Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Branch-and-Browse claims that an LLM web agent can be made dramatically more successful and faster by combining subtask-aware tree search with per-page action memory, achieving 35.8% success on WebArena versus 19.2% for tree search.

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 09:15 UTC pith:YJRKFWYV

load-bearing objection A plausible integration of tree search, memory, and replay for web agents, but the headline gain over Tree Search rests on a baseline that was not re-run in the same harness. the 4 major comments →

arxiv 2510.19838 v2 pith:YJRKFWYV submitted 2025-10-18 cs.AI cs.CLcs.LG

Branch-and-Browse: Efficient and Controllable Web Exploration with Tree-Structured Reasoning and Action Memory

Shiqi He , Yue Cui , Xinyu Ma , Yaliang Li , Bolin Ding , Mosharaf Chowdhury This is my paper
classification cs.AI cs.CLcs.LG
keywords web agentsLLM agentstree searchsubtask decompositionaction memoryWebArenaweb explorationbacktracking
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 argues that LLM web agents fail at long-horizon tasks because they either reason in a straight line with no way to back up, or explore states so coarsely that every branch redoes the same contextual work. Its answer, Branch-and-Browse, decomposes the instruction into subtasks, explores pages as a tree, and attaches a memory to every page URL that records which actions were tried and what happened. On the WebArena benchmark, the framework reports 35.8% task success, up from 19.2% for the published tree-search baseline, while cutting average time per successful task from 20.8 to 12.4 minutes. A sympathetic reader would take this as evidence that structured, memory-rich exploration—not simply more model calls—is what makes web agents practical.

Core claim

Branch-and-Browse establishes that a web agent can be modeled as a subtask-aware tree search: each node is a visited web page, each edge is an elementary action such as click, type, or navigate, and the search proceeds in reason–act–evaluation cycles under a subtask manager that advances when a success predicate is met. To make branching affordable, the framework adds a page action memory that maintains per-URL records of objectives, progress summaries, reason–action history, DOM snapshots, and attempted actions with success flags, shared across branches and sessions. It also introduces two acceleration mechanisms: nearest-URL state replay, which restores an earlier state by loading the clos

What carries the argument

The central object is the exploration tree combined with page-level action memory. Nodes are web pages; edges are browser actions. The mechanism that carries the argument is the page action memory: every visited URL stores the active subtask, a progress summary, the sequence of reasoning and actions with element references and results, and a categorized log of attempted actions. This memory feeds two accelerators—nearest-URL replay (reload the closest cached page, then replay only the actions needed to reconstruct a target state) and background reasoning (an offline pass that scores unexplored nodes and pre-expands deterministic link clicks)—so branches share context instead of re-exploring

Load-bearing premise

The headline gains assume the published Tree Search baseline (19.2% success, 20.8 minutes per task) was measured under the same model, prompts, and action set as Branch-and-Browse; the paper cites published results without re-running that baseline in its own harness.

What would settle it

Re-run the published Tree Search algorithm on the same 812 WebArena tasks with the same model (gpt-4o-2023-12-12), the same exploration budget, and the same browser-automation interface as Branch-and-Browse; if the baseline also reaches roughly 35% success or a similar 12.4-minute average, the claimed architectural advantage disappears.

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

If this is right

  • Long-horizon web tasks become feasible for a single agent run, because early mistakes are recovered by branching and backtracking instead of restarting from scratch.
  • Exploration budgets scale: under the same 10-step budget, deeper and wider searches improve success because replay and background reasoning keep added costs sublinear.
  • Page-level action memory turns a one-session agent into one that accumulates reusable knowledge about a site, making repeated visits faster.
  • Search-based and policy-based agent designs are complementary; combining them could plausibly exceed the best published numbers on WebArena.
  • The 40.4% reduction in time per successful task indicates tree search can be made cost-competitive with simple linear agents.

Where Pith is reading between the lines

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

  • A testable prediction the paper does not run: if page action memory is kept across sessions on the same WebArena site, later task success should rise even with the same exploration budget.
  • Background reasoning is effectively model-based lookahead; a distilled value model trained offline could replace the extra LLM calls and cut cost further.
  • The d=0, b=1 configuration (ReAct plus memory) already beats the 19.2% tree-search baseline, suggesting memory alone may drive much of the gain—an ablation isolating memory would test this.
  • If the method generalizes to live websites, the privacy and consent safeguards the authors recommend become necessary product requirements, not optional extras.

