Research map / Large language models / Agents, code and math
MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier
Zonglin Yang, Lidong Bing
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on inspiration retrieval data (152k examples) | R1-Distilled-Qwen-7B · TOMATO-Star inspiration retrieval | accuracy: 28.42% -> 54.37% | Improved |
| fine-tune on hypothesis composition data (97k samples) | R1-Distilled-Qwen-7B · TOMATO-Star hypothesis composition | M3 total: 4.05 -> 4.68 | Improved |
| add bounded composition training data (ablation) | MS-HC-7B · hypothesis composition with ground-truth inspiration | M3 total: 4.68 -> 4.74 (1x), 4.73 (2x) | No clear effect |
| add bounded composition training data (ablation) | MS-HC-7B · hypothesis composition with Hard proxy inspirations | M3 total: 2.39 -> 2.42 (1x), 2.56 (2x) | Improved |
| jointly train on IR and HC data (ablation) | MS-7B vs MS-IR-7B / MS-HC-7B · TOMATO-Star | IR accuracy and HC M3 total: IR 54.37% -> 54.34%; HC 4.68 -> 5.02 | Improved |
| use hierarchical best-first search instead of tournament search (ablation) | MS-IR-7B routing · 3,035 inspiration corpus | IR inference calls: 218.00 -> 67.78 | Improved |
| add detailed motivation to hierarchical search (ablation) | hierarchical search · 3,035 inspiration corpus | IR inference calls: 67.78 -> 63.05 | Improved |
| add simple motivation to hierarchical search (ablation) | hierarchical search · 3,035 inspiration corpus | IR inference calls: 67.78 -> 63.80 | Improved |
| increase number of required inspirations k from 1 to 3 (ablation) | brute-force end-to-end sampling · per-attempt pass rate | pass rate: 9.46% -> 0.23% | Hurt |
| increase number of required inspirations k from 1 to 3 (ablation) | decomposed HC and IR subtasks · per-attempt pass rate | pass rate: HC 18.88% -> 25.75%; IR 27.84% -> 24.22% | No clear effect |
| scale training samples from 10^0 to 10^5 (ablation) | MS-IR-7B · inspiration retrieval | accuracy: 28.4% -> 54.4% | Improved |
| increase inference budget with hierarchical search vs brute-force sampling | MOOSE-Star vs brute force · 109 test papers | fraction of papers solved: 100% @ 5979 steps vs 38.5% @ 9500 | Improved |
Closest papers by meaning
- MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 2025 · similarity 0.92)
- SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models (NeurIPS 2024 · similarity 0.88)
- DiscoveryBench: Towards Data-Driven Discovery with Large Language Models (ICLR 2025 · similarity 0.88)
- Position: Data-driven Discovery with Large Generative Models (ICML 2024 · similarity 0.88)
- LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models (ICML 2025 · similarity 0.87)
- CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
