Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Web agents stay on track longer when every plan step states what would prove it wrong.

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 · grok-4.5

2026-07-31 22:04 UTC pith:WKETHSKT

load-bearing objection Solid systems paper: FCUs with explicit falsifiers plus plan–test–repair deliver a real WebArena gain; mechanism evidence is thinner than the headline but not hollow. the 4 major comments →

arxiv 2607.24167 v1 pith:WKETHSKT submitted 2026-07-27 cs.AI

Falsifiable Commitment Planning for Self-Correcting Web Agents

Guangyi Liu , Huan Zhao , Quanming Yao This is my paper
classification cs.AI
keywords web agentslong-horizon planningfalsifiable commitmentsplan-test-repairdrift detectionskill reusescope-aware repairWebArena
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.

Long web tasks often fail quietly: each click still looks reasonable while the agent has already left the path that can finish the job. This paper argues that ordinary plans leave that validity check implicit, so agents keep following stale routines until the end. FCPAgent turns each plan step into a falsifiable commitment—a subgoal, optional reusable skill, confirming evidence, falsifying evidence, and a confidence score—and runs a plan–test–repair loop around it. Before and after browser actions, a hybrid tester matches evidence cheaply and escalates only ambiguous or risky cases to an LLM diagnosis; when a commitment is falsified, repair revises the smallest implicated piece (action, skill, or plan). On WebArena the method lifts average success 13.8% over the strongest baseline, with the biggest gains on the longest tasks, and it still helps zero-shot on a harder chore benchmark.

Core claim

Stable long-horizon web agents need plans that know when they are wrong. Representing each step as a Falsifiable Commitment Unit and executing it under online hybrid testing plus scope-aware repair raises WebArena average success from 57.4% to 65.3% (13.8% relative), with especially large gains on long Shopping and GitLab tasks and a 9.6% relative zero-shot gain on WebChoreArena.

What carries the argument

Falsifiable Commitment Unit (FCU): a plan step as (subgoal, optional skill, confirming evidence, falsifying evidence, confidence), run inside a plan–test–repair loop whose hybrid tester checks actions and observations and whose repair localizes contradictions to execution, skill, or planning scope.

Load-bearing premise

The confirming and falsifying evidence the planner writes, plus the fixed hybrid routing thresholds, must be accurate enough at test time to catch real drift without constant false alarms.

What would settle it

If ablating falsifying evidence or online commitment testing erases the long-horizon gains on WebArena (or drift detection on held-out trajectories falls near chance), the claim that explicit falsifiable commitments drive the improvement fails.

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

If this is right

  • Long web tasks benefit more from explicit progress and falsification criteria than from more planning or memory alone.
  • Drift should be attributed to action, skill, or plan so recovery revises the smallest adequate part.
  • Hybrid fast evidence matching plus selective LLM diagnosis can cut verification cost without losing success.
  • Skills from successes and failure-repair memories from failures both sharpen future commitments and recoveries.
  • Cross-benchmark reuse of fixed experience libraries remains useful when commitments carry their own validity tests.

Where Pith is reading between the lines

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

  • The same commitment-testing interface could transfer to other partially observed tool environments (desktop GUIs, multi-app workflows) where local actions look fine after the plan has already failed.
  • Early-detection benchmarks—measuring how soon recoverable drift is caught, not only final success—would pressure-test the method where the authors say evaluation is still thin.
  • Learned or calibrated falsifiers could relax dependence on one-shot planner-written evidence and fixed score gates.

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 / 5 minor

Summary. The paper proposes FCPAgent, a web-agent framework in which each plan step is a Falsifiable Commitment Unit (FCU): a subgoal, an optional retrieved skill, staged confirming evidence, hierarchical falsifying evidence (execution/skill/planning level), and a confidence score. Execution proceeds as a plan–test–repair loop: a hybrid tester (lightweight NLI/ITM evidence matching plus an LLM verifier for escalated cases) checks candidate actions and post-action states against the active FCU, and a scope-aware repairer revises the smallest implicated component using a frozen failure-repair library. Skill and failure-repair libraries are distilled offline from 191 WebArena training tasks and frozen for the 621-task test split; all methods share the Qwen3.5-397B-A17B backbone. On WebArena, FCPAgent reaches 65.3% average success vs 57.4% for the strongest baseline (ColorBrowserAgent), a 13.8% relative gain concentrated in long-horizon Shopping/GitLab buckets, plus a 9.6% relative zero-shot gain on WebChoreArena. Ablations attribute gains to planning, testing, and repair; Table 10 reports 80% F1 drift detection on 30 annotated trajectories. The central claim is that the FCU-based testing loop causes the gains via drift detection and scoped repair.

