MoRI: Learning Motivation-Grounded Reasoning for Scientific Ideation in Large Language Models

Chenyang Gu; Jiahao Cheng; Meicong Zhang; Pujun Zheng; Jinquan Zheng; Guoxiu He

cs.CLMar 19, 2026

MoRI: Learning Motivation-Grounded Reasoning for Scientific Ideation in Large Language Models

Authors: Chenyang Gu, Jiahao Cheng, Meicong Zhang, Pujun Zheng, Jinquan Zheng, Guoxiu He

Organizations: School of Economics and Management, East China Normal University

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 MoRI (Motivation-grounded Reasoning for Scientific Ideation), 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 GitHub.

Figures & tables

Figure 1: Conceptual comparison. Unlike existing approaches that rely on pattern recombination or computationally expensive external scaffolding, MoRI internalizes scientific ideation through learning motivation-grounded reasoning. It initially identifies a Motivation ( $m$ ) from a given Context ( $x$ ), then generates a Reasoning Trajectory ( $z$ ) to deduce a grounded Methodology ( $y$ ), which is optimized via our composite RL rewards.

Figure 2: Overview of MoRI. Our framework optimizes reasoning via GRPO (Bottom) using composite rewards: Entropy-Aware Information Gain (Left) for high-entropy explanation and technical depth and Contrastive Semantic Gain (Right) for logical direction alignment, modulated by Length Anchoring (Center) to enforce reasoning depth.

Source	Overall		Novelty		Rigor		Feasibility
Source	H	L	H	L	H	L	H	L
MoRI	2.76	3.07	2.73	3.20	2.73	3.07	2.80	2.93
AI-Scientist-V2	2.17	2.76	2.14	2.57	2.00	2.71	2.36	3.00
GPT-4o	1.75	2.17	1.50	2.00	1.75	2.25	2.00	2.25
Ground Truth	3.29	3.47	3.33	3.53	3.47	3.80	3.07	3.07
Overall	2.67	3.03	2.65	3.02	2.67	3.12	2.69	2.94

Table 1: Human-LLM alignment analysis. We report mean scores from human experts (H) and LLM judge (L) across 60 samples from three dimensions.

Commercial Models
	ICLR Test Set					NeurIPS OOD Test Set					Overall
Model	Nov.	Rig.	Feas.	Mean	$\Delta$	Nov.	Rig.	Feas.	Mean	$\Delta$	Nov.	Rig.	Feas.	Mean	$\Delta$
GPT-4o	2.51	2.78	2.79	2.69	+18.6%	2.57	2.87	2.88	2.77	+13.7%	2.54	2.83	2.84	2.74	+16.1%
Claude-3.5-Sonnet	3.39	3.07	2.82	3.09	+3.2%	3.42	3.03	2.94	3.13	+0.6%	3.40	3.05	2.89	3.11	+2.3%
Agentic Frameworks
AI-Scientist-V2 † (Sonnet)	3.45	3.08	2.89	3.14	+1.6%	–	–	–	–	–	–	–	–	–	–

Table 2: Main comparison across test sets. We report results on the in-domain ICLR test set (83 papers), the out-of-domain NeurIPS test set (67 papers), and their combination (150 papers).

\Delta

denotes the relative improvement of MoRI over each baseline. Best results (excluding Ground Truth) are in bold . † denotes the use of a stronger backbone ( claude-3-5-sonnet ) for AI-Scientist-V2. For agentic frameworks, the default backbone is gpt-4o-2024-08-06 , with Haiku referring to claude-3-5-haiku-2024-10-22 .

Model	Nov.	Rig.	Feas.	Mean
Full-SFT (Direct)	2.80	2.69	2.81	2.75
Full-SFT (Two-stage)	2.85	2.76	2.94	2.85
MoRI	3.30	3.00	3.16	3.15

Table 3: Ablation on motivation conditioning (NeurIPS OOD test set). Direct:

x\rightarrow y

; Two-stage:

x\rightarrow m\rightarrow y

; MoRI adds RL optimization on top of the two-stage formulation.

Figure 3: Training Dynamics. Moving average of CoT Length (a), Shaped EAIG (b), and Shaped Semantic Score (c).

