MAIL: Memory-driven, Adaptive, Incremental, and Literature-grounded Framework for Hypothesis Generation in Chemistry

Mahdi Babaei; Xueshen Li; Yutao Kuang; Jolene P. Reid; Yu Gan

cs.AIAug 28, 2026

MAIL: Memory-driven, Adaptive, Incremental, and Literature-grounded Framework for Hypothesis Generation in Chemistry

Authors: Mahdi Babaei, Xueshen Li, Yutao Kuang, Jolene P. Reid, Yu Gan

Organizations: Biomedical Engineering, Stevens Institute of Technology, Hoboken, NJ 07030, USA · Department of Chemistry, University of British Columbia, Vancouver, BC V6T 1Z1, Canada · Department of Bioengineering, University of Maryland, College Park, MD 20742, USA · Artificial Intelligence Interdisciplinary Institute at Maryland, University of Maryland, College Park, MD 20742, USA

Abstract

The ever-expanding volume of the chemical literature offers unprecedented opportunities to generate novel and impactful hypotheses. However, the bottleneck lies in efficiently navigating this vast knowledge base to formulate high-quality, experimentally meaningful insights. While Large Language Models (LLMs) show promise for this task, existing methods often rely on static inspiration corpora, predefined heuristics, or laborious human-in-the-loop pipelines and decision-support frameworks that limit scalability and novelty. In this work, we propose an automated approach, a Memory-augmented, Adaptive, Incremental, and Literature-grounded (MAIL) framework for hypothesis generation in chemistry. Our MAIL method formulates hypothesis generation as a temporally grounded, memory-driven reasoning process, where hypotheses emerge from an evolving conceptual path that continuously accumulates and reinterprets prior knowledge. We evaluated the MAIL framework on a public TOMATO-Chem dataset and a newly curated and disseminated high-novelty nature/science challenge (HN-NS) dataset. Across both datasets, MAIL generates structurally coherent and mechanistically plausible hypotheses, achieves the highest MIOS and MPOS by more effectively recovering the central ideas and methodological elements of the historical target hypotheses, and obtains the highest overall expert-evaluation scores for scientific quality. These results demonstrate the potential of LLMs to autonomously explore chemical domains and generate hypotheses that are both innovative and chemically plausible.

Figures & tables

Figure 1: High-level overview of the proposed autonomous hypothesis generation method. At round $t$ , the research background and retrieved literature are used to generate the current hypothesis $h_{t}$ . The LLM then evaluates this hypothesis and produces internal feedback $f_{t}$ . In the following round, the current hypothesis, LLM-generated feedback, and newly retrieved literature $I^{(t+1)}$ are jointly used to generate the refined hypothesis $h_{t+1}$ . This iterative process continues without human feedback until the final hypothesis is selected.

Figure 2: Overview of the MAIL framework. Starting from the fixed background $b$ , the model constructs a query and performs temporally filtered literature retrieval to obtain $I^{(1)}$ . The selected inspirations $S^{(1)}$ are then used to initialize $N$ candidate hypothesis trajectories. At each subsequent round $t$ , the preceding hypothesis $h_{j}^{(t-1)}$ is self-evaluated to produce structured feedback $f_{t-1}$ . This feedback, together with $b$ , the dynamically retrieved literature $I^{(t)}$ , and a compressed memory summary of previously selected inspirations, guides the selection of $S^{(t)}$ and refinement of each candidate into $h_{j}^{(t)}$ . After $T$ rounds, a target-blind selection prompt chooses the final hypothesis $h^{*}$ from the refined candidates. All feedback, refinement, and selection operations are performed automatically without human intervention during inference.

Dimension	Definition	Scoring Tips
Chemical Plausibility	Is the hypothesis chemically sound and consistent with known mechanisms or principles?	0 = Contradicts known chemistry or is implausible; 1 = Speculative but not ruled out by current understanding; 2 = Strongly consistent with known mechanisms or literature precedent.
Physical Feasibility	Can the hypothesis be tested using accessible methods or tools?	0 = Requires inaccessible, unscalable, or unknown techniques; 1 = Needs careful optimization, special reagents, or advanced equipment; 2 = Readily testable using standard lab procedures.
Novelty – Technical	Does it introduce new catalysts, substrates, experimental conditions, or materials?	0 = Uses standard techniques or well-known combinations; 1 = Applies known tools in a slightly modified or new way; 2 = Introduces a new catalyst, scaffold, platform, or methodology.
Novelty – Conceptual	Does it introduce a new chemical concept, reaction logic, or mechanistic rationale?	0 = Based entirely on established chemical concepts and reaction logic, without reinterpretation or mechanistic novelty; 1 = Recombines familiar mechanistic ideas or reaction types creatively or unexpectedly, but without introducing fundamentally new logic; 2 = Proposes a fundamentally new chemical concept, novel reaction logic, or mechanistic rationale that could reshape how related systems are understood.
Completeness	Is the hypothesis specific and testable as stated?	0 = Vague idea without concrete components or outcomes; 1 = Some elements are defined, but others are missing; 2 = Clearly specifies inputs, conditions, and measurable outputs.
Scientific Impact	Could it change how chemists think, plan, or execute reactions?	0 = Minor or incremental advancement; 1 = Could improve performance or solve a niche problem; 2 = May shift strategy, or design at a broader level.