Significance. If the attribution questions are resolved, this is a solid contribution to reliable long-horizon web agents. Strengths worth naming: (i) clean experimental controls — one backbone across all methods, experience libraries frozen at test time, three runs per method, equal data budgets for memory-based baselines; (ii) external, non-circular success metrics (WebArena/WebChoreArena validators); (iii) a rare zero-shot cross-benchmark transfer result (Table 3); (iv) component and fine-grained ablations (Tables 4, 8) plus honest runtime instrumentation (Tables 9, 10, Figure 10) that most agent papers omit; (v) an efficiency accounting showing the hybrid design trades ~20% LLM time for nothing lost in success; (vi) the paper itself attempts a falsifiable, task-level diagnostic of its own mechanism rather than resting on final success alone — the right instinct, even though the execution (Table 10) is currently underpowered. The FCU formulation (subgoal + skill + staged confirming evidence + hierarchical falsifiers + confidence) is a genuinely reusable representation.

major comments (4)
  1. [§C.4, Table 10] Table 10 is the only direct validation of the claimed mechanism (hybrid commitment testing detects drift), and it is too weak to carry that weight. (a) The sample is described as "randomly sampled" but is stratified 20/30 successful; drift detection matters most on failing trajectories, where recall determines whether the loop rescues runs the baseline loses, so this sample cannot measure the mechanism where the claimed gain is earned. (b) Annotation appears single-annotator with no adjudication or inter-rater statistic, and the counting rule is lenient (any one confirmed diagnosis per trajectory = detected). (c) With n=13 drift cases, recall 76.9% has a binomial CI of roughly 46–95% — statistically compatible with a mediocre detector. Please expand the annotated set, oversample failing trajectories, add a second annotator with an agreement statistic, and report CIs.
  2. [§B.5, Eq. (5)–(6)] The fast-tester routing decision (Eq. 6) depends on at least eight fixed constants: τ0+=0.65, τ0−=0.55, δ+=0.18, δ−=0.25, η=0.1, πlow=0.5, margin 0.08, and λ=0.8. No sensitivity sweep is reported anywhere. This is load-bearing because the textual scorer, nli-deberta-v3-base, is MNLI-trained and here scores entailment between accessibility-tree serializations and planner-written evidence strings — far outside its training distribution — while λ=0.8 weights this channel heavily. If the main results are fragile to modest perturbations of these gates, the contribution is the thresholds, not the architecture. A sweep over the base gates and λ (reporting success and escalation rate) is needed; ideally also state how the constants were chosen.
  3. [§6.5, Tables 4, 5, 9] Two results complicate the causal attribution to FCU evidence matching. First, Table 9 shows the state-level fast tester escalates 72–77% of states to the slow verifier; the lightweight layer therefore skips the LLM call only about a quarter of the time, and Table 5 shows slow-only testing matches full success (66.9 vs 67.6; 68.8 vs 68.8) while saving only ~20% time. The drift detection itself thus appears to be done almost entirely by the slow LLM verifier, with the FCU evidence/scoring layer contributing efficiency rather than detection quality. Second, in Table 4 the "w/o falsifiable plan" variant (skills+repair, no FCUs) still reaches 60.4/61.7/71.2 — above ColorBrowserAgent on all three domains — so a substantial part of the 13.8% gain comes from the memory machinery, not the falsifiable-commitment mechanism. The abstract's framing should be revised to attribute gains accordingly, o
  4. [§6.2–6.3, Tables 2–3] Table 2 reports averages over three runs but no per-run variance or significance assessment. The headline 65.3% vs 57.4% is a ~7.9pp gap over 621 tasks; whether this exceeds run-to-run variance under temperature-1 decoding is not established. This matters most for the smaller domains (Cross, n=37, where the relative gain is largest at 23%) and for WebChoreArena (Table 3, 9.6% relative, absolute gap 2.8pp, no repetition protocol stated). Please report per-run standard deviations and a paired significance test over tasks or runs.
minor comments (5)
  1. [§6.4, Figure 3] Bucketing by the baseline's trajectory length as a horizon proxy introduces a selection effect: tasks on which the baseline fails often have longer (looping) trajectories, which can mechanically inflate the relative-gain numbers in the long buckets. A length proxy independent of the evaluated systems (e.g., human demonstration length or minimum steps) would be more convincing.
  2. [Table 5] The per-domain numbers (219s vs 281s Shopping; 529s vs 648s Admin) don't obviously map to the stated "464.5s to 374.0s, 19.5%" average; please clarify the averaging. Also note the 19.5% saving coexists with 72–77% escalation in Table 9 — worth reconciling in the text.
  3. [Figure 2] "Sever drift" should be "Severe drift". The abstract and §1 describe WebArena gains well, but the paper would benefit from a short limitations paragraph on Map-domain validator instability (currently only in §6.2).
  4. [§A.1, Table 6] The absence of numerical comparison with SkillTracer and ContractSkill is justified by implementation availability, but since Table 6 positions them as the closest methods (EDD+AGR), a qualitative discussion of expected overlap — e.g., what FCUs add beyond contract-based skills — would strengthen the novelty positioning.
  5. [§C.4] Table 10's sampling is described as "randomly sampled" but is stratified by outcome (20/30 successful). Please relabel and describe the sampling frame (which of the 621×3 runs were eligible).