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. This paper presents Branch-and-Browse, a web-agent framework that combines subtask-aware tree-structured exploration, page-level action memory, nearest-URL replay, and background reasoning. On WebArena (812 tasks), it reports a 35.8% success rate and a 40.4% reduction in average completion time per successful task relative to a published Tree Search baseline (19.2% SR; 20.8 → 12.4 minutes). Ablations show that background reasoning contributes more to time savings than replay, and sensitivity studies examine depth and branching factor. The central claim is that the framework improves both success and efficiency over linear ReAct-style and prior tree-search methods.

Significance. The framework design is plausible and clearly described; the ablation and sensitivity analyses are useful. If the headline numbers hold under controlled re-evaluation, Branch-and-Browse would be a solid contribution to LLM-based web agents, demonstrating that fine-grained subtasks and memory can make tree search tractable. However, the evidence as presented is not yet sufficient: the main comparator is a published number rather than a same-harness re-run, and the efficiency metric is computed on different successful-task subsets without variance. The contribution is promising but requires a more rigorous evaluation before the claims can be accepted. The paper honestly lists limitations (single-browser session, no parallel branch exploration), which we view as acceptable caveats for a first report but important for interpreting scalability.

major comments (4)
  1. [Table 2 / §4.1–4.2] The headline comparison to Tree Search is not apples-to-apples. The Tree Search row is labeled as a published baseline in the upper block, and §4.1 only describes the authors' own implementation. There is no statement that Tree Search was re-run in the same AgentScope + Playwright MCP harness with the same GPT-4o snapshot, observation format, prompts, action space, or exploration budget. The 16.6-point SR difference and the 8.4-minute time reduction therefore conflate architectural improvements with implementation differences. Please re-run Tree Search (and at least one ReAct-style baseline) in the same harness, or provide strong evidence that the published number is directly comparable.
  2. [Figure 4 / §4.3] The efficiency comparison is confounded. The caption states 'Failed tasks are excluded,' so Branch-and-Browse and Tree Search are averaged over different subsets of the 812 tasks. A method that succeeds only on easy tasks would appear faster. Also, no standard errors or repeated seeds are reported, although LLM-based agents are stochastic. Please report time on the intersection of tasks that both methods solve, or include all tasks with a sensible cap, and provide variance over at least a few runs.
  3. [Abstract / §4.2] The phrase 'state-of-the-art methods' is not supported. Table 2 lists several baselines with higher success rates (e.g., SteP 33.3%, AWM 35.5%, API Hybrid Agent 38.9%, WebPilot 37.2%, AgentOccam-Judge 45.7%). The 35.8% success rate of Branch-and-Browse is not state of the art on SR. Please qualify the claims as 'among search-based methods' or 'relative to a published Tree Search baseline,' both in the abstract and in the introduction/conclusion.
  4. [§4.4, Table 3] The statement 'All experiments are conducted under the same overall search budget as the default setting (equivalent to the number of ReAct steps performed by the baseline)' is ambiguous. It is unclear whether the budget is the 10-step exploration budget from §4.1, and how the baseline's ReAct steps are counted. Please specify the budget in node expansions or LLM calls, and clarify the relationship with the depth and branching factor settings.
minor comments (6)
  1. [§3.1 heading] Typo: 'Fined-Grained Structured Planning' should be 'Fine-Grained Structured Planning'.
  2. [Introduction, para. 2] The sentence 'search-based methods such as Tree Search ... expand multiple status' should read 'expand multiple states'.
  3. [Table 3 header] The column header 'Time (↑)' is confusing: lower time is better, so the arrow should point down or be removed. Also clarify in the caption how time is measured (per successful task, per all tasks, median, mean?).
  4. [§4.1 Implementation Details] The model identifier 'gpt-4o-2023-12-12' appears implausible; GPT-4o versions typically use a 2024 date (e.g., gpt-4o-2024-05-13). Please verify and correct.
  5. [Table 2 baseline label] The baseline 'BrowserGym (Drouin et al., 2024)' appears inaccurate: Drouin et al. 2024 is the WorkArena paper, and BrowserGym is an environment, not an agent. Please correct the citation/label and clarify where the 23.5% number comes from.
  6. [Table 2 / evaluation set] The 'Multisite' column is not defined in the text. If it refers to tasks spanning multiple WebArena sites, please say so and report the number of tasks per site for context.

