Pith. sign in

REVIEW 3 major objections 4 minor 22 references

Failure-mined offline RL lifts web agents on all three benchmarks

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

2026-08-04 00:41 UTC pith:GBZ2HTNY

load-bearing objection Solid three-part recipe for compact web agents, but the missing SFT-on-mined control means the RL-specific claim isn't isolated. the 3 major comments →

arxiv 2608.00335 v1 pith:GBZ2HTNY submitted 2026-07-31 cs.AI cs.CLcs.LG

RMSWeb: Reflection, Failure-Mode Mining, and Salvage-DS for Web Agent Reinforcement Learning

Chengbo Liu , Lifang Zhou , Ruijie Yan , Pei Tan , Ao Sun , Haojun Huang , Guichun Hua , Sining Wei
show 3 more authors
Yining Chen Yingying He Yutao Xie
This is my paper
classification cs.AI cs.CLcs.LG
keywords web agentsreinforcement learningoffline RLfailure-mode miningreflection-conditioned collectionaction-semantic rewardcritical-step miningvision-language models
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.

The paper tries to establish that a compact vision-language web agent, after supervised fine-tuning (SFT), can be reliably improved by an offline reinforcement-learning recipe that never needs a live browser during training. The recipe has three parts: reflection-conditioned retries that make trajectory collection yield more successes with shorter paths; a failure-mode miner that extracts only the critical states where a successful run and a failed or detoured run of the same task diverge; and Salvage-DS, a step-level RL objective that scores actions against verified targets with a polarized web-action reward, admits relative updates only for groups with both reward contrast and a competent rollout, and routes rejected groups to an action-only supervised anchor. If correct, the paper would show that both training data and RL signal can be concentrated where a policy actually fails, and that this transfers to live-web success and step efficiency at 8B and 32B scales.

Core claim

The central claim is that the SFT policy's own failures identify where RL should be spent. The paper pairs verified successful trajectories with failed or inefficient ones, finds the last state-aligned divergence whose surrounding evidence matches an assigned failure mode, and keeps only the successful-side observation-action pair as an RL prompt-target. On WebVoyager, Online-Mind2Web, and WebTailBench, this recipe improves over SFT by 3.7/7.0/2.4 points at 8B and 1.2/2.0/7.7 points at 32B; reflection-collected data reduces action steps on solved tasks by up to 19.7%. Under the paper's fixed evaluation harness, the 8B model reaches 86.39% on WebVoyager and 66.44% on Online-Mind2Web.

What carries the argument

The load-bearing object is the failure-mode-mined 'critical divergence': for each task, a verified successful trajectory is contrasted with a failed or detoured one, and the last aligned pre-action state pair is selected where states are comparable, actions differ, and local evidence matches an assigned failure category. Only the successful-side observation and its verified action enter RL. Salvage-DS then scores sampled actions against that verified action with a four-level polarized reward (1.0 exact, 0.7 core-correct, 0.3 right action type with wrong key parameter, 0.0 wrong), applies group-relative advantages only when within-group reward spread and a competent-rollout threshold both pas

Load-bearing premise

Everything rests on the assumption that the LLM-judged, state-aligned 'critical divergence' between a successful run and a failed/detoured run identifies the causally correct action to reinforce; if the divergence search or failure labels are noisy, the RL can reinforce spurious actions and the reported live-web gains would not replicate.

What would settle it

Take the mined 4.5k critical states, shuffle or corrupt the failure labels (or replace the divergence search with randomly chosen same-state action pairs), and retrain under the identical Salvage-DS objective; if the gain over SFT persists within roughly a point, critical-state selection is not the driver. A complementary check is to re-adjudicate the judge's success verdicts on a labeled subset and flip the disputed ones, then see whether the RL gain collapses.

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

If this is right

  • Failure-mode-mined RL adds consistent gains over SFT on all three live-web benchmarks at both 8B and 32B, making offline RL on a compact critical-state corpus a workable alternative to RL over the full trajectory set.
  • Reflection-conditioned collection shortens successful trajectories by 12.9–19.7% under matched decoding, so data collection can buy step efficiency without sacrificing accuracy on shorter-horizon tasks.
  • Ablations show that full-corpus RL, anchor-free RL, and a strong group-relative baseline each trail the full recipe, indicating that both critical-state selection and the salvage anchor contribute independently.
  • The salvage anchor turns rejected groups into supervised teaching on hard states and anti-forgetting on mastered ones, so the algorithm keeps learning even when a sampled group is unsuitable for a relative update.

