MoRI: Learning Motivation-Grounded Reasoning for Scientific Ideation in Large Language Models
MoRI employs reinforcement learning with entropy-aware information gain and semantic contrast to enhance scientific ideation depth and validity.
Key Findings
Methodology
MoRI initializes the model via supervised fine-tuning to generate motivations from context, then employs a composite reward combining entropy-aware information gain (EAIG) and contrastive semantic gain (CSG). EAIG encourages detailed technical elaboration by focusing on high-entropy tokens in ground-truth methodologies, while CSG aligns reasoning trajectories with scientific validity via semantic similarity metrics. The training uses Group Relative Policy Optimization (GRPO) to optimize reasoning paths, with regularization terms like length anchoring and format constraints to prevent reward hacking. Experiments on ICLR 2024-2025 papers demonstrate that MoRI surpasses commercial models and agentic baselines in novelty, rigor, and feasibility.
Key Results
- On datasets from ICLR 2024-2025, MoRI achieves an average score of 3.18 out of 4, outperforming the baseline fine-tuned models by 8.5%. It maintains strong generalization on out-of-domain NeurIPS data with a mean of 3.15, showing robustness. The reward components effectively promote technical depth and logical direction, as ablation studies confirm. Human and automatic evaluations show high correlation (r=0.715), validating the quality of generated ideas.
- The entropy-aware information gain component effectively guides the model to produce detailed technical reasoning, while the semantic contrast ensures the generated methodology moves meaningfully from problem to solution. Combining both yields ideas that are both deep and aligned with scientific principles, outperforming single-reward variants.
- Internalized reasoning capabilities enable MoRI to generate scientifically grounded ideas that go beyond superficial pattern matching, demonstrating potential for AI-driven scientific discovery in complex research scenarios.
Significance
This work advances AI's role in scientific innovation by internalizing the reasoning process, addressing the limitations of external scaffolding and heuristic methods. MoRI enhances the depth, coherence, and scientific validity of AI-generated ideas, offering a new paradigm for automated research and hypothesis generation. Its dual-reward system and training strategy provide a robust framework for future AI systems aiming to emulate human-like scientific reasoning. The ability to generalize across domains and maintain high-quality reasoning marks a significant step toward autonomous scientific discovery, with broad implications for academia and industry.
Technical Contribution
MoRI introduces a novel composite reward mechanism integrating entropy-aware information gain and contrastive semantic gain, tailored for scientific reasoning tasks. It leverages GRPO to optimize reasoning trajectories, internalizing the logic from motivation to methodology. The approach effectively balances technical depth with conceptual alignment, providing theoretical guarantees for reasoning quality and robustness. This framework surpasses existing imitation-based or external scaffolding methods, establishing a new standard for AI-driven scientific ideation.
Novelty
This research is the first to embed motivation-grounded reasoning within large language models using a dual-reward system that emphasizes technical detail and semantic alignment. Unlike prior work relying on external agents or superficial pattern recombination, MoRI internalizes the reasoning process, focusing on the logical connection from research motivation to methodology. Its innovative reward design and training pipeline set it apart as a pioneering approach in AI scientific reasoning.
Limitations
- The model depends heavily on large-scale scientific literature datasets, which may introduce biases and limit applicability in underrepresented fields. Its performance in niche or emerging domains remains uncertain.
- Training with reinforcement learning incurs high computational costs and long convergence times, restricting real-time deployment or rapid iteration.
- The reward mechanisms, while effective, can still be exploited or produce unintended shortcuts, necessitating further refinement for robustness in complex reasoning tasks.
Future Work
Future efforts will focus on integrating multi-modal data, such as experimental results and datasets, to enrich reasoning capabilities. Developing more sophisticated validation and explanation mechanisms will improve interpretability and trustworthiness. Extending the framework to broader scientific fields and incorporating domain-specific knowledge bases will enhance generalization. Additionally, optimizing training efficiency and exploring unsupervised pretraining strategies could make the approach more scalable for practical applications.
AI Executive Summary
Scientific discovery is a complex cognitive process that involves formulating hypotheses, designing experiments, and deriving methodologies grounded in scientific principles. Despite advances in large language models (LLMs), current AI systems largely emulate human research workflows superficially, often producing ideas that lack technical depth and scientific grounding. This gap limits their utility in genuine scientific innovation.
Chenyang Gu and colleagues address this challenge with MoRI, a novel framework that internalizes scientific reasoning within LLMs. The core idea is to model scientific ideation as a motivation-driven process: the model first identifies a research motivation from the context, then generates a reasoning trajectory that logically connects this motivation to a feasible methodology. This approach moves beyond simple context completion, emphasizing deliberate, motivation-based problem solving.
To achieve this, MoRI employs supervised fine-tuning to establish foundational skills, followed by reinforcement learning with a carefully designed composite reward. The reward combines entropy-aware information gain (EAIG), which encourages detailed technical elaboration on high-entropy tokens, with contrastive semantic gain (CSG), which aligns the reasoning trajectory with scientifically valid solutions. This synergy ensures the generated ideas are both technically rigorous and conceptually aligned.
Experimental results on datasets derived from recent ICLR papers demonstrate that MoRI outperforms leading commercial models and agentic frameworks across multiple metrics, including novelty, rigor, and feasibility. It maintains strong generalization on out-of-domain datasets like NeurIPS, indicating robustness. Human evaluations confirm the high quality and scientific validity of the generated ideas.
This work significantly advances AI's capacity for scientific reasoning, providing a scalable, internally grounded approach that bridges the gap between superficial pattern matching and genuine scientific insight. Despite high computational costs and some remaining challenges in reward robustness, MoRI opens new avenues for automated scientific discovery, promising to accelerate innovation and deepen our understanding of complex scientific problems.
Deep Dive
Abstract
Scientific ideation aims to propose novel solutions within a given scientific context. Existing LLM-based agentic approaches emulate human research workflows, yet inadequately model scientific reasoning, resulting in surface-level conceptual recombinations that lack technical depth and scientific grounding. To address this issue, we propose \textbf{MoRI} (\textbf{Mo}tivation-grounded \textbf{R}easoning for Scientific \textbf{I}deation), a framework that enables LLMs to explicitly learn the reasoning process from research motivations to methodologies. The base LLM is initialized via supervised fine-tuning to generate a research motivation from a given context, and is subsequently trained under a composite reinforcement learning reward that approximates scientific rigor: (1) entropy-aware information gain encourages the model to uncover and elaborate high-complexity technical details grounded in ground-truth methodologies, and (2) contrastive semantic gain constrains the reasoning trajectory to remain conceptually aligned with scientifically valid solutions. Empirical results show that MoRI consistently outperforms strong commercial LLMs and complex agentic baselines across multiple dimensions, including novelty, technical rigor, and feasibility. The code is available on \href{https://github.com/ECNU-Text-Computing/IdeaGeneration}{GitHub}.