Circularity Check

0 steps flagged

No significant circularity: the headline results are externally measured benchmark outcomes, and the author-overlap citations are not load-bearing.

full rationale

Branch-and-Browse is an empirical systems paper rather than a formal derivation. The headline claim—35.8% success rate and 12.4 minutes per successful task versus Tree Search's 19.2% and 20.8 minutes—is a measured outcome on the WebArena benchmark, with success determined by WebArena's external programmatic reward. No reported quantity that is framed as a prediction is defined in terms of the method's own fitted parameters or scoring rules. The components (subtask decomposition, tree-structured exploration, nearest-URL replay, background reasoning, and page action memory) are presented as mechanisms and evaluated through ablations and sensitivity sweeps; their performance is not equivalent by construction to their inputs. The only author-overlap citations—AgentScope (Gao et al., 2025) and the Cui et al. related-work entries—appear in implementation or surveying contexts and are not load-bearing: AgentScope is used as an orchestration harness, not as evidence for the efficiency or success-rate claims. The appended Limitations section concedes scope constraints (no parallel multi-branch execution, no policy-search integration) but does not admit or create a circular step. The main validity concerns—that the Tree Search baseline is taken from published work rather than re-run in the same harness, and that hyperparameters are selected on the same benchmark—are experimental-control and generalization issues, not circularity. Therefore no circularity steps are identified.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper introduces no physical or mathematical entities; the free parameters are search hyperparameters tuned on the WebArena benchmark. The axioms are the empirical comparability and metric choices that the central claims depend on.

free parameters (3)
  • search depth d = 5 (default)
    Chosen via sensitivity analysis on WebArena (Table 3); d=5,b=5 gives the best reported SR of 35.8%.
  • branching factor b = 5 (default)
    Chosen via sensitivity analysis on WebArena (Table 3); higher b improves SR but increases runtime.
  • exploration budget = 10 steps per task
    Set as default in §4.1 with no derivation; controls how many tree nodes are expanded.
axioms (5)
  • domain assumption WebArena success rate is a valid proxy for web-agent quality.
    The whole evaluation treats the WebArena benchmark's programmatic reward as ground truth (§4.1).
  • domain assumption Published baselines are comparable to runs in the authors' own harness.
    Table 2 compares Branch-and-Browse to published numbers without same-harness re-runs of baselines (§4.1, Table 2).
  • domain assumption Time per successful task, excluding failed tasks, is a meaningful efficiency metric.
    The 40.4% runtime reduction is measured only on successful trajectories (§4.3, Figure 4).
  • domain assumption The LLM's node scoring, subtask decomposition, and background-reasoning judgments are reliable without calibration.
    GPT-4o is used both as the actor and as the evaluator/prioritizer (§3.1, §3.2), yet no accuracy checks on these scores are provided.
  • domain assumption A cached URL plus replayed actions accurately reconstructs an earlier browser state.
    Nearest-URL replay assumes intermediate state (forms, tabs, dynamic DOM) can be restored from a saved URL and remaining actions (§3.2).

pith-pipeline@v1.3.0-alltime-deepseek · 12668 in / 12932 out tokens · 109470 ms · 2026-08-04T09:15:09.987193+00:00 · methodology

0 comments
read the original abstract

Autonomous web agents powered by large language models (LLMs) show strong potential for performing goal-oriented tasks such as information retrieval, report generation, and online transactions. These agents mark a key step toward practical embodied reasoning in open web environments. However, existing approaches remain limited in reasoning depth and efficiency: vanilla linear methods fail at multi-step reasoning and lack effective backtracking, while other search strategies are coarse-grained and computationally costly. We introduce Branch-and-Browse, a fine-grained web agent framework that unifies structured reasoning-acting, contextual memory, and efficient execution. It (i) employs explicit subtask management with tree-structured exploration for controllable multi-branch reasoning, (ii) bootstraps exploration through efficient web state replay with background reasoning, and (iii) leverages a page action memory to share explored actions within and across sessions. On the WebArena benchmark, Branch-and-Browse achieves a task success rate of 35.8\% and reduces execution time by up to 40.4\% relative to state-of-the-art methods. These results demonstrate that Branch-and-Browse is a reliable and efficient framework for LLM-based web agents.

