Go to ACL ARR 2026 January homepageFrom Text Generation to Structured Reasoning: Structure-Aware Hypothesis Evolution for Scientific Hypothesis Generation
Keywords: Scientific Hypothesis Generation, Evolutionary Algorithms, Structure-Aware Reasoning, Materials Science
Abstract: Large language models (LLMs) have shown promise for automated scientific hypothesis generation, yet most existing approaches formulate hypothesis generation at the text level, without explicitly modeling the structured reasoning processes underlying scientific discovery.
As a result, hypothesis generation and refinement are often weakly grounded in scientific logic and misaligned with expert reasoning.
We propose a structure-centric framework that models scientific hypotheses as hierarchical reasoning chains.
To support structure-aware modeling and evaluation, we introduce HSRC-5000, a large-scale materials science dataset constructed by decomposing scientific papers into explicit reasoning components.
Building on this representation, we construct the Hierarchical Scientific Reasoning Graph and propose Structure-Aware Hypothesis Evolution (SAHE), a dimension-aware evolutionary framework that enables controlled and causally consistent hypothesis generation.
Experiments show that explicitly modeling scientific reasoning structure consistently improves hypothesis generation quality, yielding an average absolute improvement of about one point in overall score over strong retrieval- and evaluation-based baselines, with particularly pronounced gains in logical coherence and multi-dimensional novelty.
Qualitative analyses further indicate close alignment between the generated reasoning chains and human expert scientific logic.
The code and dataset will be publicly released.
Paper Type: Long
Research Area: Natural Language Generation
Research Area Keywords: Generation,inference methods,domain adaptation,
Contribution Types: NLP engineering experiment, Data resources, Theory
Languages Studied: English
Submission Number: 10097
Loading