PaperlayerView exact revision on arXiv

Source-linked AI summary

BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions

Tao Yu, Zhengbo Zhang, Zhiheng Lyu, Junhao Gong, Hongzhu Yi, Xinming Wang, Yuxuan Zhou, Jiabing Yang, Ping Nie, Yan Huang, Wenhu Chen

arXiv:2510.10666v2cs.CLcs.AI

TL;DR

Web agents often rely on external tools to turn dynamic web pages into static text, despite the value of interactive browsing. BrowserAgent uses Playwright-based browser actions, explicit memory, and two-stage training; it reports stronger results than Search-R1 across Open-QA tasks, especially multi-hop QA, with less training data.

  • Problem

    Many web-agent frameworks use external parsing and summarization tools to convert dynamic web information into static text, limiting in-depth page interaction and adding cost.

    Source: §1 · p. 2

  • Method

    BrowserAgent interacts with dynamic pages through human-inspired browser actions and combines explicit memory with a two-stage SFT and rejection fine-tuning pipeline.

    Source: §1 · p. 2, §2.1 · p. 4, §2.4 · p. 5

  • Results

    BrowserAgent outperforms existing methods such as Search-R1 using less training data, with particular strength on multi-hop reasoning tasks.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    The results present BrowserAgent as an efficient, scalable framework for interactive web agents that handle complex reasoning chains.

    Source: §5 · pp. 9–10

    The authors identify cross-website generalization, more intelligent memory, multi-agent collaboration, and continual learning as future-work areas.

    Source: §5 · p. 10

Abstract

from arXiv · show

Efficiently solving real-world problems with LLMs increasingly hinges on their ability to interact with dynamic web environments and autonomously acquire external information. While recent research like Search-R1 and WebDancer demonstrates strong performance in solving web tasks, they heavily rely on additional tools to convert the interactive web environment into static text content. This is in contrast to human browsing behaviors, which involve diverse interactions with the browser, such as scrolling, clicking, and typing. In this paper, we propose BrowserAgent, a more interactive agent that solves complex tasks through human-inspired browser actions. BrowserAgent operates directly on raw web pages via Playwright through a set of predefined browser actions. We adopt a two-stage training (Supervised Fine-Tuning (SFT) and Rejection Fine-Tuning (RFT)) to improve the model's generalization abilities. Despite using significantly less training data than Search-R1, BrowserAgent achieves more competitive results across different Open-QA tasks. Additionally, we introduce an explicit memory mechanism to store key conclusions across steps, further enhancing the model's reasoning capabilities for long-horizon tasks. Notably, BrowserAgent-7B can achieve around 20\% improvement over Search-R1 on multi-hop QA tasks like HotpotQA, 2Wiki, and Bamboogle. These results indicate that BrowserAgent can serve as a more advanced framework for more interactive and scalable web agents.

1 INTRODUCTION

Prior web-agent frameworks turn dynamic pages into static text with external parsing and summarization tools, limiting direct interaction and adding cost. BrowserAgent instead uses human-inspired browser actions and scalable collection infrastructure, with explicit memory and strong reported Open-QA performance using less training data.

1 INTRODUCTION

  • Prior web-agent frameworks rely on HTML parsing and summarization tools, limiting in-depth interaction with web pages and incurring high cost.The paper identifies these tools as converting dynamic web information into static, short documents.Source: §1 · p. 2
  • BrowserAgent uses human-inspired actions such as clicking, typing, and scrolling to interact with dynamic web pages through Playwright.This approach lets agents share a representation space and action set with human browser users.Source: §1 · p. 2
  • Ray-parallelized orchestration raises collection throughput from 1–2 to 50+ episodes per minute on one 32-core server, cutting cost by over an order of magnitude.The infrastructure addresses Playwright’s limited parallelization, which had made large-scale training infeasible.Source: §1 · p. 2
  • BrowserAgent combines raw webpage interaction, atomic browser operations, SFT followed by rejection sampling, and explicit memory for multi-turn reasoning.The paper presents this as a lightweight alternative requiring less data and infrastructure than complex reinforcement-learning strategies.Source: §1 · pp. 2–3
  • With 5.3K training samples, BrowserAgent reports significant outperformance over Search-R1 across a wide range of Open-QA tasks.The paper presents its experimental results as evidence that BrowserAgent is an alternative to existing frameworks.Source: §1 · p. 3

2 BROWSERAGENT

BrowserAgent is trained through real-time browser interactions, using defined browser operations, a multi-turn collection process, and a two-stage SFT–RFT pipeline. Its training data combines basic and multi-hop question answering, while the browser environment and collection infrastructure support those interactive workflows.