Where Pith is reading between the lines

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

  • Editorial extension: if the divergence signal transfers, the same success-versus-failure contrast could be applied to other sequential decision domains such as OS-level GUI control, converting sparse task-level outcomes into dense step-level training targets.
  • Editorial extension: the coarse 1.0/0.7/0.3/0.0 reward spacing implies a testable experiment — vary the spacing while holding the contrast and competence gates fixed and observe whether accepted-group ratios or final accuracy move.
  • Editorial extension: the paper itself notes the reward is a proxy for task success, not success itself, so the strongest test of the routing mechanism is swapping the verified-action target for an outcome-based judge in a live browser; if gains persist, the gating mechanism rather than the offline proxy is doing the work.

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

3 major / 4 minor

Summary. RMSWeb proposes a three-part recipe for post-SFT offline RL training of compact web agents: (1) reflection-conditioned trajectory collection to increase yield and reduce detours, (2) failure-mode-guided mining that contrasts successful and failed/detoured executions to select critical states, and (3) Salvage-DS, an offline RL objective combining a polarized action-semantic reward, contrast- and competence-gated dynamic sampling, and an action-only salvage anchor. Experiments on Qwen3-VL-Instruct 8B/32B report pass@1 gains over SFT on WebVoyager, Online-Mind2Web, and WebTailBench (2.4–7.0 points at 8B; 1.2–7.7 points at 32B), step-efficiency gains up to 19.7%, and ablations showing full-corpus RL, anchor-free RL, and DAPO trail the full recipe.

Significance. If the results hold, RMSWeb would provide a practical data-centric recipe for improving compact web agents. The paper's strengths are its same-harness internal comparisons, the ablation set (full-corpus RL, anchor-free, DAPO), the detailed evaluation protocol and contamination checks, and the training-dynamics analysis supporting the stability claims. The collection-side yield analysis and reflection ablation are also well designed. However, the experimental design does not yet isolate the RL objective from the mined data selection, and the external comparability of the WebTailBench numbers is weakened by the judge substitution. These issues are fixable but require additional experiments before the central claims are fully supported.

major comments (3)
  1. [Experiments — Ablation studies (Table 2); Salvage-DS Eq. (2)] The ablations compare the full recipe against full-corpus RL, anchor-free RL, and DAPO, but not against continued supervised fine-tuning on the same 4.5k mined critical-state prompts. Since the paper's central claim is that the RL objective improves over SFT, and the mined data already contains verified action targets, plain cross-entropy on those prompts could reproduce the gains. If it does, the improvement is attributable to mining/selection, not to the Salvage-DS objective. Add an SFT-on-mined control (and ideally SFT on the full corpus) with the same training budget and report the comparison. This is load-bearing because the salvage term in Eq. (2) is itself supervised cross-entropy; without the control the reader cannot tell whether the relative-RL term adds anything beyond data selection.
  2. [Evaluation protocol — Cross-harness check; Table 1; Figure 1] WebTailBench is scored with a substituted WebVoyager-style DOM-aware judge rather than Fara's native verifier. This does not affect the same-harness internal 8B/32B comparisons, but Table 1 and Figure 1 place these numbers beside external WebTailBench scores produced under Fara's native judge. The re-adjudication documents false negatives in the native judge, but it does not establish that the substituted judge is not also more lenient in the opposite direction. At minimum, report one external model (e.g., Fara-7B) scored with the same DOM-aware judge, or otherwise calibrate the two judges, before claiming a 'leading' accuracy–cost frontier on WebTailBench.
  3. [Table 1; Main results] The claim that RL 'improves all three benchmarks at both scales' rests partly on small differences: +1.2 on WebVoyager and +2.0 on Online-Mind2Web at 32B. No variance, confidence intervals, or per-evaluation scores are reported. For pass@1 on 588 or 300 tasks, binomial standard errors are roughly 1–2 percentage points, so the 32B gains on those benchmarks may be within noise. Report the three per-evaluation scores, confidence intervals, and/or a paired significance test (e.g., McNemar on per-task outcomes for SFT vs. RL) to support the cross-scale claim.
minor comments (4)
  1. [Table 1] The table mixes task sets: external WebVoyager rows report F595 while our rows are the 588-task retained set. A footnote caveats this, but the layout invites direct comparison; consider separating external rows visually or adding a task-set column.
  2. [Figure 1 and cost accounting] Repeated-rollout cost curves use the nominal budget C_k = k*C_1 without early stopping, while the paper's deployment is single-run. Clarify how the external curves were reconstructed and state whether the same no-early-stopping convention is used for all systems.
  3. [Equation (2)] The clipping bounds ϵ_lo and ϵ_hi are introduced after the equation rather than before, and it would help to state explicitly whether the min is applied per token or per group. Small clarity fix.
  4. [Table 3] The text says reflection leaves accuracy essentially unchanged on WebVoyager and Online-Mind2Web, but the WebVoyager accuracy is numerically lower with reflection (82.65 vs. 83.28). This is consistent with noise, but given the small differences, consider flagging non-significance explicitly.

