The year/Independent research

Paper 2603.03756

MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier

Published
Mar 2026
Research lab
Independent
Citations
3
GitHub
90 stars

01 In brief

Summary

The paper introduces MOOSE-Star, a framework for tractable training of LLMs for scientific discovery by directly modeling P(hypothesis|background).

The authors argue that end-to-end training is intractable due to combinatorial complexity O(N^k) from retrieving k inspirations from a large knowledge base.

MOOSE-Star decomposes this into sequential subtasks: inspiration retrieval and hypothesis composition, reducing complexity to O(kN).

Further innovations include bounded composition (robustness to noisy retrieval), hierarchical search (O(log N) complexity), and motivation planning (pruning search space).

They release TOMATO-Star, a dataset of 108,717 papers (38,400 GPU hours).

Experiments show MOOSE-Star scales with training data and inference budget, while brute-force sampling hits a 'complexity wall' for multi-inspiration hypotheses.

The framework enables scalable training and test-time inference for scientific discovery.

02 From the paper

Abstract

While large language models (LLMs) show promise in scientific discovery, existing research focuses on inference or feedback-driven training, leaving the direct modeling of the generative reasoning process, $P(\text{hypothesis}|\text{background})$ ($P(h|b)$), unexplored. We demonstrate that directly training $P(h|b)$ is mathematically intractable due to the combinatorial complexity ($O(N^k)$) inherent in retrieving and composing inspirations from a vast knowledge base. To break this barrier, we introduce MOOSE-Star, a unified framework that enables tractable and scalable training of $P(h|b)$, while supporting more scalable inference. In the best case, MOOSE-Star reduces complexity from exponential to logarithmic ($O(\log N)$) by (1) training on decomposed subtasks derived from the probabilistic equation of discovery, (2) employing motivation-guided hierarchical search to enable logarithmic retrieval and prune irrelevant subspaces, and (3) utilizing bounded composition for robustness against retrieval noise. To facilitate this, we release TOMATO-Star, a dataset of 108,717 decomposed papers (38,400 GPU hours) for training. Empirically, MOOSE-Star scales continuously with training data and inference budget, whereas direct brute-force sampling hits a complexity wall.