2.1 FRAMEWORK DESIGN

  • BrowserAgent defines browser operations for human-like interaction and uses an iterative thinking–summarizing–acting cycle with memory across turns.Its interaction process records reasoning–action pairs and necessary conclusions until an answer is reached or the step limit is exceeded.Source: §2 · p. 3, §2.1 · p. 4

2.2 DATASET SYNTHESIS

  • The 5.3K-trajectory dataset combines 4K NQ and HotpotQA examples with 1.3K additional HotpotQA trajectories for multi-hop reasoning.Basic-capability interactions run up to 6 steps, while challenging data collection runs up to 30 steps.Source: §2.2 · pp. 4–5

2.3 ENVIRONMENT SETTING

  • The browser environment uses Playwright accessibility-tree observations and Ray-based serving, with 64 concurrent browsers deployed on one 32-CPU machine.The authors note potential to scale the framework to larger CPU clusters.Source: §2.3 · p. 5

2.4 TRAINING

  • Training first applies SFT to Qwen2.5-7B-Instruct on 5.3K generated question-answer pairs, then RFT selects correct answers with the most reasoning steps.The RFT data mixes selected reasoning examples with sampled SFT data to retain answer-format behavior while enhancing reasoning.Source: §2.4 · p. 5

3 EXPERIMENTS

The experiments evaluate BrowserAgent on general and multi-hop QA benchmarks using exact match and LLM judgment, against retrieval, fine-tuning, and reinforcement-learning baselines. Results and ablations examine its performance, memory, model size, and interaction-step scaling.

3.1 BENCHMARKS

  • The six benchmarks cover general QA (NQ, PopQA) and multi-hop QA (HotpotQA, 2WikiMultiHopQA, Musique, Bamboogle).Source: §3.1 · p. 6

3.2 EVALUATION METRICS

  • Evaluation combines exact match with consensus LLM judgment, classifying an answer as valid when at least two of three models judge it correct.The paper says exact match can undercount factually correct answers with different surface forms.Source: §3.2 · p. 6

3.3 BASELINES

  • Baselines span direct inference, retrieval-augmented methods, SFT, and Search-R1, with the same pretrained LLM and aligned training-data sources used for fairness.Source: §3.3 · p. 7

3.4 EXPERIMENTAL SETUP

  • Experiments use Qwen2.5-7B-Instruct, a 2022 Wikipedia dump, and a maximum of 30 browser-interaction rounds.Source: §3.4 · p. 7

3.5 EXPERIMENTAL RESULTS

  • Around 20% improvement over Search-R1-Instruct on multi-hop QA tasks; BrowserAgent also consistently outperforms strong baselines across in- and out-of-distribution evaluations.The reported comparison uses Table 2; the authors note that EM alone may not fully capture performance without explicit format-based rewards.Source: §3.5 · p. 7, Table 2 · p. 7

3.6 ABLATION

  • Training the model to retain intermediate conclusions is reported to prevent information loss and reasoning-chain disconnection while improving performance and generalization.The multi-round memory stores key conclusions in a structured format to remove redundant context; the authors report scaling to 30 reasoning steps without exceeding context limits.Source: §3.6 · p. 8, Table 3 · p. 8
  • The 7B model consistently outperforms the 3B model across all evaluation benchmarks.The comparison was limited to 3B and 7B models because of computational-resource constraints.Source: §3.6 · p. 8
  • Performance improves as the maximum interaction-step limit grows beyond around six turns, while average reasoning steps remain stable even with a 30-step limit.At around six turns, performance is slightly below Search-R1.Source: §3.6 · p. 8

4 RELATED WORK

Related work discusses rejection-sampling fine-tuning, agentic information seeking, and web-agent research. BrowserAgent extends these directions with direct browser interaction and a scalable training pipeline.

4.1 REJECTION SAMPLING FINE-TUNING

  • Rejection sampling fine-tuning retains top-scoring candidate outputs and uses them for further training; the paper applies this strategy to browser agents.Source: §4.1 · p. 9

4.2 AGENTIC INFORMATION SEEKING

  • Agentic information seeking lets an agent actively gather and integrate information through browser actions; prior approaches are described as limited in long-context modeling and reliant on external resources.The authors propose incorporating page scrolling into an end-to-end model to support active reading of web content.Source: §4.2 · p. 9

4.3 AGENTIC SEARCH

  • Prior web-agent work has emphasized benchmarks rather than direct training, partly because browser interaction is computationally expensive; the paper presents a scalable two-stage training pipeline to reduce interaction overhead.The cited benchmark-focused works include Zhou et al. (2024), Yao et al. (2023a), Koh et al. (2024), Yoran et al. (2024), and Yang et al. (2025).Source: §4.3 · p. 9

5 CONCLUSION AND FUTURE WORK

BrowserAgent combines Playwright-based web operations with parallel tool serving, explicit memory, and SFT plus RFT. The conclusion reports performance, scalability, and data-efficiency advantages, while naming several directions for future work.

5 CONCLUSION AND FUTURE WORK

  • BrowserAgent uses fine-grained Playwright actions to acquire and integrate information directly from web pages, combined with parallel serving, explicit memory, and two-stage SFT plus RFT.The conclusion says the approach reduces redundant context and scales along complex reasoning chains.Source: §5 · pp. 9–10
  • The authors report advantages in performance, scalability, and data efficiency, including better results than existing methods such as Search-R1 with less training data, especially on multi-hop reasoning.Source: §5 · p. 10
  • Future work will explore more intelligent memory, cross-website generalization, multi-agent collaboration, and continual learning from interaction logs.Source: §5 · p. 10

ETHICS STATEMENT

The appendices state the study’s ethical position and LLM uses, specify browser-agent and answer-evaluation prompts, and illustrate an exact-match scoring mismatch in an NQ case.