Circularity Check

0 steps flagged

No significant circularity: external WebArena validators and frozen train-only experience libraries keep the reported gains independent of the FCU machinery.

full rationale

FCPAgent is an empirical systems paper, not a first-principles derivation. The central claim is a measured success-rate lift on WebArena (57.4% → 65.3%) and zero-shot WebChoreArena under the benchmarks’ own functional validators V(q, τ), which are independent of FCU confirming/falsifying evidence. Skill and failure-repair libraries are built only from the 191-task training split and frozen at test time; ablations remove planning, testing, or repair rather than refitting the metric to the method. Hybrid-routing thresholds (τ⁰₊, τ⁰₋, λ, etc.) are fixed design hyperparameters, not parameters fitted to the reported test success and then re-presented as predictions. Related-work citations (WebDART, ColorBrowserAgent, ContractSkill, etc.) are used for comparison, not as load-bearing uniqueness theorems that force the result. No equation equates the claimed gain to an input by construction. Weaknesses in mechanistic evidence (small author-annotated Table 10, threshold sensitivity) are correctness/robustness concerns, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 4 invented entities

This is an engineering/empirical agent paper. Load-bearing premises are domain assumptions about web tasks and LLM planners, plus many hand-chosen routing and retrieval hyperparameters—not physical free parameters. Invented entities are software abstractions (FCU, libraries, hybrid tester), not new natural kinds.