Table 1: Evaluation rubric for research hypotheses in chemistry. Each dimension is scored on a 0–2 scale based on the provided definitions and scoring tips.

Method	Top MIOS	Avg MIOS	Top MPOS	Avg MPOS
MOOSE [ 31 ]	3.18	2.52	1.94	1.85
Zero-shot Proposer [ 6 ]	2.49	2.12	1.74	1.53
SciMON [ 12 ]	2.71	2.17	1.91	1.73
NOVA [ 33 ]	2.33	2.04	1.59	1.28
ResearchBench [ 44 ]	3.41	2.84	2.30	1.87
MOOSE-Chem [ 8 ]	4.01	3.15	3.56	2.14

Table 2: Comparison of MAIL and baseline methods on two MIOS and MPOS on the TOMATO-Chem dataset. Bold values indicate the best performance, and underlined values indicate the second-best performance.

Iterative Refinement	Feedback Reasoning	Context-Aware Prompts	Prompt Optimization	Avg MIOS	Avg MPOS
				3.69	2.38
$\times$				2.59	1.74
	$\times$			3.03	2.07
		$\times$		2.35	1.46
			$\times$	2.29	1.51

Table 3: Ablation study. The average of MIOS and MPOS over the benchmark. Each variant removes one key component of the full method. Bold values indicate the best performance, and underlined values indicate the second-best performance

	Chemical Plausibility	Physical Feasibility	Novelty (Technical)	Novelty (Conceptual)	Completeness	Scientific Impact	Broader Utility	Total Score
MOOSE-Chem	1.36	1.20	1.64	1.28	1.68	1.68	1.48	10.32
MAIL (Ours)	1.44	1.32	1.56	1.36	1.72	1.76	1.64	10.80

Table 4: Expert evaluation results on TOMATO-Chem dataset. Scores range from 0 to 2 per criterion. Bold values indicate the best performance.

Num. Papers (K) per query	Top MIOS	Avg MIOS	Top MPOS	Avg MPOS
1	3.64	2.95	3.15	1.93
3	3.91	3.27	3.47	2.05
5	4.27	3.69	3.91	2.38
10	3.36	3.01	2.90	1.72

Table 5: Experiments on the effect of the number of retrieved papers (

K

) per query on TOMATO-Chem dataset. Bold values indicate the best performance, and underlined values indicate the second-best performance.

Figure 3: Performance progression across iterative refinement rounds. The final column represents the average scores of the selected hypotheses.

Figure 4: Pearson correlation matrix of expert evaluation scores for generated hypotheses on TOMATO-Chem dataset. Cells are color-coded from blue (negative correlation) to red (positive correlation), with the intensity indicating the magnitude of the association.

Temperature	Top MIOS	Avg MIOS	Top MPOS	Avg MPOS
0.0	3.85	1.85	3.01	1.93
0.3	3.88	3.34	3.27	2.02
0.5	4.04	3.32	3.93	2.28
0.7	4.27	3.69	3.91	2.38
1.0	4.29	3.48	3.35	2.09

Table 6: Experiments on the effect of different temperature values on hypothesis generation performance. Bold values indicate the best performance, and underlined values indicate the second-best performance.

Method	Top MIOS	Avg MIOS	Top MPOS	Avg MPOS
MOOSE [ 31 ]	2.41	1.78	1.62	1.55
Zero-shot Proposer [ 6 ]	1.88	1.50	1.39	1.34
SciMON [ 12 ]	2.05	1.49	1.53	1.38
NOVA [ 33 ]	1.76	1.44	1.27	1.12
ResearchBench [ 44 ]	2.78	2.11	1.97	1.64
MOOSE-Chem [ 8 ]	3.03	2.23	2.85	1.88

Table 7: Performance evaluation on HN-NS Dataset. Each value reflects the top-scoring and average performance per method. Bold values indicate the best performance, and underlined values indicate the second-best performance.

	Chemical Plausibility	Physical Feasibility	Novelty (Technical)	Novelty (Conceptual)	Completeness	Scientific Impact	Broader Utility	Total Score
MOOSE-Chem	1.21	1.08	1.08	1.00	1.33	1.42	1.21	8.33
MAIL (Ours)	1.42	1.08	1.17	1.00	1.25	1.46	1.25	8.63

Table 8: Expert evaluation results on HN-NS Dataset. Scores range from 0 to 2 per criterion. Bold values indicate the best performance.