Figures

Figures reproduced from arXiv: 2510.19838 by Bolin Ding, Mosharaf Chowdhury, Shiqi He, Xinyu Ma, Yaliang Li, Yue Cui.

Figure 1
Figure 1. Figure 1: Comparison of web agent strategies. The left [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of the Branch-and-Browse framework. Given an example task (Zhou et al., 2023), the subtask manager first decomposes the goal into three subtasks: (1) identify the page providing the product sales report, (2) filter data for the Q1 2022 period, and (3) aggregate brand sales to select the top-1 brand. Each subtask is explored through a tree-structured exploration consisting of iterative Reason–Act–E… view at source ↗
Figure 3
Figure 3. Figure 3: Example of reason context in branch-and-browse. During tree exploration, the agent on Page 3 reasons to execute a web interaction (browser_click), acts on the referenced element, and evaluates the outcome (score = 0), leading to Page 4. The page-level context records the task objective, progress summary, reason–action history, and page snapshot. The Page 3 Action Memory is updated by appending the new acti… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation results on WebArena (812 tasks): average time per successful task across all sites. Failed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns

    cs.AI 2026-06 unverdicted novelty 6.0

    SkillMigrator reduces LLM-action counts by 8-10% on WebArena and Mind2Web by transferring web skills via layout-matched transferable interaction patterns.

  2. MetaResearcher: Scaling Deep Research via Self-Reflective Reinforcement Learning in Adversarial Virtual Environments

    cs.AI 2026-06 unverdicted novelty 5.0

    MetaResearcher is a proposed multi-component framework for scaling deep research agent training via adversarial virtual worlds, discovery tasks, meta-rewards, and multi-agent collaboration.

Reference graph

Works this paper leans on

51 extracted references · 29 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Deepak Bhaskar Acharya, Karthigeyan Kuppan, and B Divya. 2025. Agentic ai: Autonomous intelligence for complex goals--a comprehensive survey. IEEe Access

    2025

  2. [2]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    Pith/arXiv arXiv 2023

  3. [3]

    Saaket Agashe, Jiuzhou Han, Shuyu Gan, Jiachen Yang, Ang Li, and Xin Eric Wang. 2024. Agent s: An open agentic framework that uses computers like a human. arXiv preprint arXiv:2410.08164

    Pith/arXiv arXiv 2024

  4. [4]

    L \'e o Boisvert, Megh Thakkar, Maxime Gasse, Massimo Caccia, Thibault de Chezelles, Quentin Cappart, Nicolas Chapados, Alexandre Lacoste, and Alexandre Drouin. 2024. Workarena++: Towards compositional planning and reasoning-based common knowledge work tasks. Advances in Neural Information Processing Systems, 37:5996--6051

    2024

  5. [5]

    Chaoran Chen, Bingsheng Yao, Ruishi Zou, Wenyue Hua, Weimin Lyu, Yanfang Ye, Toby Jia-Jun Li, and Dakuo Wang. 2025. Towards a design guideline for rpa evaluation: A survey of large language model-based role-playing agents. arXiv preprint arXiv:2502.13012

    Pith/arXiv arXiv 2025

  6. [6]

    Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Li YanTao, Jianbing Zhang, and Zhiyong Wu. 2024. Seeclick: Harnessing gui grounding for advanced visual gui agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9313--9332

    2024

  7. [7]

    Yue Cui, Liuyi Yao, Zitao Li, Yaliang Li, Bolin Ding, and Xiaofang Zhou. 2025 a . Efficient leave-one-out approximation in llm multi-agent debate based on introspection. arXiv preprint arXiv:2505.22192

    Pith/arXiv arXiv 2025

  8. [8]

    Yue Cui, Liuyi Yao, Shuchang Tao, Weijie Shi, Yaliang Li, Bolin Ding, and Xiaofang Zhou. 2025 b . Enhancing tool learning in large language models with hierarchical error checklists. arXiv preprint arXiv:2506.00042

    Pith/arXiv arXiv 2025

  9. [9]

    Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. 2023. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36:28091--28114

    2023

  10. [10]

    Alexandre Drouin, Maxime Gasse, Massimo Caccia, Issam H Laradji, Manuel Del Verme, Tom Marty, L \'e o Boisvert, Megh Thakkar, Quentin Cappart, David Vazquez, and 1 others. 2024. Workarena: How capable are web agents at solving common knowledge work tasks? arXiv preprint arXiv:2403.07718