free parameters (4)
  • Hybrid tester base gates τ⁰₊, τ⁰₋, margins δ⁺/δ⁻, π_low, η, uncertainty margin = τ⁰₊=0.65, τ⁰₋=0.55, δ⁺=0.18, δ⁻=0.25, π_low=0.5, η=0.1
    App. B.5 sets τ⁰₊=0.65, τ⁰₋=0.55, δ⁺=0.18, δ⁻=0.25, π_low=0.5, η=0.1, uncertainty 0.08; these hand-chosen thresholds control when cheap matching escalates to LLM diagnosis and thus affect measured robustness/efficiency.
  • Evidence fusion weight λ = 0.8
    State-level score α± mixes NLI and image–text matching with λ; main setup uses λ=0.8, biasing textual entailment.
  • Retrieval depths K skills and P failure cases; BM25/embedding mix = K=3, P=2; σ_S=0.4/0.6; σ_F=0.5/0.5
    K=3, P=2; skill score 0.4 BM25 + 0.6 embedding; failure score 0.5 query + 0.5 diagnosis. These control experience conditioning of planning and repair.
  • Max interaction length and decoding temperature = B=20, temperature=1
    Budget B=20 steps and temperature 1 for all methods; horizon and stochasticity of reported success depend on these choices.
axioms (6)
  • domain assumption Final task feedback is delayed/sparse, so intermediate confirming and falsifying evidence is necessary to judge plan validity online.
    Stated in §4 Problem Formulation and motivates the entire FCU interface.
  • domain assumption LLM planners can generate useful task-specific E⁺/E⁻ and confidences κ from query, state, skills, and failure memories.
    §5.1 commitment generation; without this, testing has nothing reliable to match.
  • domain assumption Violations can be attributed to execution-, skill-, or planning-level scopes, and the smallest adequate revision is preferable.
    §5.3 scope-aware repair hierarchy; underpins multi-scope MSFA/AGR claims in Table 1.
  • domain assumption Offline skills and failure-repair cases distilled from a fixed 191-task training split remain valid and non-leaking for the 621-task test split and zero-shot WebChoreArena.
    §6.1 and App. B.6; libraries frozen at test time; fairness claim vs baselines with same data budget.
  • domain assumption Lightweight NLI (DeBERTa) and ITM (SigLIP2) scores are informative routing signals for commitment support/contradiction on accessibility trees and screenshots.
    §5.2 hybrid testing equation for α±_t; efficiency claims rest on this.
  • ad hoc to paper TOTE/goal-directed testing (Miller, Galanter, Pribram 1960) is an appropriate conceptual model for web plan steps.
    Used as inspiration in §1 and §3; not derived, but frames falsifiable commitments.
invented entities (4)
  • Falsifiable Commitment Unit (FCU) independent evidence
    purpose: Represent each plan step as (g, s, E⁺, E⁻, κ) so runtime evidence can support proceed vs repair.
    Definition 1 §4; core abstraction of the method.
  • Hybrid commitment tester (fast evidence matching + slow LLM diagnosis) independent evidence
    purpose: Check actions/states against FCU evidence cheaply and escalate only completion-like, risky, or ambiguous cases.
    §5.2 and App. B.5; operationalizes online testing.
  • Scope-aware repairer with failure-repair library independent evidence
    purpose: Localize falsification to execution/skill/plan (or restart) and revise the smallest adequate scope using retrieved failure cases.
    §5.3; ties diagnosis to recovery.
  • Skill library and failure-repair library distilled from trajectories independent evidence
    purpose: Provide procedural priors and negative guardrails for FCU generation and repair.
    §5.1, App. B.3–B.4; standard memory pattern specialized to FCU evidence writing.

pith-pipeline@v1.2.0-grok45-kimik3 · 23467 in / 4311 out tokens · 84617 ms · 2026-07-31T22:04:22.486209+00:00 · methodology

0 comments
read the original abstract

Long-horizon web agents often go off track before final failure: a trajectory can remain locally plausible even after the current state, reused skill, or plan assumption no longer supports the user instruction. Existing agents can plan, reflect, or reuse experience, but their plans rarely specify the evidence under which an active step should still be trusted. We propose FCPAgent, a falsifiable commitment planning framework for robust long-horizon web agents. FCPAgent represents each plan step as a Falsifiable Commitment Unit (FCU): a subgoal grounded in a reusable skill, together with confirming evidence, falsifying evidence, and a confidence score. Execution is organized as a plan-test-repair loop. The hybrid commitment testing module checks candidate actions before they modify the browser and checks observations after execution; for efficiency, it combines lightweight evidence matching with LLM-based diagnostic verification. When evidence falsifies a commitment, scope-aware repair localizes the contradiction to the execution, skill, or planning level and revises the smallest adequate part. On WebArena, FCPAgent achieves a 13.8% relative improvement in average success over the strongest baseline, with especially large gains on long-horizon tasks.

Figures

Figures reproduced from arXiv: 2607.24167 by Guangyi Liu, Huan Zhao, Quanming Yao.

