =

    2024

  28. [28]

    arXiv preprint arXiv:2307.13854 , year=

    Webarena: A realistic web environment for building autonomous agents , author=. arXiv preprint arXiv:2307.13854 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2411.04468 , year=

    Magentic-one: A generalist multi-agent system for solving complex tasks , author=. arXiv preprint arXiv:2411.04468 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2407.18901 , year=

    Appworld: A controllable world of apps and people for benchmarking interactive coding agents , author=. arXiv preprint arXiv:2407.18901 , year=

    Pith/arXiv arXiv

  31. [31]

    GitHub repository , howpublished =

    FoundationAgents , title =. GitHub repository , howpublished =. 2024 , publisher =

    2024

  32. [32]

    The Benefits of a Concise Chain of Thought on Problem-Solving in Large Language Models , url=

    Renze, Matthew and Guven, Erhan , year=. The Benefits of a Concise Chain of Thought on Problem-Solving in Large Language Models , url=. doi:10.1109/fllm63129.2024.10852493 , booktitle=

    arXiv 2024

  33. [33]

    Vicinagearth , volume=

    A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges , author=. Vicinagearth , volume=. 2024 , publisher=

    2024

  34. [34]

    arXiv 2025 , author=

    Large language model agent: A survey on methodology, applications and challenges. arXiv 2025 , author=. arXiv preprint arXiv:2503.21460 , year=

    Pith/arXiv arXiv 2025

  35. [35]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  36. [36]

    International Conference on Learning Representations (ICLR) , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. International Conference on Learning Representations (ICLR) , year=

  37. [37]

    Proceedings of the 29th Symposium on Operating Systems Principles , pages =

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author =. Proceedings of the 29th Symposium on Operating Systems Principles , pages =. 2023 , doi =

    2023

  38. [38]

    The First Conference on Language Modeling , year =

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , author =. The First Conference on Language Modeling , year =

  39. [39]

    Advances in Neural Information Processing Systems , volume =

    Attention Is All You Need , author =. Advances in Neural Information Processing Systems , volume =. 2017 , publisher =

    2017

  40. [40]

    International Conference on Learning Representations , year =

    Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer , author =. International Conference on Learning Representations , year =

  41. [41]

    Advances in Neural Information Processing Systems , volume =

    Language Models are Few-Shot Learners , author =. Advances in Neural Information Processing Systems , volume =

  42. [42]

    2026 , howpublished =

    Firecrawl: The API to Search, Scrape, and Crawl the Web , author =. 2026 , howpublished =

    2026

  43. [43]

    2026 , howpublished =

    Wikipedia , author =. 2026 , howpublished =

    2026

  44. [44]

    2024 , howpublished =

    LangChain , title =. 2024 , howpublished =

    2024

  45. [45]

    DoVer: Intervention-Driven Auto Debugging for

    Ming Ma and Jue Zhang and Fangkai Yang and Yu Kang and Qingwei Lin and Saravan Rajmohan and Dongmei Zhang , booktitle=. DoVer: Intervention-Driven Auto Debugging for. 2026 , url=

    2026

  46. [46]

    arXiv preprint arXiv:2408.15978 , year=

    WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration , author=. arXiv preprint arXiv:2408.15978 , year=

    Pith/arXiv arXiv

  47. [47]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , year=

    WebWalker: Benchmarking LLMs in Web Traversal , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , year=

  48. [48]

    International Conference on Learning Representations , year=

    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework , author=. International Conference on Learning Representations , year=

  49. [49]

    arXiv preprint arXiv:2308.08155 , year=

    Autogen: Enabling next-gen llm applications via multi-agent conversation , author=. arXiv preprint arXiv:2308.08155 , year=

    Pith/arXiv arXiv

  50. [50]

    The Thirteenth International Conference on Learning Representations (ICLR) , year=

    AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? , author=. The Thirteenth International Conference on Learning Representations (ICLR) , year=

  51. [51]

    arXiv preprint arXiv:2605.08715 , year=

    AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems , author=. arXiv preprint arXiv:2605.08715 , year=

    Pith/arXiv arXiv

  52. [52]

    Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Raffles: Reasoning-based attribution of faults for llm systems , author=. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  53. [53]

    International Conference on Learning Representations , volume=

    Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement , author=. International Conference on Learning Representations , volume=

  54. [54]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  55. [55]

    Advances in neural information processing systems , volume=

    Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=