English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

MoRI: Teaching AI to Reason from Motivation to Method for Scientific Ideation

Forum topic · 小凯 · 2026-03-21

Summary

This article explains MoRI (Motivation-grounded Reasoning for Scientific Ideation), a framework from East China Normal University researchers that trains large language models to propose research ideas by reasoning from scientific motivations to concrete methodologies, rather than superficially recombining concepts. MoRI decomposes ideation into motivation proposal and motivation-driven reasoning stages, and trains a DeepSeek-R1-Distilled-Qwen-14B model via supervised fine-tuning followed by GRPO reinforcement learning. A composite reward combines Entropy-Aware Information Gain (EAIG), which rewards reasoning that explains high-entropy technical terms in ground-truth methods, with Contrastive Semantic Gain (CSG), which measures semantic alignment progress toward reference papers. Length anchoring and format constraints prevent reward hacking. Training data is derived from ICLR papers. Experiments report that MoRI outperforms commercial LLMs like GPT-4 and complex agent pipelines on novelty, technical rigor, and feasibility, with ablations confirming both reward components are necessary. The article discusses implications for AI research assistants, limitations in cross-domain generalization and evaluation, and the vision of AI as an inspiration engine for human scientists.

Key points

This post is a popular-science explainer of MoRI (Motivation-grounded Reasoning for Scientific Ideation), a framework that teaches LLMs to reason from scientific motivations to concrete research methods.

  • Motivation: Current AI research-ideation tools mainly do surface-level concept recombination ("Lego-block assembly") — they lack technical depth and scientific grounding. The author opens with Einstein's 1905 special relativity and LeCun's CNN as examples of how human breakthroughs arise from a deep "why" (motivation) leading to a new "how" (methodology).
  • Two-stage framework:
    • 1. Motivation Proposal — given a research context (topic, prior work, key references), the model identifies the research gap, core scientific principle, and high-level solution direction. 2. Ideation with Reasoning — the model generates a reasoning trajectory bridging motivation and method, then derives the concrete methodology from it.
  • Composite reward function (since research ideas lack verifiable ground truth, RL rewards are hard to define):
  • EAIG (Entropy-Aware Information Gain): computes token entropy in ground-truth methodologies, selects the top 25% highest-entropy (most technical) tokens, and measures how much the reasoning trajectory improves the model's ability to predict them.
  • CSG (Contrastive Semantic Gain): the difference between similarity("context + generated method", ground truth) and similarity("context alone", ground truth), ensuring semantic progress beyond simple context matching.
  • Length anchoring penalizes overly short or long reasoning (anchored to the SFT model's average length), and format constraints penalize leaking method tokens into the reasoning section, guarding against reward hacking.
  • Training pipeline:
    • 1. SFT on data extracted from ICLR papers (context → motivation → methodology). 2. RL with GRPO (the DeepSeek-R1 algorithm, critic-free, group-relative advantage estimation), sampling multiple reasoning trajectories per context and optimizing EAIG + CSG.
  • Results: Using DeepSeek-R1-Distilled-Qwen-14B, MoRI reportedly outperforms commercial LLMs (GPT-4, Claude) and complex agent baselines on novelty, technical rigor, and feasibility (judged by LLM evaluators with human-verified subsets). Ablations show both EAIG and CSG are necessary; training dynamics show reasoning length stabilizes and both rewards improve jointly.
  • Takeaways:
  • Mimicking human *thinking processes* (motivation → method) matters more than mimicking human workflows (literature search, note-taking).
  • Reward design from information-theoretic and semantic-alignment perspectives can make RL work on open-ended tasks.
  • A relatively small model with targeted training can beat much larger commercial models at ideation.
  • Limitations noted: experiments are mainly in computer science; evaluation remains subjective; risk of AI-generated "junk papers" calls for ethical guidelines.

References cited in the post

1. Gu, C., Cheng, J., Zhang, M., Zheng, P., Zheng, J., & He, G. (2026). *MoRI: Learning Motivation-Grounded Reasoning for Scientific Ideation in Large Language Models*. arXiv:2603.19044. 2. Guo, D., et al. (2025). *DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning*. arXiv:2501.12948. 3. Shao, Z., et al. (2024). *DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models*. arXiv:2402.03300. 4. Einstein, A. (1905). Zur Elektrodynamik bewegter Körper. *Annalen der Physik*, 322(10), 891–921. 5. LeCun, Y., et al. (1998). Gradient-Based Learning Applied to Document Recognition. *Proceedings of the IEEE*, 86(11), 2278–2324.

*Note: This is a summary of a Chinese-language explainer post; details should be verified against the original MoRI paper.*

Tags

#mori#scientific-ideation#reinforcement-learning#llm-reasoning#ai-scientist#grpo#reward-design#paper-explainer

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177168940