Research map / Large language models / Retrieval and knowledge
SCRIBES: Web-Scale Script-Based Semi-Structured Data Extraction with Reinforcement Learning
Shicheng Liu, Kai Sun, Lisheng Fu, Xilun Chen, Xinyuan Zhang, Zhaojiang Lin, Rulin Shao, Yue Liu, Anuj Kumar, Wen-tau Yih, Xin Luna Dong
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 |
|---|---|---|---|
| RL training with group cross-page reward vs agentic prompting | Qwen2.5-14B · SemiBench script gen (All) | F1 (LLM-judged): 8.0 -> 19.9 | Improved |
| RL training with SCRIBES reward vs agentic baseline | Qwen2.5-32B · SemiBench script gen (All) | F1 (LLM-judged): 19.4 -> 28.1 | Improved |
| use group-averaged reward instead of self-score only reward (ablation) | Qwen2.5-14B · SemiBench Holdout | F1 (LLM-judged): 9.5 -> 16.7 | Improved |
| use group-averaged reward instead of self-score only reward (ablation) | Qwen2.5-14B · SemiBench Example pages | F1 (LLM-judged): 27.9 -> 26.7 | Mixed |
| add CommonCrawl synthetic-annotation training after annotated data (ablation) | Qwen2.5-32B · SemiBench (All) | F1 (LLM-judged): 28.1 -> 33.2 | Improved |
| add CommonCrawl training (ablation) | Qwen2.5-14B · SemiBench (All) | F1 (LLM-judged): 19.9 -> 21.8 | Improved |
| restrict CC training to failure cases instead of all CC (ablation) | Qwen2.5-32B · SemiBench (All) | F1 (LLM-judged): 29.7 -> 33.2 | Improved |
| restrict CC training to failure cases instead of all CC (ablation) | Qwen2.5-14B · SemiBench (All) | F1 (LLM-judged): 22.0 vs 21.8 | No clear effect |
| train directly on CC data or 1:1 CC+annotated mix (ablation) | Qwen2.5 · SemiBench | F1 (LLM-judged) | No clear effect |
| add SCRIBES triples to QA input | GPT-4o · SemiBench QA | QA accuracy (%): 82.5 -> 86.6 | Improved |
| add SCRIBES triples to QA input | Qwen2.5-14B · SemiBench QA | QA accuracy (%): 74.2 -> 77.3 | Improved |
| add SCRIBES triples to QA input | Qwen2.5-3B · SemiBench QA | QA accuracy (%): 53.8 -> 54.3 | No clear effect |
Closest papers by meaning
- Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels (ICLR 2026 · similarity 0.87)
- DeepDiver: Adaptive Web-Search Intensity Scaling via Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- WARC-Bench: Web Archive based Benchmark for GUI Subtask Executions (ICLR 2026 · similarity 0.85)
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.85)
- WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.84)
- Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.84)
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.