Figure 1
Figure 1. Figure 1: An off-track execution example. An ordinary plan [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FCPAgent casts web-agent execution as online testing of falsifiable commitments. Hybrid commitment testing checks actions and observations; scope-aware repair revises the smallest contradicted component. falsifiable plan step associates its intended operation with ob￾servable evidence that either confirms or falsifies its continued validity. In the GitLab example, entering group-related pages without expos… view at source ↗
Figure 3
Figure 3. Figure 3: Success rate (%) by task length. The strongest gains [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: GitLab success rate as the amount of training [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A WebArena-Shopping case study showing how [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of a falsifiable commitment sequence. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example record from the skill library. Before planning, FCPAgent may retrieve failure-repair cases using only the task query, providing high-level lessons that help the planner specify stronger falsifiers and avoid known failure modes. When a commitment is falsified during execution, the repair controller retrieves cases using both the task query and the current violation diagnosis. The retrieved records a… view at source ↗
Figure 8
Figure 8. Figure 8: Example record from the failure-repair library. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Efficiency and success rate on Shopping tasks under two runtime measures. The y-axis reports task success rate, while [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of test-time cost across the modules of [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of repair strategies selected by the scope-aware repairer. [PITH_FULL_IMAGE:figures/full_fig_p017_11.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

35 extracted references · 2 linked inside Pith

  1. [1]

    Zhou, Shuyan and Xu, Frank F. and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Ou, Tianyue and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , booktitle =. 2024 , url =

    2024

  2. [2]

    Le Sellier De Chezelles, Thibault and Gasse, Maxime and Drouin, Alexandre and Caccia, Massimo and Boisvert, L. The. arXiv preprint arXiv:2412.05467 , year =

    Pith/arXiv arXiv

  3. [3]

    Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in

    Yang, Jianwei and Zhang, Hao and Li, Feng and Zou, Xueyan and Li, Chunyuan and Gao, Jianfeng , journal =. Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in. 2023 , url =

    2023

  4. [4]

    2023 , url =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , url =

    2023

  5. [5]

    2023 , url =

    Deng, Xiang and Gu, Yu and Zheng, Boyuan and Chen, Shijie and Stevens, Samuel and Wang, Boshi and Sun, Huan and Su, Yu , booktitle =. 2023 , url =

    2023

  6. [6]

    2022 , url =

    Yao, Shunyu and Chen, Howard and Yang, John and Narasimhan, Karthik , booktitle =. 2022 , url =

    2022

  7. [7]

    2024 , url =

    Koh, Jing Yu and Lo, Robert and Jang, Lawrence and Duvvur, Vikram and Lim, Ming Chong and Huang, Po-Yu and Neubig, Graham and Zhou, Shuyan and Salakhutdinov, Ruslan and Fried, Daniel , booktitle =. 2024 , url =

    2024

  8. [8]

    International Conference on Machine Learning , year =

    Agent Workflow Memory , author =. International Conference on Machine Learning , year =

  9. [9]

    2025 , url =

    Yang, Ke and Liu, Yao and Chaudhary, Sapana and Fakoor, Rasool and Chaudhari, Pratik and Karypis, George and Rangwala, Huzefa , booktitle =. 2025 , url =

    2025

  10. [10]

    International Conference on Learning Representations , year =

    Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation , author =. International Conference on Learning Representations , year =

  11. [11]

    2026 , url =

    Prabhu, Viraj and Dai, Yutong and Fernandez, Matthew and Ramakrishnan, Krithika and Gu, Jing and Luo, Yanqi and Savarese, Silvio and Xiong, Caiming and Li, Junnan and Chen, Zeyuan and Xu, Ran , booktitle =. 2026 , url =

    2026

  12. [12]

    2026 , url =

    Zhou, Jiamu and Wang, Jihong and Zhang, Weiming and Liu, Weiwen and Zhang, Zhuosheng and Lou, Xingyu and Zhang, Weinan and Deng, Huarong and Wang, Jun , journal =. 2026 , url =

    2026

  13. [13]

    2026 , url =

    Wang, Hanyu and Cao, Yuanpu and Lin, Lu and Chen, Jinghui , journal =. 2026 , url =

    2026

  14. [14]

    2025 , url =

    Yang, Jingbo and Hou, Bairu and Wei, Wei and Chang, Shiyu and Bao, Yujia , journal =. 2025 , url =

    2025

  15. [15]

    2026 , url =

    Lu, Zijian and Zuo, Yiping and Nie, Yupeng and He, Xin and Fan, Weibei and Dai, Chen , journal =. 2026 , url =

    2026

  16. [16]

    arXiv preprint arXiv:2603.07024 , year =

    Enhancing Web Agents with a Hierarchical Memory Tree , author =. arXiv preprint arXiv:2603.07024 , year =

    arXiv

  17. [17]

