Evidence-Informed LLM Beliefs for Continual Scientific Discovery
Abstract Overview
This paper studies continual scientific discovery with LLMs and argues that prior work, especially AutoDiscovery, measures surprise against a static prior even though beliefs should evolve as evidence accumulates. The authors introduce evidence-informed LLM beliefs, where the prior for a new hypothesis is updated using outcomes from previously tested hypotheses, yielding a non-stationary notion of surprisal. They compare several in-context belief-update mechanisms and find that embedding-based retrieval over prior discoveries best aligns prior beliefs with eventual posteriors. They also modify the search process with belief-update filtering and diversity maximization so that the system avoids redundant hypotheses and explores regions that remain surprising under updated beliefs.
Novelty
The main novelty is reframing LLM-driven scientific discovery around non-stationary surprisal, computed from priors that are continually updated with prior evidence rather than treated as fixed. The work also introduces an evidence-informed search procedure that couples belief-update filtering with diversity maximization to make search responsive to those updated beliefs.
Results
Embedding-based top-k retrieval performed best among the tested belief-update mechanisms, identifying about 37.5% of static surprisals as spurious. Across five discovery domains, the full evidence-informed search method increased accumulated non-stationary surprisal by 30.62% over standard static-reward search, demonstrating gains beyond online deduplication alone.
Key Points
- Static surprisal can overcount discovery progress because hypotheses implied by earlier findings may still appear surprising when priors are not updated.
- Among the tested context representations, top-k embedding-based retrieval was the most effective way to construct evidence-informed priors while being more efficient than full-context conditioning.
- The search improvements come from explicit hypothesis-selection changes—belief-update filtering, diversity maximization, and online deduplication—rather than simply swapping the reward from static to non-stationary surprisal.