$\boldsymbol{w_{s}}$	$\boldsymbol{w_{e}}$	Description	Novelty	Tech. Rigor	Feasibility	Mean
0.0	1.0	EAIG Only	2.68	2.22	2.63	2.51
1.0	0.0	CSG Only	3.16	2.93	3.06	3.05
0.5	0.5	Balanced	3.32	2.96	2.98	3.09
0.7	0.3	Optimal	3.34	3.04	3.07	3.15

Table 4: Ablation on reward composition (Step 100 w/ Length Anchoring).

w_{s}

and

w_{e}

denote weights for CSG and EAIG respectively.

Figure 4: Entropy Mask Dynamics. Moving average of CoT Length (a), Shaped EAIG (b), and Shaped Semantic Score (c).

Ent. Mask	Novelty	Tech. Rigor	Feasibility	Mean	$\boldsymbol{\Delta}$
Top 50%	3.16	2.78	3.00	2.98	–
Top 25%	3.32	2.96	2.98	3.09	+3.7%

Table 5: Ablation on Entropy Filtering strategies (config

w_{s}

=0.5,

w_{e}

=0.5, Step 100).

Figure 5: Impact of Length Anchoring. Moving average of CoT Length (a), Shaped EAIG (b), and Shaped Semantic Score (c).

Config	LA	Novelty	Tech. Rigor	Feasibility	Mean	$\boldsymbol{\Delta}$
$w_{s}$ =0.7, $w_{e}$ =0.3	✗	3.22	2.94	3.00	3.05	–
$w_{s}$ =0.7, $w_{e}$ =0.3	✓	3.34	3.04	3.07	3.15	+3.2%
$w_{s}$ =0.5, $w_{e}$ =0.5	✗	2.96	2.76	2.95	2.89	–
$w_{s}$ =0.5, $w_{e}$ =0.5	✓	3.32	2.96	2.98	3.09	+6.9%

Table 6: Impact of Length Anchoring (Step 100). LA : Length Anchoring.

Figure 6: Overview of the Data Construction Pipeline. The process operates in three distinct phases: (1) Data Ingestion & Preprocessing , where raw ICLR PDFs are converted into text using MinerU; (2) Extraction & Cleaning , which parses the research context (

x

) and motivation (

m

), while extracting and de-symbolizing the method section (

y^{*}

) to remove notation-specific noise; and (3) Posterior Reconstruction , where a reasoning chain (

z

) is synthesized by rewriting an initial thought trace (

z_{\text{init}}

) to logically align with the ground-truth method.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 7 — Figure 7: End-to-end example of one training/inference instance. Part (A) shows the extracted context $x$ provided to the model; part (B) shows an abbreviated excerpt of the corresponding held-out ground-truth method $y^{*}$ ; part (C) shows the inference prompt template. Only the context in part (A) is exposed as model input, which prevents leakage of the target method.

Sub-domain	ICLR	NeurIPS
Perception & Multimodal	25.3%	16.4%
Language Reasoning & Alignment	22.9%	19.4%
Generative Modeling	16.9%	7.5%
Learning Efficiency & Foundations	16.9%	28.4%
Scientific & Interdisciplinary AI	10.8%	17.9%
Decision, Control & Safety	7.2%	10.4%

Appendix

Table 7: Sub-domain distribution of the two test sets.

Statistic	Value
Mean $\pm$ Std	1.26 $\pm$ 1.0
P25 / Median / P75 / P90	0.40 / 1.14 / 1.93 / 2.64
EAIG threshold (top-25%)	1.93
Selected tokens	36,573 (25.1%)

Appendix

Table 8: Token entropy distribution statistics over 100 training samples.

Category	Count	Sel. Rate	Mean Ent.
Technical terms	82,896	34.2%	1.52
Other	6,307	17.2%	0.93
Common words	40,279	14.5%	0.95
Whitespace	615	12.5%	0.86
Math symbols	2,240	10.8%	0.73
Punctuation	12,461	7.7%	0.87

Appendix

Table 9: Selection rate by token category under the top-25% entropy mask. Technical terms are selected at more than twice the rate of common words.

