MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search

TL;DR

The paper addresses the challenge of generating fine-grained, experimentally actionable scientific hypotheses, formalizing the task as combinatorial optimization and introducing the MOOSE-Chem2 framework with Hierarchical Heuristic Search (HHS) to exploit LLM internal heuristics. It demonstrates that hierarchical decomposition smooths the reward landscape and enables more effective optimization, outperforming flat baselines on a post-2024 chemistry benchmark with expert-annotated ground-truth hypotheses. Key findings include better alignment with ground-truth content, higher recall of methodological details, and that repeated evaluations by the strongest single LLM outperform diverse ensembles, with identical-LLM aggregation improving novelty and recall. The work suggests broad generalizability across disciplines via discipline-specific hierarchies and highlights directions for integrating experimental feedback and benchmark expansion.

Abstract

Large language models (LLMs) have shown promise in automating scientific hypothesis generation, yet existing approaches primarily yield coarse-grained hypotheses lacking critical methodological and experimental details. We introduce and formally define the new task of fine-grained scientific hypothesis discovery, which entails generating detailed, experimentally actionable hypotheses from coarse initial research directions. We frame this as a combinatorial optimization problem and investigate the upper limits of LLMs' capacity to solve it when maximally leveraged. Specifically, we explore four foundational questions: (1) how to best harness an LLM's internal heuristics to formulate the fine-grained hypothesis it itself would judge as the most promising among all the possible hypotheses it might generate, based on its own internal scoring-thus defining a latent reward landscape over the hypothesis space; (2) whether such LLM-judged better hypotheses exhibit stronger alignment with ground-truth hypotheses; (3) whether shaping the reward landscape using an ensemble of diverse LLMs of similar capacity yields better outcomes than defining it with repeated instances of the strongest LLM among them; and (4) whether an ensemble of identical LLMs provides a more reliable reward landscape than a single LLM. To address these questions, we propose a hierarchical search method that incrementally proposes and integrates details into the hypothesis, progressing from general concepts to specific experimental configurations. We show that this hierarchical process smooths the reward landscape and enables more effective optimization. Empirical evaluations on a new benchmark of expert-annotated fine-grained hypotheses from recent literature show that our method consistently outperforms strong baselines.

Figures (5)

• Figure 1: Hierarchies designed for chemistry and material science by PhD-level domain expert.
• Figure 2: Overview of the proposed Hierarchical Heuristic Search (HHS) framework.
• Figure 3: The smoothing effect of hierarchy on the reward landscape of hypothesis space.
• Figure 4: Hierarchical design as a low-pass filtering over the spectrum of the reward landscape.
• Figure 5: Overview of the input and output of the Hierarchical Heuristic Search (HHS) framework, also known as MOOSE-Chem2.