[Paper Review] A Multi-task Large Reasoning Model for Molecular Science

Pengfei Liu, Shuang Ge+2 more|arXiv (Cornell University)|Mar 13, 2026

Machine Learning in Materials Science0 citations

TL;DR

The paper presents a multi-task large reasoning model with a multi-specialist architecture and chain-of-thought reasoning, enhanced by reinforcement learning, achieving strong multi-task molecular performance with data-efficient learning.

ABSTRACT

Advancements in artificial intelligence for molecular science are necessitating a paradigm shift from purely data-driven predictions to knowledge-guided computational reasoning. Existing molecular models are predominantly proprietary, lacking general molecular intelligence and generalizability. This underscores the necessity for computational methods that can effectively integrate scientific logic with deep learning architectures. Here we introduce a multi-task large reasoning model designed to emulate the cognitive processes of molecular scientists through structured reasoning and reflection. Our approach incorporates multi-specialist modules to provide versatile molecular expertise and a chain-of-thought (CoT) framework enhanced by reinforcement learning infused with molecular knowledge, enabling structured and reflective reasoning. Systematic evaluations across 10 molecular tasks and 47 metrics demonstrate that our model achieves an average 50.3% improvement over the base architecture, outperforming over 20 state-of-the-art baselines, including ultra-large-parameter foundation models, despite using significantly fewer training data and computational resources. This validates that embedding explicit reasoning mechanisms enables high-efficiency learning, allowing smaller-scale models to surpass massive counterparts in both efficacy and interpretability. The practical utility of this computational framework was validated through a case study on the design of central nervous system (CNS) drug candidates, illustrating its capacity to bridge data-driven and knowledge-integrated approaches for intelligent molecular design.

Motivation & Objective

• Motivate integrating chemical knowledge with deep learning for molecular tasks beyond pure prediction.
• Develop a multi-specialist, task-adaptive framework that embeds chemical logic into CoT reasoning.
• Achieve data-efficient learning by combining data synergy and specialist synergy with reinforcement learning.
• Demonstrate superior multi-task performance across ten molecular tasks with limited training data and resources.
• Show practical utility via a CNS drug design case study linking generation, prediction, and synthesis.

Proposed method

• Construct a multi-specialist layer inside a pre-trained LLM (DeepSeek-7B base) with a router to coordinate eight specialist groups by task type.
• Train prediction specialists on a 93K instruction dataset and inference (CoT) specialists on a 3.5K high-quality CoT dataset.
• Incorporate Low-Rank Adaptation (LoRA) to enable efficient parameter updates.
• Apply reinforcement learning with task-specific molecular science rewards to align reasoning with chemical validity.
• Use three-step training: representation learning via instruction fine-tuning on 74.5K data, CoT fine-tuning on 3.6K data, and knowledge-alignment RL.
• Leverage data synergy (joint training of related tasks) and specialist synergy (prediction + inference specialist collaboration) to enhance reasoning.

Figure 1: Overview of the reasoning framework. (a) Current LLMs in multi-task molecular science tasks, showcasing their core capabilities and inherent challenges. (b) Molecular multi-task reasoning framework, detailing the process from user query to inference, featuring tokenization and embedding, s

Experimental results

Research questions

• RQ1 Can a multi-task molecular reasoning model outperform state-of-the-art baselines across diverse tasks by embedding chemical knowledge into chain-of-thought reasoning?
• RQ2 What is the impact of data synergy and specialist synergy on multi-task molecular performance and reasoning alignment?
• RQ3 How does reinforcement learning with knowledge-guided rewards affect the consistency between prediction and inference specialists?
• RQ4 Is it possible to deliver high accuracy and interpretable reasoning with a smaller, knowledge-infused model in CNS drug design scenarios?

Key findings

• Average improvement of 50.3% over the base architecture across 10 molecular tasks.
• Outperforms over 20 state-of-the-art baselines, including ultra-large parameter models, while using fewer training data and resources.
• Shows near 6% improvement over the strong multi-task model LLaSMol on task metrics.
• Demonstrates robust reasoning interpretability via chain-of-thought and a CNS drug design case study.
• Reveals that data and specialist synergy plus CoT RL substantially boost performance compared to instruction-only or CoT-only variants.
• Identifies Lipophilicity as a task where the model is slightly behind the baseline, indicating specialization limits.

Figure 2: Comprehensive evaluation of model performance and synergy. (a) Overall performance against baselines, comparing our model with over ten representative LLMs across core metrics as detailed in (b), with all radar plot metrics normalized. (b) Detailed metric values across the ten tasks. (c) A