Circularity Check

0 steps flagged

No significant circularity: benchmark gains are measured outcomes, not entailed by the training objective or mining criteria.

full rationale

The paper's central claims are empirical: RL on a failure-mode-mined corpus improves measured task success on three live-web benchmarks, and reflection-collected data reduces action steps. These are not derived quantities that reduce to the paper's inputs. Equation (2) defines a training loss that combines a gated relative-RL surrogate, a salvage anchor, and a KL term; it is a training objective, not a prediction mechanism whose outputs are logically implied by its reward levels or gate thresholds. The mined corpus is produced by a separate, benchmark-excluded procedure: the contamination check rejects any training task lexically overlapping the benchmarks, the failure-mode taxonomy comes from a held-out development split, and the RL reward is step-level action matching rather than the benchmark judge. The reported WebVoyager / Online-Mind2Web / WebTailBench numbers are measured under a fixed harness and are not forced by the objective or by the mining thresholds. The paper adopts rather than claims novelty for action matching, citing GUI-R1 and SRL, so no ansatz is smuggled in as an external result. There are no author-overlapping uniqueness theorems or load-bearing self-citations; the cited algorithmic components (PPO, GRPO, DAPO) are external and explicit. The acknowledged limitation that the reward is a proxy for task success is a validity caveat, not a circularity. The absence of an SFT-on-mined-states control and the WebTailBench judge substitution are evaluation-design concerns, but they do not make the derivation circular. Therefore the paper exhibits no significant circularity under the stated criteria.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim depends on hand-set reward levels, gating thresholds, and mining thresholds that are not varied or justified by a tuning procedure. It also depends on the domain assumptions that Browser Use observations, LLM judges, and offline action matching are faithful proxies for live task success. No new physical or conceptual entities (e.g., new forces, particles, or conserved quantities) are introduced.

free parameters (5)
  • Polarized reward levels {1.0, 0.7, 0.3, 0.0} = 1.0 / 0.7 / 0.3 / 0.0
    Hand-specified action-semantic reward levels; spacing ≥0.3 chosen to expose semantic contrast; no sensitivity analysis or tuning procedure given (appendix “Salvage-DS reward details”).
  • Contrast gate threshold ε_σ = 0.10
    std gate accepts only groups with σ_g ≥ 0.10; chosen without reported sensitivity analysis (Salvage-DS RL configuration).
  • Competence gate threshold r_comp = 0.55
    Minimum competent rollout reward; hand-set; no ablation varying it is reported (Salvage-DS RL configuration).
  • Salvage weight λ_salvage = 0.10
    Weight on the action-only NLL anchor; no sensitivity analysis reported (Salvage-DS RL configuration).
  • Mining selection thresholds = S_total ≥ 0.62, S_category ≥ 0.45, state score ≥ 0.60, 30% per-domain cap, 10% technical cap
    Hand-set thresholds for critical-step mining; they directly determine the RL corpus composition and cannot be validated from released artifacts (appendix “Critical-step mining details”).
axioms (4)
  • domain assumption Browser Use interface faithfully renders live web pages and captures sufficient state for action selection.
    All collection, mining, and evaluation rely on a shared Browser Use fork; if observations are lossy, the offline action-matching signal is misaligned with task success (Introduction, Evaluation protocol details).
  • domain assumption LLM judges (GPT-4o) provide valid and stable task-success labels across benchmarks.
    All headline numbers are LLM-judged success; judge errors or bias directly determine reported scores (Evaluation protocol details).
  • domain assumption Step-level action matching against verified ground truth is an adequate proxy reward for task success.
    Acknowledged in Limitations: “the reward is a proxy for task success rather than success itself”; the RL objective optimizes this proxy, not live outcome.
  • ad hoc to paper The failure-mode taxonomy and LLM-assisted divergence labels are accurate enough to select causally important states.
    The taxonomy is derived from a held-out dev split and applied via an LLM judge with confidence thresholds; the paper calls these rules “heuristic and LLM-assisted” in Limitations. If labels are noisy, mining selects spurious contrasts.