Figure 5: Pearson correlation matrix of expert evaluation scores for generated hypotheses on HN-NS dataset. Cells are color-coded from blue (negative correlation) to red (positive correlation), with the intensity indicating the magnitude of the association.

		TOMATO-Chem Dataset		HN-NS Dataset
Method	Model	Avg MIOS	Avg MPOS	Avg MIOS	Avg MPOS
MAIL (Ours)	GPT-4o (Baseline)	3.69	2.38	2.61	2.09
MOOSE-Chem	GPT-4o (Baseline)	3.15	2.14	2.23	1.88
MAIL (Ours)	Qwen2.5-72B	3.61	2.32	2.55	2.04
MOOSE-Chem	Qwen2.5-72B	3.04	2.11	2.20	1.86
MAIL (Ours)	Llama-3.1-70B	3.56	2.27	2.51	2.00

Table 9: Performance evaluation of MAIL and MOOSE-Chem across varying LLMs on the TOMATO-Chem and HN-NS datasets. Bold values indicate the best performance for a given model and dataset.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Round	MIOS	MPOS
Round 1	1.30	0.00
Round 2	3.00	2.00
Round 3	4.40	4.00

Appendix

Table A.1: Scores for each refinement round based on the two evaluation metrics.

Scientific Hypothesis Generation

Scientific hypothesis generation develops methods for proposing testable explanations and research directions from literature, data, and experiments.

See all 90 papers

Learning Mechanistic Reasoning for Chemical Reactions with Large Language Models
Jul 14, 2026Xingyu Dang, Haocheng Tang, Junmei Wang +1LLM Reasoning
Department of Computer Science, Princeton University, Princeton, NJ, 08540, United States · School of Pharmacy, University of Pittsburgh, Pittsburgh, PA, 15261, United States · Khoury College of Computer Science, Northeastern University, Boston, MA, 02115, United States +2
Towards Diverse Scientific Hypothesis Search with Large Language Models
Jun 9, 2026Haorui Wang, Parshin Shojaee, Kazem Meidani +7Automated Algorithm Discovery Scientific Hypothesis Generation
Georgia Institute of Technology · Virginia Tech · Carnegie Mellon University +3
ProjectionBench: Evaluating Scientific Hypothesis Generation in LLMs Under Progressive Information Disclosure
May 28, 2026A. J. Lew, Y. Cao, M. J. BuehlerLLM Evaluation Scientific Reasoning
Unreasonable Labs · Unreasonable Labs Mountain View, CA 94043, USA

Jul 14, 2026J/K move · Enter open · S save

Learning Mechanistic Reasoning for Chemical Reactions with Large Language Models

Xingyu Dang, Haocheng Tang, Junmei Wang, Yanjun Li

Department of Computer Science, Princeton University, Princeton, NJ, 08540, United States · School of Pharmacy, University of Pittsburgh, Pittsburgh, PA, 15261, United States · Khoury College of Computer Science, Northeastern University, Boston, MA, 02115, United States+2

LLM Reasoning

Open paper AskPDF arXiv

Reaction mechanisms consist of the step-by-step sequences of elementary reactions that explain chemical transformations. Learning the mechanism logic is therefore essential for enhancing the fundamental chemical intelligence of large language models (LLMs). The stepwise deduction of reaction mechanism aligns naturally with the reasoning paradigms of reasoning LLMs. However, current chemical LLMs primarily emphasize coarse-grained name reactions for product prediction and retrosynthesis, often leading to physical inconsistencies and hallucinations. In contrast, specialized small-scale generative models for mechanism inference typically suffer from restricted generalization capacity across diverse chemical spaces. To overcome these limitations, we built a novel, large-scale reasoning dataset of reaction mechanisms. Furthermore, we established the FukuyamaBench, a difficult benchmark derived from Fukuyama's Advanced Organic Reaction Mechanism book, to rigorously evaluate model performance on hierarchical mechanism reasoning. Our fine-tuned Qwen3-30B-A3B achieves 8.3% exact pathway match on FukuyamaBench Set~A, surpassing the specialized FlowER model (5.1%), demonstrating that mechanism-aware training substantially enhances chemical reasoning in language models.

Abstract

Scientific Hypothesis Generation

Explore similar work

Learning Mechanistic Reasoning for Chemical Reactions with Large Language Models

Towards Diverse Scientific Hypothesis Search with Large Language Models

ProjectionBench: Evaluating Scientific Hypothesis Generation in LLMs Under Progressive Information Disclosure