PaperSearchQA: Learning to Search and Reason over Scientific Papers with RLVR
TLDR
Trains search agents to reason over scientific papers using RLVR, releasing a biomedical corpus and QA dataset.
Reasoning
Strengths include novel RLVR training for scientific search, large corpus, and dataset. Weaknesses are limited to biomedical factoid QA and not full scientific discovery.
Read-first score
Read-first score 54.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.
Field roles
Rank sensitivity
Stability: volatile; rank range: 70.
Keyword Scores
Deep Analysis
Innovations
- Training search agents to search and reason over scientific papers using RLVR, targeting technical QA in science/engineering/medicine.
- Release of a 16M biomedical paper abstract search corpus and a 60k-sample factoid QA dataset (PaperSearchQA) with benchmarks.
- Demonstration that RLVR-trained agents outperform non-RL retrieval baselines.
- Observation of agent behaviors like planning, reasoning, and self-verification.
- Scalable data creation methods extendable to other scientific domains.
Methodology
The authors construct a search corpus of 16 million biomedical paper abstracts and a factoid QA dataset (PaperSearchQA) with 60k samples answerable from the corpus. They train search agents using reinforcement learning with verifiable rewards (RLVR) in this environment, comparing against non-RL retrieval baselines. They also perform quantitative analysis of agent behaviors.
Key Results
RLVR-trained search agents outperform non-RL retrieval baselines, and agents exhibit planning, reasoning, and self-verification behaviors.