pith-pipeline@v1.3.0-alltime-deepseek · 18034 in / 12652 out tokens · 112907 ms · 2026-08-04T00:41:14.416761+00:00 · methodology

0 comments
read the original abstract

Compact web agents can reduce deployment cost, but training them poses challenges in both data collection and post-SFT reinforcement learning (RL). Successful trajectories are expensive to collect and often contain inefficient detours. After supervised fine-tuning (SFT), full trajectory corpora are dominated by routine states; moreover, when group-relative RL is applied to web actions, inadequately designed action-level rewards can yield weak or misleading relative updates, while groups rejected as unsuitable for such updates receive no fallback learning signal. We present RMSWeb, a three-part recipe for Qwen3-VL-Instruct at 8B and 32B. Reflection-conditioned retries increase collection yield and shorten successful trajectories; failure-mode mining concentrates offline RL on critical states exposed by the SFT policy; and Salvage-DS combines an action-semantic polarized reward, contrast-and-competence-gated dynamic sampling, and an action-only anchor for rejected groups. Policies trained with reflection-collected data use up to 19.7% fewer action steps on solved tasks. On WebVoyager, Online-Mind2Web, and WebTailBench, RMSWeb improves over SFT by 2.4-7.0 points at 8B and 1.2-7.7 points at 32B. Our 8B model also achieves the strongest reported Online-Mind2Web result among similarly sized open-weight models in our comparison and a leading reported accuracy-cost trade-off on WebVoyager and WebTailBench, with the caveat that external evaluation protocols differ.

Figures

Figures reproduced from arXiv: 2608.00335 by Ao Sun, Chengbo Liu, Guichun Hua, Haojun Huang, Lifang Zhou, Pei Tan, Ruijie Yan, Sining Wei, Yingying He, Yining Chen, Yutao Xie.

Figure 1
Figure 1. Figure 1: Pass@k accuracy–inference-cost trade-offs on WebVoyager and WebTailBench. Curves show independent-rollout pass@1/2/3. Our 8B model uses token-accounted single-run costs; external single-run accuracies and costs are reported by Fara-7B, with pass@2/3 accuracies reconstructed from its figures. All repeated-rollout costs use the nominal budget Ck = kC1 without early stopping. Evaluation protocols differ acros… view at source ↗
Figure 2
Figure 2. Figure 2: Reflective trajectory collection. A reflection handler distills current trajectory [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Salvage-DS maps each optimization pathology to [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Salvage-DS training flow. For each failure-critical state, the policy samples K actions, applies the polarized reward, and computes group std σg and maximum reward mg = maxi rg,i. Groups passing both gates are accepted into relative RL; rejection by either gate routes the group to the action-only salvage anchor. Refill supplies exactly 64 accepted groups, whereas the salvage branch retains up to 64 rejecte… view at source ↗
Figure 6
Figure 6. Figure 6: Reflective-collection yield on round-1 failure tasks. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Reflection prompt used during reflection-conditioned trajectory collection. It turns one round’s trajectory (and, from [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of the task-query set by categories. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training dynamics at both scales (top: 8B; bottom: 32B). (a) The std-gate rejection share rises as mined states become mastered, while competence-gate rejection falls. Red shades split std-gate rejections into consistently solved (maxi rg,i ≥ rcomp) and consistently unsolved (maxi rg,i < rcomp) groups. (b) Training raises the correct action’s log-probability and reverses the initial correct-minus-wrong gap… 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

22 extracted references · 17 linked inside Pith

  1. [3]