    Pith/arXiv arXiv 2024

  11. [11]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv--2407

    2024

  12. [12]

    Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2024. Large language models empowered agent-based modeling and simulation: A survey and perspectives. Humanities and Social Sciences Communications, 11(1):1--24

    2024

  13. [13]

    Dawei Gao, Zitao Li, Yuexiang Xie, Weirui Kuang, Liuyi Yao, Bingchen Qian, Zhijian Ma, Yue Cui, Haohao Luo, Shen Li, and 1 others. 2025. Agentscope 1.0: A developer-centric framework for building agentic applications. arXiv preprint arXiv:2508.16279

    Pith/arXiv arXiv 2025

  14. [14]

    Yu Gu, Kai Zhang, Yuting Ning, Boyuan Zheng, Boyu Gou, Tianci Xue, Cheng Chang, Sanjari Srivastava, Yanan Xie, Peng Qi, and 1 others. 2024. Is your llm secretly a world model of the internet? model-based planning for web agents. arXiv preprint arXiv:2411.06559

    Pith/arXiv arXiv 2024

  15. [15]

    Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. 2023. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856

    Pith/arXiv arXiv 2023

  16. [16]

    Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. 2024. Webvoyager: Building an end-to-end web agent with large multimodal models. arXiv preprint arXiv:2401.13919

    Pith/arXiv arXiv 2024

  17. [17]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    Pith/arXiv arXiv 2024

  18. [18]

    Jihyung Kil, Chan Hee Song, Boyuan Zheng, Xiang Deng, Yu Su, and Wei-Lun Chao. 2024. Dual-view visual contextualization for web navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14445--14454

    2024

  19. [19]

    Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. 2024 a . Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649

    Pith/arXiv arXiv 2024

  20. [20]

    Jing Yu Koh, Stephen McAleer, Daniel Fried, and Ruslan Salakhutdinov. 2024 b . Tree search for language model agents. arXiv preprint arXiv:2407.01476

    arXiv 2024

  21. [21]

    Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, and 1 others. 2024. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5295--5306

    2024

  22. [22]

    Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. 2024. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814

    Pith/arXiv arXiv 2024

  23. [23]

    Laser: Llm agent with state-space exploration for web navigation

    Kaixin Ma, Hongming Zhang, Hongwei Wang, Xiaoman Pan, and Dong Yu. Laser: Llm agent with state-space exploration for web navigation. In NeurIPS 2023 Foundation Models for Decision Making Workshop

    2023

  24. [24]

    Tula Masterman, Sandi Besen, Mason Sawtell, and Alex Chao. 2024. The landscape of emerging ai agent architectures for reasoning, planning, and tool calling: A survey. arXiv preprint arXiv:2404.11584

    Pith/arXiv arXiv 2024

  25. [25]

    Liangbo Ning, Ziran Liang, Zhuohang Jiang, Haohao Qu, Yujuan Ding, Wenqi Fan, Xiao-yong Wei, Shanru Lin, Hui Liu, Philip S Yu, and 1 others. 2025. A survey of webagents: Towards next-generation ai agents for web automation with large foundation models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 6140--6150

    2025

  26. [26]

    Runliang Niu, Jindong Li, Shiqi Wang, Yali Fu, Xiyu Hu, Xueyuan Leng, He Kong, Yi Chang, and Qi Wang. 2024. Screenagent: A vision language model-driven computer control agent. arXiv preprint arXiv:2402.07945

    Pith/arXiv arXiv 2024

  27. [27]

    OpenAI . 2023. GPT-4V(ision) System Card . https://openai.com/index/gpt-4v-system-card/. Accessed: 2025-09-28

    2023

  28. [28]

    Jiayi Pan, Yichi Zhang, Nicholas Tomlin, Yifei Zhou, Sergey Levine, and Alane Suhr. 2024. Autonomous evaluation and refinement of digital agents. arXiv preprint arXiv:2404.06474

    Pith/arXiv arXiv 2024

  29. [29]

    Aske Plaat, Max van Duijn, Niki van Stein, Mike Preuss, Peter van der Putten, and Kees Joost Batenburg. 2025. Agentic large language models, a survey. arXiv preprint arXiv:2503.23037

    arXiv 2025

  30. [30]