Figure 8: Annotated excerpt from a ground-truth method section. Token color indicates entropy (blue=low, red=high).

Parameter	Value	Description
max_generations	1	Ideas per topic
num_reflections	3	Reflection rounds
max_results	10	Papers per query
year_filter	-2024	Publication cutoff
overlap_threshold	0.8	Leakage filter

Appendix

Table 10: AI-Scientist-V2 parameter settings.

Stage	Original	Adapted
Problem ID	3-round iteration	Bypassed
Method Dev	3-round iteration	Preserved
Experiment	3-round iteration	Bypassed
Input	API retrieval	Regex parsing

Appendix

Table 11: ResearchAgent pipeline modifications.

Parameter	Original	Adapted
Discussion rounds	7	4
Team members	3	4
Discussion model	LLaMA-8B	GPT-4o-mini
Review model	LLaMA-70B	Qwen-72B
Final output	Abstract	Method

Appendix

Table 12: VirSci configuration comparison.

Panel of Figure 9 — Figure 9: Complete prompt template for context-aware LLM evaluation. The prompt instructs the judge to adopt a domain expert mindset and provides detailed rubrics for each evaluation dimension.

Dimension	Pearson $r$
Novelty	0.613
Technical Rigor	0.684
Feasibility	0.629

Appendix

Table 13: Human–LLM alignment on an additional 60-sample NeurIPS subset.

Config	EAIG Weight	Final CoT Length
s0-e1	100%	Shortest (collapses)
s5-e5	50%	Short
s7-e3	30%	Moderate
s1-e0	0%	Preserved

Appendix

Table 14: Observed CoT length under different EAIG weights w/o Length Anchoring.

Training Dynamics
Parameter	Value
Global Batch Size	8
PPO Mini-Batch Size	4
PPO Micro-Batch Size (per GPU)	2
PPO Epochs per Rollout	1
Learning Rate	$5\times 10^{-7}$

Appendix

Table 15: Hyperparameters for MoRI training.

Model	Novelty	Tech. Rigor	Feasibility	Overall
GPT-4o	2.54 [2.38, 2.70]	2.83 [2.71, 2.94]	2.84 [2.72, 2.97]	2.74 [2.64, 2.84]
Claude-3.5-Sonnet	3.40 [3.28, 3.52]	3.05 [2.92, 3.18]	2.89 [2.78, 2.99]	3.11 [3.02, 3.20]
AIS-v2 (GPT-4o)	2.38 [2.23, 2.53]	2.60 [2.48, 2.71]	2.82 [2.69, 2.95]	2.60 [2.49, 2.70]
ResearchAgent	2.63 [2.46, 2.80]	2.39 [2.22, 2.56]	2.50 [2.34, 2.67]	2.50 [2.35, 2.66]
VirSci	2.19 [2.08, 2.30]	2.26 [2.16, 2.36]	2.27 [2.18, 2.37]	2.24 [2.16, 2.32]
Full-SFT	2.99 [2.85, 3.13]	2.82 [2.71, 2.93]	2.99 [2.87, 3.11]	2.93 [2.84, 3.03]

Appendix

Table 16: Combined results (ICLR + NeurIPS, 150 papers) with 95% bootstrap confidence intervals (10,000 resamples). Best results (excluding Ground Truth) are in bold .

MoRI vs	$\boldsymbol{\Delta}$	$\boldsymbol{p}$	Cohen’s $\boldsymbol{d}$
VirSci	+0.93	$<$ 0.001	2.000 (large)
ResearchAgent	+0.67	$<$ 0.001	0.939 (large)
AIS-v2 (GPT-4o)	+0.58	$<$ 0.001	1.083 (large)
GPT-4o	+0.44	$<$ 0.001	0.874 (large)
Full-SFT	+0.24	0.003	0.458 (small)
Claude-3.5-Sonnet	+0.07	n.s.	—

Appendix

Table 17: Pairwise significance tests on the combined set (150 papers).

\Delta

: difference in overall mean.

p

-values are Bonferroni-corrected.

Panel of Figure 10 — Figure 10: Generated Motivation by MoRI for the example research idea on plan-based reasoning improvement.