    Deng, X.; et al

    The Browser- Gym Ecosystem for Web Agent Research.arXiv preprint arXiv:2412.05467. Deng, X.; et al

    Pith/arXiv arXiv

  2. [4]

    Gupta,T.;Wolters,P.;Ma,Z.;Sushko,P.;Pang,R.Y.;Llanes, D.; Yang, Y.; Anderson, T.; Zheng, B.; et al

    Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents.arXiv preprint arXiv:2411.06559. Gupta,T.;Wolters,P.;Ma,Z.;Sushko,P.;Pang,R.Y.;Llanes, D.; Yang, Y.; Anderson, T.; Zheng, B.; et al

    Pith/arXiv arXiv

  3. [6]

    In International Conference on Learning Representations

    A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis. In International Conference on Learning Representations. He,H.;etal.2024a.OpenWebVoyager:BuildingMultimodal Web Agents via Iterative Real-World Exploration, Feedback and Optimization.arXiv preprint arXiv:2410.19609. He, H.; et al. 2024b. WebVoyager: Building an End-to-...

    Pith/arXiv arXiv

  4. [7]

    Weblica: Scalable and Reproducible Train- ingEnvironmentsforVisualWebAgents.arXiv:2605.06761. Koh, J. Y.; et al. 2024a. Tree Search for Language Model Agents.arXiv preprint arXiv:2407.01476. Koh, J. Y.; et al. 2024b. VisualWebArena: Evaluating Mul- timodal Agents on Realistic Visual Web Tasks. InAnnual Meeting of the Association for Computational Linguist...

    Pith/arXiv arXiv

  5. [8]

    InInternational Confer- ence on Machine Learning

    WebLINX: Real-World Website Nav- igation with Multi-Turn Dialogue. InInternational Confer- ence on Machine Learning. Luo,R.;etal.2025. GUI-R1:AGeneralistR1-StyleVision- Language Action Model for GUI Agents.arXiv preprint arXiv:2504.10458. Murty, S.; Zhu, H.; Bahdanau, D.; and Manning, C. D

    Pith/arXiv arXiv 2025

  6. [10]

    InFind- ings of the Association for Computational Linguistics: ACL 2025, 22522–22538

    GUI Agents: A Survey. InFind- ings of the Association for Computational Linguistics: ACL 2025, 22522–22538. Vienna, Austria: Association for Com- putational Linguistics. Ou,T.;Xu,F.F.;Madaan,A.;Liu,J.;Lo,R.;Sridhar,A.;Sen- gupta,S.;Roth,D.;Neubig,G.;andZhou,S.2024. Synatra: Turning Indirect Knowledge into Direct Demonstrations for Digital Agents at Scale....

    2025

  7. [11]

    Qi, Z.; et al

    Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents.arXiv preprint arXiv:2408.07199. Qi, Z.; et al

    Pith/arXiv arXiv

  8. [12]

    Schulman, J.; et al

    UI-TARS: Pioneering Automated GUI Interaction with Native Agents.arXiv preprint arXiv:2501.12326. Schulman, J.; et al

    Pith/arXiv arXiv

  9. [14]

    Shinn, N.; et al

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.arXiv preprint arXiv:2402.03300. Shinn, N.; et al

    Pith/arXiv arXiv

  10. [15]

    Sun, Q.; Cheng, K.; Ding, Z.; Jin, C.; Wang, Y.; Xu, F.; Wu, Z.;Jia,C.;Chen,L.;Liu,Z.;Kao,B.;Li,G.;He,J.;Qiao,Y.; and Wu, Z

    Learn-by-Interact: A Data-Centric Framework for Self-Adaptive Agents.arXiv preprint arXiv:2501.10893. Sun, Q.; Cheng, K.; Ding, Z.; Jin, C.; Wang, Y.; Xu, F.; Wu, Z.;Jia,C.;Chen,L.;Liu,Z.;Kao,B.;Li,G.;He,J.;Qiao,Y.; and Wu, Z

    Pith/arXiv arXiv

  11. [16]

    Wei,Z.;etal.2025

    Agent Workflow Memory.arXiv preprint arXiv:2409.07429. Wei,Z.;etal.2025. WebAgent-R1:TrainingWebAgentsvia End-to-End Multi-Turn Reinforcement Learning. InConfer- ence on Empirical Methods in Natural Language Processing. Xie, T.; et al