    Changle Qu, Sunhao Dai, Xiaochi Wei, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, Jun Xu, and Ji-Rong Wen. 2025. Tool learning with large language models: A survey. Frontiers of Computer Science, 19(8):198343

    2025

  31. [31]

    Kunal Singh, Shreyas Singh, and Mukund Khanna. 2025. Trishul: Towards region identification and screen hierarchy understanding for large vlm based gui agents. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 170--179

    2025

  32. [32]

    Paloma Sodhi, SRK Branavan, Yoav Artzi, and Ryan McDonald. 2023. Step: Stacked llm policies for web actions. arXiv preprint arXiv:2310.03720

    Pith/arXiv arXiv 2023

  33. [33]

    Yueqi Song, Frank Xu, Shuyan Zhou, and Graham Neubig. 2024 a . Beyond browsing: Api-based web agents. arXiv preprint arXiv:2410.16464

    Pith/arXiv arXiv 2024

  34. [34]

    Zirui Song, Yaohang Li, Meng Fang, Yanda Li, Zhenhao Chen, Zecheng Shi, Yuan Huang, Xiuying Chen, and Ling Chen. 2024 b . Mmac-copilot: Multi-modal agent collaboration operating copilot. arXiv preprint arXiv:2404.18074

    Pith/arXiv arXiv 2024

  35. [35]

    Qiushi Sun, Kanzhi Cheng, Zichen Ding, Chuanyang Jin, Yian Wang, Fangzhi Xu, Zhenyu Wu, Chengyou Jia, Liheng Chen, Zhoumianze Liu, and 1 others. 2024. Os-genesis: Automating gui agent trajectory construction via reverse task synthesis. arXiv preprint arXiv:2412.19723

    Pith/arXiv arXiv 2024

  36. [36]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others. 2024 a . A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345

    2024

  37. [37]

    Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. 2024 b . Agent workflow memory. arXiv preprint arXiv:2409.07429

    Pith/arXiv arXiv 2024

  38. [38]

    Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, and 1 others. 2024. Os-atlas: A foundation action model for generalist gui agents. arXiv preprint arXiv:2410.23218

    Pith/arXiv arXiv 2024

  39. [39]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, and 1 others. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences, 68(2):121101

    2025

  40. [40]

    Renjun Xu and Jingwen Peng. 2025. A comprehensive survey of deep research: Systems, methodologies, and applications. arXiv preprint arXiv:2506.12594

    Pith/arXiv arXiv 2025

  41. [41]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    Pith/arXiv arXiv 2025

  42. [42]

    Ke Yang, Jiateng Liu, John Wu, Chaoqi Yang, Yi R Fung, Sha Li, Zixuan Huang, Xu Cao, Xingyao Wang, Yiquan Wang, and 1 others. 2024 a . If llm is the wizard, then code is the wand: A survey on how code empowers large language models to serve as intelligent agents. arXiv preprint arXiv:2401.00812

    Pith/arXiv arXiv 2024

  43. [43]

    Ke Yang, Yao Liu, Sapana Chaudhary, Rasool Fakoor, Pratik Chaudhari, George Karypis, and Huzefa Rangwala. 2024 b . Agentoccam: A simple yet strong baseline for llm-based web agents. arXiv preprint arXiv:2410.13825

    Pith/arXiv arXiv 2024

  44. [44]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

    2023

  45. [45]

    Haopeng Zhang, Philip S Yu, and Jiawei Zhang. 2025 a . A systematic survey of text summarization: From statistical methods to large language models. ACM Computing Surveys, 57(11):1--41

    2025

  46. [46]

    Yao Zhang, Zijian Ma, Yunpu Ma, Zhen Han, Yu Wu, and Volker Tresp. 2025 b . Webpilot: A versatile and autonomous multi-agent system for web task execution with strategic exploration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23378--23386

    2025

  47. [47]

    Zhuosheng Zhang and Aston Zhang. 2023. You only look at screens: Multimodal chain-of-action agents. arXiv preprint arXiv:2309.11436

    Pith/arXiv arXiv 2023

  48. [48]

    Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. 2024. Gpt-4v (ision) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614

    Pith/arXiv arXiv 2024

  49. [49]

    Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, and 1 others. 2023. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854

    Pith/arXiv arXiv 2023

  50. [50]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  51. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...