Research map / Large language models / Instruction and fine-tuning
Weasel: Out-of-Domain Generalization for Web Agents via Importance-Diversity Data Selection
Fatemeh Pesaran Zadeh, Seyeon Choi, Xing Han Lù, Siva Reddy, Gunhee Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select 10K steps with Weasel instead of full 52K SFT | Qwen2.5-7B-Instruct · WebArena-Lite (AgentTrek) | success rate: 10.9 -> 14.5 | Improved |
| select 10K steps with Weasel instead of uniform sampling | Qwen2.5-7B-Instruct · WebArena-Lite | success rate: 9.1 -> 14.5 | Improved |
| add target-centered pruning to full SFT (ablation) | Qwen3-8B · WebArena / MiniWob | success rate: MiniWob 59.4 -> 40.3 | Hurt |
| add target-centered pruning to full SFT (ablation) | Qwen2.5-7B-Instruct · WebArena-Lite | success rate: 10.9 -> 9.7 | No clear effect |
| use target-centered pruning instead of Prune-by-Bid (ablation) | Qwen2.5-7B-Instruct · WebArena-Lite | success rate: 8.5 -> 10.9 | Improved |
| combine target-centered with semantic pruning (ablation) | Target-centered pruning · WebArena-Lite | success rate: 10.9 -> 7.3 | Hurt |
| move retained AXTree window away from target (ablation) | Target-centered pruning · WebArena-Lite | success rate: 11.5% -> 7.3% | Hurt |
| add reasoning synthesis to Weasel (ablation) | Qwen3-8B · WebArena-Lite | success rate: 17.0 -> 21.2 | Improved |
| add data selection on top of reasoning synthesis (ablation) | Qwen3-8B · WebArena-Lite | success rate: 18.2 -> 21.2 | Improved |
| add pairwise diversity to unary importance (ablation) | Weasel · WebArena-Lite | success rate: 10.9 -> 14.5 | Improved |
| add unary importance to diversity-only selection (ablation) | Weasel · WebArena-Lite | success rate: 7.9 -> 14.5 | Improved |
| use max-composition of state and answer diversity (ablation) | Weasel · WebArena-Lite | success rate: 9.7 -> 14.5 | Improved |
Closest papers by meaning
- How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.90)
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.89)
- WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.89)
- Prune4Web: DOM Tree Pruning Programming for Web Agent (AAAI 2026 · similarity 0.88)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.88)
- A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.88)
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.