ETHICS STATEMENT

  • The authors report using only publicly available benchmark data and no personal or sensitive data, high-risk scenarios, or ethically controversial datasets or models.They state that the research aims to advance web-based agent systems and presents no significant ethical risks.Source: ETHICS STATEMENT · p. 10

A THE USAGE OF LLMS

  • LLMs were used for data generation, experimental result evaluation, paper polishing, visualization generation, and related-work retrieval.The authors say they proofread the paper to check for hallucinated content.Source: A THE USAGE OF LLMS · p. 14

B.3 SYSTEM PROMPT

  • The browser-agent prompt supplies the task, page accessibility tree, open tabs, prior actions, and information already found, then instructs the model to act step by step.The prompt lists page operations, tab management, URL navigation, and a stop action; it also requires usable intermediate conclusions and correctly formatted actions.Source: §B.3 · pp. 14–16
  • A separate evaluator prompt asks whether an answer matches the question’s ground truth and requires a yes-or-no output..Source: §B.3 · p. 15

B.4 CASE STUDY

  • In an NQ case, BrowserAgent answered “King Edward IV” for the ground truth “Edward IV of England,” a semantically equivalent answer that exact match could not match.The example uses three interaction steps.Source: §B.4 · p. 16

Case Study

The case studies illustrate three-step browser interactions on NQ and HotpotQA questions, including answers whose meaning is correct despite differing from the ground-truth wording.

Case Study

  • The NQ example takes three interaction steps and returns “King Edward IV” for “Edward IV of England,” which the authors describe as semantically equivalent but unmatched by exact match.The agent follows a link to the Princes in the Tower article and uses its statement that the brothers were sons of King Edward IV and Elizabeth Woodville.Source: Case Study · p. 17
  • The examples show the agent carrying earlier findings into later steps, including retaining Skin Yard’s U.S. origin while checking Ostava’s origin.The trace records the two origins before giving the combined answer.Source: Case Study · p. 19
  • For the HotpotQA question about Skin Yard and Ostava, the agent concludes that Skin Yard was from the U.S. and Ostava from Bulgaria, matching the ground-truth negative answer.The case description says the agent used three steps to reach its answer.Source: Case Study · pp. 18–19

B.5 FILTERING STRATEGY AND DATA RATIO

The RFT dataset combines filtered reasoning examples with most of the SFT data to preserve the original distribution and answer-format learning.

B.5 FILTERING STRATEGY AND DATA RATIO

  • String matching filters out RFT data that does not conform to the expected format.The stated purpose is to retain examples in the required format.Source: §B.5 · p. 20
  • The final RFT dataset retains a randomly sampled 80% of SFT data to preserve its original distribution and mitigate forgetting of answer formatting.The formatting being preserved was learned during SFT.Source: §B.5 · p. 20
  • The authors add 1,073 selected reasoning examples: 400 NQ samples and 673 HotpotQA samples.These were randomly selected from the filtered original RFT data to maintain consistency in data volume and proportion between stages.Source: §B.5 · p. 20

B.6 EXPERIMENTAL SETUP

The experiments train on synthesized data and evaluate generalization across six public question-answering datasets using exact match and LLM-based judgment.

B.6 EXPERIMENTAL SETUP

  • Training uses 5.3K synthesized examples for two epochs of SFT at 5e-6 and two epochs of RFT at 1e-6 on eight NVIDIA A100 GPUs.The SFT and RFT stages use different learning rates.Source: §B.6 · p. 20
  • Evaluation covers test or validation sets from six publicly available question-answering datasets to assess in-domain and out-of-domain generalization..Source: §B.6 · p. 20
  • The evaluation metrics are exact match and LLM-based judgment, intended to reflect accuracy and real-world applicability..Source: §B.6 · p. 20

B.7 EXPERIMENTAL RESULTS

The experimental-results section includes an ablation study of BrowserAgent-RFT and reports average step counts for that study.

B.7.1 TEST STEP SCALING

  • Table 4 presents results of the ablation study on BrowserAgent-RFT.Source: Table 4 · p. 20
  • Table 5 reports the average number of steps in the BrowserAgent-RFT ablation study.Source: Table 5 · p. 20

B.7.2 TRIVIAQA RESULTS

On TriviaQA, the paper reports a discrepancy between exact-match and LLM-judge scores for its two models, which it attributes to missing format-specific supervision.

B.7.2 TRIVIAQA RESULTS

  • Table 6 shows a significant discrepancy between EM and LLM-judge scores for the two models on TriviaQA.Source: §B.7.2 · p. 20, Table 6 · p. 21
  • The authors attribute the TriviaQA score gap to absent format-specific supervision, which can leave factually correct answers failing exact-match criteria because of formatting.Source: §B.7.2 · p. 20

Related papers

Loading 2510.10666v2…