Panel of Figure 11 — Figure 11: Generated Reasoning by MoRI (representative excerpts with key insights highlighted).

Panel of Figure 12 — Figure 12: Generated Method by MoRI (Part 1: Overview, Core Idea, and Training).

Panel of Figure 13 — Figure 13: Generated Method by MoRI (Part 2: Solution Generation, Evaluation, and DPO Training).

Panel of Figure 14 — Figure 14: Generated Method by AI-Scientist-V2 for the same research problem.

Panel of Figure 15 — Figure 15: Generated Method by Claude-3.5-Sonnet for the same research problem.

Panel of Figure 16 — Figure 16: Behavior A: Goal Decomposition with Constraint Enumeration. The model declares an objective, enumerates required properties, and narrows toward a concrete mechanism in a top-down manner.

Panel of Figure 17 — Figure 17: Behavior B: Hypothesize–Critique–Revise Loop. The model actively discovers shortcomings in its own proposals and iterates, demonstrating self-verification rather than single-shot generation.

Panel of Figure 18 — Figure 18: Behavior C: Paradigm Questioning via Reverse Framing. The model challenges the default methodological stance of the field before committing to a solution direction, enabling non-obvious research angles.

Scientific Reasoning in Language Models

Scientific reasoning in language models studies how AI systems use evidence, domain knowledge, and tools to solve problems and support discovery.

See all 132 papers

Debate as Reward: A Multi-Agent Reward System for Scientific Ideation via RL Post-Training
Apr 17, 2026Moein Salimi, Babak Hosseini Mohtasham, Amin Aghakasiri +6Reward Hacking RL for Language Models
Sharif University of Technology · University of Tehran · Sharif University of Technology[cs.AI]
Agentic-Ideation: Sample Efficient Agentic Trajectories Synthesis for Scientific Ideation Agents
Jun 30, 2026Keyu Zhao, Lingyan Kong, Fengli Xu +1Multi-Agent LLM Systems AI Agents for Scientific Discovery
Department of Electronic Engineering, Tsinghua University.
IdeaScientist: Orchestrating Agents for Grounded Scientific Ideation
Oct 2, 2026Jiarui Liu, Renjie Tao, Yiwei Liao +15Research Idea Generation Autonomous Research Agents
Carnegie Mellon University · Meta Reality Labs · FAIR at Meta +5

Apr 17, 2026J/K move · Enter open · S save

Debate as Reward: A Multi-Agent Reward System for Scientific Ideation via RL Post-Training

Moein Salimi, Babak Hosseini Mohtasham, Amin Aghakasiri, Mahdi Naieni+5

Sharif University of Technology · University of Tehran · Sharif University of Technology[cs.AI]

Reward Hacking RL for Language Models Multi-Agent LLM Systems Scientific Hypothesis Generation Large Language Model-Guided Scientific Discovery

Open paper AskPDF arXiv

Large Language Models (LLMs) have demonstrated potential in automating scientific ideation, yet current approaches relying on iterative prompting or complex multi-agent architectures often suffer from hallucination or computational inefficiency. A critical bottleneck in applying Reinforcement Learning (RL) to this open-ended domain is reward hacking -- where models exploit imperfect evaluation proxies to maximize scores without producing genuine scientific innovation. To address these limitations, we propose an RL framework explicitly tailored for high-quality scientific idea generation. We propose the first multi-agent reward function designed to serve as a judge, decoupling methodological validation from implementation details while providing strict binary rewards that are robust to reward hacking. To effectively optimize against this sparse signal, we utilize an unbiased variant of Group Relative Policy Optimization to mitigate artificial length bias. We grounded our training in ICLR-320, a curated dataset of problem-solution pairs extracted from ICLR 2024 proceedings. Experiments demonstrate that our framework significantly outperforms state-of-the-art baselines across expert-evaluated metrics of novelty, feasibility, and effectiveness.

Abstract

Scientific Reasoning in Language Models

Explore similar work

Debate as Reward: A Multi-Agent Reward System for Scientific Ideation via RL Post-Training

Agentic-Ideation: Sample Efficient Agentic Trajectories Synthesis for Scientific Ideation Agents

IdeaScientist: Orchestrating Agents for Grounded Scientific Ideation