Iterative Hypothesis Generation for Scientific Discovery with Monte Carlo Nash Equilibrium Self-Refining Trees

Paper · 2025 · Gollam Rabby, Diyana Muhammed, Prasenjit Mitra et al.

Summary

Purely LLM-based methods for generating scientific hypotheses often produce ideas that are either unoriginal or poorly grounded. The authors attribute this to two gaps: little iterative refinement and a weak balance between exploring new ideas and exploiting promising ones. To close these gaps they propose MC-NEST (Monte Carlo Nash Equilibrium Self-Refine Tree). It treats hypothesis generation as a game between competing search strategies. A Monte Carlo Tree Search, guided by Nash Equilibrium and UCT, traverses a tree of candidate hypotheses. Adaptive sampling and LLM self-critique then iteratively improve the selected candidates. The framework was tested on biomedicine, social science and computer science datasets with several LLMs. Outputs were judged by BERTScore and by ratings of novelty, clarity, significance and verifiability, including a human evaluation. MC-NEST scored consistently above prompt-based baselines. A peptide-engineering case study, visualized with AlphaFold, shows the approach in practice. The work matters because it frames search-driven, human-supervised LLM hypothesis generation as a reusable method that works across domains.

Key points

Averaged over novelty, clarity, significance and verifiability (1–3 scale), MC-NEST scored 2.65 on social science, 2.74 on computer science and 2.80 on biomedicine. The best prompt-based baselines scored 2.36, 2.51 and 2.52.
Three ways of choosing candidates are compared within the framework: a Greedy approach, plain Importance Sampling, and a pairwise variant called Pairwise Importance Sampling. It was run with GPT-4o, Deepseek-7B and Deepseek-32B, alongside ZSCoT prompting.
In the human evaluation, the ZSCoT variant reached 2.62 on social science and 2.37 on computer science. Importance Sampling with Deepseek-32B scored 2.44 on biomedicine.
Extending rollouts from 4 to 8 steps improved both BERTScore and the qualitative ratings. Pairwise Importance Sampling gained the most in novelty and significance.
In the case study, the system suggested lysine-for-arginine swaps in an NLS and alanine-for-glycine swaps in a linker. The authors report that these changes improved solubility while preserving nuclear localization.
The authors list as a limitation that the dataset leans toward computer science papers. Code, datasets and evaluation protocols are announced for release.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields, topics and concepts

📚 Shelves 4Computer Science · 16,667 works Artificial Intelligence · 16,194 works Arts and Humanities · 16,039 works Language and Linguistics · 15,891 works

Cite

Cite as Gollam Rabby, Diyana Muhammed, Prasenjit Mitra and Soren Auer. 2025. Iterative Hypothesis Generation for Scientific Discovery with Monte Carlo Nash Equilibrium Self-Refining Trees. arXiv:2503.19309.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2503.19309 · arXiv 2503.19309

@misc{rabby2025iterative,
  title = {{Iterative Hypothesis Generation for Scientific Discovery with Monte Carlo Nash Equilibrium Self-Refining Trees}},
  author = {Gollam Rabby, Diyana Muhammed, Prasenjit Mitra and Soren Auer},
  year = {2025},
  doi = {10.48550/arxiv.2503.19309},
  eprint = {2503.19309},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2503.19309}
}

References

3 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shya…
Springer (2006) 10. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. Advances in neural information processing systems 35, 22199–22213 (2022) 11. Lake, B.M., Salakhutdinov, R., Tenenbaum, J.B.: Human-level concept learning through probabilistic… — Large Language Models as Biomedical Hypothesis Generators: A Comprehensive Evaluation
G. Rabby et al. 18. Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al.: Mastering the game of go with deep neural networks and tree search. nature 529(7587), 484–489 (2016) 19. Wang, H., Fu, T.… find

Updated 31 October 2026

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rabby-2025-iterative-hypothesis-generation-scientific-discovery/