Research map / Multi-agent systems and games / LLM agents
WebDancer: Towards Autonomous Information Seeking Agency
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents WebDancer, a four-stage pipeline (browsing QA synthesis, trajectory rejection sampling, SFT cold start, DAPO-based RL) for training ReAct-style web information-seeking agents, evaluated on GAIA and WebWalkerQA.
Problem
Existing web information-seeking agents are trained on simple, saturated QA datasets (e.g., 2Wiki above 80%) and lack an effective end-to-end training recipe for deep, multi-step research behavior comparable to Deep Research.
Approach
A data-centric, training-stage pipeline: synthesize deep QA via crawled web pages (CrawlQA) and easy-to-hard refinement (E2HQA), sample Short-CoT (GPT-4o) and Long-CoT (QwQ-Plus) ReAct trajectories with rejection sampling and multi-stage filtering, cold-start with SFT that masks observation tokens, then on-policy RL with DAPO using a binary LLM-judge answer reward.
Key results
- WebDancer built on QwQ-32B reaches 51.5% Pass@1 on GAIA text-only validation (vs 37.8% vanilla ReAct) and 47.9% on WebWalkerQA (vs 24.1%).
- WebDancer on Qwen-2.5-32B reaches 40.7% on GAIA, and on Qwen-2.5-7B reaches 31.0%, versus vanilla ReAct 31.0% and 18.4% respectively.
- Best model reports 64.1% Pass@3 on GAIA and 62.0% Pass@3 on WebWalkerQA; a memorization stress test on Qwen-7B yields only 37.4% greedy accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply WebDancer training rather than Vanilla ReAct | QwQ-32B · GAIA | Pass@1 (Avg.): 37.8 -> 51.5 | Helped |
| apply WebDancer training rather than Vanilla ReAct | QwQ-32B · WebWalkerQA | Pass@1 (Avg.): 24.1 -> 47.9 | Helped |
| apply WebDancer training rather than Vanilla ReAct | Qwen-2.5-7B · GAIA | Pass@1 (Avg.): 18.4 -> 31.0 | Helped |
| apply WebDancer training rather than Vanilla ReAct | Qwen-2.5-7B · WebWalkerQA | Pass@1 (Avg.): 24.2 -> 36.0 | Helped |
| apply WebDancer training rather than Vanilla ReAct | Qwen-2.5-32B · GAIA | Pass@1 (Avg.): 31.0 -> 40.7 | Helped |
| apply WebDancer training rather than Vanilla ReAct | Qwen-2.5-32B · WebWalkerQA | Pass@1 (Avg.): 31.9 -> 38.4 | Helped |
| remove SFT cold start before RL (ablation) | QwQ agent · GAIA | Pass@3: only 5% | Hurt |
| train with Long-CoT rather than Short-CoT (ablation) | QwQ-32B · GAIA | Pass@3: 44.66 -> 58.25 | Helped |
| train with Long-CoT rather than Short-CoT (ablation) | Qwen2.5-7B · GAIA | invalid rate: 0.65% -> 21.36% | Hurt |
| mix Short-CoT and Long-CoT training data (ablation) | WebDancer · GAIA | performance | No clear effect |
| use the final filtered trajectory dataset rather than less-filtered variants (ablation) | QwQ agent SFT · GAIA | performance | Helped |
Surprising findings
- RL-only training from QwQ without SFT cold start reaches only about 5% Pass@3 on GAIA.
- Long-CoT trajectories from reasoning models transfer poorly to instruction-tuned models, which show higher invalid rates due to repetition.
Limitations
- Only two tools (search and visit) are integrated.
- RL uses only a small subset (~5,000 pairs) due to compute and stability constraints; rollout cost is high.
- Evaluation focuses on short-answer tasks and a small GAIA text-only validation set (103 questions).
Experimental setup
Datasets: CrawlQA E2HQA MuSiQue Bamboogle PopQA 2Wiki HotpotQA SimpleQA-style seed QA
Benchmarks: GAIA WebWalkerQA BrowseComp BrowseComp-zh
Metrics: Pass@1 Pass@3 Cons@3
Baselines: Vanilla ReAct OpenAI Deep Research Search-o1 WebThinker R1-Searcher Simple DS GPT-4o DeepSeek-R1-671B QwQ-32B RAG
Builds on: ReAct DAPO Search-o1 WebThinker R1-Searcher SimpleDeepSearcher
Closest papers by meaning
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.88)
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.88)
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.87)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.87)
- AgentWebBench: Benchmarking Multi-Agent Coordination in Agentic Web (ICML 2026 · similarity 0.87)
- Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.