Generating Literature-Driven Scientific Theories at Scale
TLDR
Automated theory generation from scientific literature at scale, using 13.7k papers to synthesize 2.9k theories, outperforming parametric methods.
Reasoning
The paper presents a novel approach to automated scientific discovery by focusing on theory building rather than experiment generation, using large-scale literature grounding. Strengths include empirical validation with real papers and future predictions; weaknesses may include limited discussion of theory quality metrics and potential biases in literature selection.
Read-first score
Read-first score 60.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 71.
Field roles
Rank sensitivity
Stability: volatile; rank range: 26.
Keyword Scores
Deep Analysis
Innovations
- Formulation of theory synthesis from scientific literature as a novel problem in automated scientific discovery
- Large-scale study generating 2.9k theories from 13.7k source papers
- Comparison of literature-grounding versus parametric LLM knowledge for theory generation
- Evaluation of accuracy-focused versus novelty-focused generation objectives
Methodology
The authors synthesize theories from a corpus of 13.7k scientific papers, generating 2.9k theories. They compare generation using literature-grounding (retrieval from papers) against parametric LLM memory, and examine accuracy-focused versus novelty-focused objectives. Evaluation measures how well theories match existing evidence and predict results from 4.6k subsequently-written papers.
Key Results
Literature-supported theory generation significantly outperforms parametric LLM memory in matching existing evidence and predicting future results from later papers.