    Advances in Neural Information Processing Systems , year =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems , year =

  18. [18]

    Advances in Neural Information Processing Systems , year =

    Self-Refine: Iterative Refinement with Self-Feedback , author =. Advances in Neural Information Processing Systems , year =

  19. [19]

    Advances in Neural Information Processing Systems , year =

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author =. Advances in Neural Information Processing Systems , year =

  20. [20]

    Transactions on Machine Learning Research , year =

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author =. Transactions on Machine Learning Research , year =

  21. [21]

    International Conference on Machine Learning , year =

    Learning Transferable Visual Models From Natural Language Supervision , author =. International Conference on Machine Learning , year =

  22. [22]

    Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics , year =

    A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference , author =. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics , year =

    2018

  23. [23]

    2026 , howpublished =

    2026

  24. [24]

    2024 , url =

    He, Hongliang and Yao, Wenlin and Ma, Kaixin and Yu, Wenhao and Dai, Yong and Zhang, Hongming and Lan, Zhenzhong and Yu, Dong , booktitle =. 2024 , url =

    2024

  25. [25]

    2024 , series =

    Zheng, Boyuan and Gou, Boyu and Kil, Jihyung and Sun, Huan and Su, Yu , booktitle =. 2024 , series =

    2024

  26. [26]

    2024 , url =

    Lai, Hanyu and Liu, Xiao and Iong, Iat Long and Yao, Shuntian and Chen, Yuxuan and Shen, Pengbo and Yu, Hao and Zhang, Hanchen and Zhang, Xiaohan and Dong, Yuxiao and Tang, Jie , booktitle =. 2024 , url =

    2024

  27. [27]

    2025 , url =

    Qi, Zehan and Liu, Xiao and Iong, Iat Long and Lai, Hanyu and Sun, Xueqiao and Sun, Jiadai and Yang, Xinyue and Yang, Yu and Yao, Shuntian and Xu, Wei and Tang, Jie and Dong, Yuxiao , booktitle =. 2025 , url =

    2025

  28. [28]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    Explorer: Scaling Exploration-driven Web Trajectory Synthesis for Multimodal Web Agents , author =. Findings of the Association for Computational Linguistics: ACL 2025 , pages =. 2025 , url =

    2025

  29. [29]

    2026 , url =

    Li, Yuyang and Dou, Yiran and Shao, Jie-Jing and Lyu, Yueming and Tsang, Ivor and Yin, Haiyan , booktitle =. 2026 , url =

    2026

  30. [30]

    and Jin, Xiaolong and Wang, Zora Zhiruo and Gandhi, Apurva and Song, Yueqi and Gu, Yu and Srinivasa, Jayanth and Liu, Gaowen and Neubig, Graham and Su, Yu , journal =

    Zheng, Boyuan and Fatemi, Michael Y. and Jin, Xiaolong and Wang, Zora Zhiruo and Gandhi, Apurva and Song, Yueqi and Gu, Yu and Srinivasa, Jayanth and Liu, Gaowen and Neubig, Graham and Su, Yu , journal =. 2025 , url =

    2025

  31. [31]

    arXiv preprint arXiv:2504.06821 , year =

    Inducing Programmatic Skills for Agentic Tasks , author =. arXiv preprint arXiv:2504.06821 , year =

    Pith/arXiv arXiv

  32. [32]

    , author =

    The empirical case for two systems of reasoning. , author =. Psychological bulletin , volume =. 1996 , publisher =

    1996

  33. [33]

    1960 , publisher =

    Plans and the Structure of Behavior , author =. 1960 , publisher =

    1960

  34. [34]

    2025 , url =

    Miyai, Atsuyuki and Zhao, Zaiying and Egashira, Kazuki and Sato, Atsuki and Sunada, Tatsumi and Onohara, Shota and Yamanishi, Hiromasa and Toyooka, Mashiro and Nishina, Kunato and Maeda, Ryoma and Aizawa, Kiyoharu and Yamasaki, Toshihiko , journal =. 2025 , url =

    2025

  35. [35]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    Devil’s advocate: Anticipatory reflection for llm agents , author =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    2024