    Pith/arXiv arXiv 2025

  12. [17]

    arXiv:2602.16855

    Mobile-Agent-v3.5: Multi- platform Fundamental GUI Agents. arXiv:2602.16855. Xu,Y.;Lu,D.;Shen,Z.;Wang,J.;Wang,Z.;Mao,Y.;Xiong, C.; and Yu, T

    arXiv

  13. [18]

    InConference on Language Modeling

    An Illusion of Progress? Assessing the Current State of Web Agents. InConference on Language Modeling. Yan, D.; Zhang, H.; Huzhang, G.; Li, Y.; Wang, Y.; Chen, Q.-G.; Xu, Z.; Luo, W.; et al. 2026.M 2: Dual-Memory Augmentation for Long-Horizon Web Agents via Trajectory Summarization and Insight Retrieval. arXiv:2603.00503. Yang, J.; et al

    arXiv 2026

  14. [20]

    arXiv:2606.02031

    OpenWebRL: DemystifyingOnlineMulti-turnReinforcementLearningfor Visual Web Agents. arXiv:2606.02031. Yao, S.; et al

    Pith/arXiv arXiv

  15. [21]

    Zhang,C.;etal.2024

    DAPO: An Open-Source LLM Re- inforcement Learning System at Scale.arXiv preprint arXiv:2503.14476. Zhang,C.;etal.2024. LargeLanguageModel-BrainedGUI Agents: A Survey.arXiv preprint arXiv:2411.18279. Zheng, B.; Gou, B.; Kil, J.; Sun, H.; and Su, Y

    Pith/arXiv arXiv 2024

  16. [22]

    InPro- ceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, 61349–61385

    GPT- 4V(ision) is a Generalist Web Agent, if Grounded. InPro- ceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, 61349–61385. PMLR. Zhou, S.; et al. 2024a. WebArena: A Realistic Web Envi- ronment for Building Autonomous Agents. InInternational Conference on Learning Representations. Zh...

    Pith/arXiv arXiv 2025

  17. [2017]

    Shao, Z.; et al

    Proximal Policy Optimization Algorithms.arXiv preprint arXiv:1707.06347. Shao, Z.; et al

    Pith/arXiv arXiv

  18. [2021]

    Nguyen, D.; Chen, J.; Wang, Y.; Wu, G.; Park, N.; Hu, Z.; Lyu, H.; Wu, J.; Aponte, R.; Xia, Y.; Li, X.; Shi, J.; Chen, H.; Lai, V

    WebGPT: Browser-Assisted Question-Answering with Human Feedback.arXiv preprint arXiv:2112.09332. Nguyen, D.; Chen, J.; Wang, Y.; Wu, G.; Park, N.; Hu, Z.; Lyu, H.; Wu, J.; Aponte, R.; Xia, Y.; Li, X.; Shi, J.; Chen, H.; Lai, V. D.; Xie, Z.; Kim, S.; Zhang, R.; Yu, T.; Tanjim, M.; Ahmed, N. K.; Mathur, P.; Yoon, S.; Yao, L.; Kveton, B.; Kil, J.; Nguyen, T....

    Pith/arXiv arXiv

  19. [2023]

    Yang, R.; Wu, Q.; Chen, Y.; Bai, H.; Yao, W.; Cheng, H.; Peng, B.; Zhang, H.; Zhang, T.; et al

    Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V.arXiv preprint arXiv:2310.11441. Yang, R.; Wu, Q.; Chen, Y.; Bai, H.; Yao, W.; Cheng, H.; Peng, B.; Zhang, H.; Zhang, T.; et al

    Pith/arXiv arXiv

  20. [2024]

    Open-sourcesoftware,https: //github.com/browser-use/browser-use

    Browser Use: Make Web- sitesAccessibleforAIAgents. Open-sourcesoftware,https: //github.com/browser-use/browser-use. Accessed: 2026-07-

    2026

  21. [2025]

    Bai, H.; et al

    Fara-7B: An Efficient Agentic ModelforComputerUse.arXiv preprint arXiv:2511.19663. Bai, H.; et al

    arXiv

  22. [2026]

    arXiv:2604.08516

    Mol- moWeb: Open Visual Web Agent and Open Data for the Open Web. arXiv:2604.08516. Gur, I.; et al

    Pith/arXiv arXiv