💬 NLP

MolReasoner: Toward Effective and Interpretable Reasoning for Molecular LLMs

MolReasoner is a novel two-stage framework that enhances the interpretability and accuracy of molecular reasoning in Large Language Models by combining knowledge-enhanced Chain-of-Thought fine-tuning with a task-adaptive reinforcement learning stage to effectively mitigate hallucinations and outperform existing baselines.

Original authors: Guojiang Zhao, Zixiang Lu, Yutang Ge, Sihang Li, Zheng Cheng, Haitao Lin, Lirong Wu, Hanchen Xia, Hengxing Cai, Wentao Guo, Hongshuai Wang, Mingjun Xu, Siyu Zhu, Guolin Ke, Linfeng Zhang, Zhifeng Gao

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guojiang Zhao, Zixiang Lu, Yutang Ge, Sihang Li, Zheng Cheng, Haitao Lin, Lirong Wu, Hanchen Xia, Hengxing Cai, Wentao Guo, Hongshuai Wang, Mingjun Xu, Siyu Zhu, Guolin Ke, Linfeng Zhang, Zhifeng Gao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, well-read librarian (a Large Language Model, or LLM) who has read every book in the world. This librarian is amazing at writing stories, answering trivia, and chatting about history. But now, you ask them to do something very specific: describe a complex chemical molecule in plain English, or build a molecule based on a written description.

If you just ask the librarian to "do their best," they might guess. They might say, "Oh, this molecule looks like a car!" when it's actually a bicycle. Or, if you ask them to build a car, they might glue a wheel to a door and call it a vehicle. In chemistry, these "guesses" are dangerous because they can lead to fake drugs or impossible materials. This is called hallucination.

The paper introduces MolReasoner, a new training method designed to turn this "guessing librarian" into a meticulous, step-by-step chemical engineer.

Here is how they did it, explained through a simple analogy:

The Problem: The "Guessing" Librarian

Before this paper, there were two main ways to teach the librarian about chemistry:

Just Asking (Prompting): You say, "Hey, describe this molecule." The librarian tries to recall patterns from their training but lacks deep chemical knowledge. They often make up facts (hallucinations) because they are just guessing based on how words usually go together.
Memorizing (Fine-tuning): You show the librarian thousands of examples of "Molecule X = Description Y." The librarian memorizes the answers perfectly. But if you show them a new molecule they've never seen, they freeze or make up a wrong answer because they didn't learn how to think, they just learned what to say.

The Solution: The Two-Stage "Apprenticeship"

The authors created MolReasoner, which teaches the model to think before it speaks. They used a two-stage training process, like a master chef training a new apprentice.

Stage 1: The "Recipe Book" Phase (Mol-SFT)

The Analogy: Imagine giving the apprentice a cookbook where every recipe doesn't just list ingredients, but explains why you add them. "First, we chop the onion because it releases flavor..."
What they did: They created a massive dataset where the model had to write out its Chain of Thought (CoT). Instead of just giving the answer, the model had to write:
1. "I see a phosphate group here."
2. "This means the molecule is likely acidic."
3. "I need to attach the carbon chain here."
The Result: The model learned the logic of chemistry, not just the vocabulary. It stopped guessing and started reasoning.

Stage 2: The "Taste Test" Phase (Mol-RL)

The Analogy: Now the apprentice is cooking. The master chef (the reward system) doesn't just say "Good" or "Bad." They taste the dish and give specific feedback: "The salt is perfect, but the sauce is too runny," or "You forgot the garlic."
What they did: They used a special scoring system (Reinforcement Learning) that checks the model's work on multiple levels:
- Format: Did you write the answer in the right box?
- Structure: Is the molecule actually buildable? (No floating atoms!)
- Fragments: Did you get the specific parts (like the "tail" or the "head" of the molecule) right?
- Function: Does it have the right chemical groups?
The Result: The model learned to self-correct. If it made a mistake, the "taste test" told it exactly where it went wrong, so it could fix it next time.

Why This Matters: The "Detective" vs. The "Magician"

The paper shows a fascinating difference in how errors happen:

Old Models (The Magician): When they fail, they fail spectacularly. They might invent a completely fake molecule that doesn't exist. It's like a magician pulling a rabbit out of a hat that isn't there. You can't trust them.
MolReasoner (The Detective): When it fails, it fails logically. It might get the length of a carbon chain wrong by one atom, but the rest of the reasoning is sound. It's like a detective who got the wrong suspect but followed the correct clues. This makes the errors diagnosable and fixable.

The Real-World Impact

Think of drug discovery as trying to build a key that fits a specific lock (a disease).

Old way: You try to guess the shape of the key. Sometimes you get lucky, but often you make a key that looks like a key but doesn't open the lock.
MolReasoner way: It measures the lock, understands the mechanism, and builds the key step-by-step. It's much more likely to create a key that actually works.

In Summary

MolReasoner is a new way to teach AI how to be a scientist rather than just a storyteller. By forcing the AI to write out its reasoning steps and then grading it on the quality of those steps (not just the final answer), the authors created a model that:

Understands chemistry deeply.
Explains its thinking clearly.
Makes fewer dangerous mistakes.

It's the difference between a student who memorized the answers to a test and a student who actually learned how to solve the problems.

1. Problem Statement

Large Language Models (LLMs) have shown promise in various domains, but their application to molecular reasoning remains underexplored and fraught with challenges. Existing approaches suffer from two primary limitations:

Prompt-based Methods: Rely on general-purpose LLMs without domain-specific adaptation. They often fail to capture deep molecular semantics, leading to "structural hallucinations" (chemically invalid molecules) and reliance on superficial linguistic cues rather than chemical logic.
Fine-tuning Methods (without explicit reasoning): Train models on molecule-text pairs using standard supervised learning. While they improve output accuracy slightly, they encourage memorization rather than genuine reasoning. This results in poor generalization to unseen structures, a lack of interpretability (black-box behavior), and an inability to internalize chemical principles.

The core research question is: Can we move beyond mere memorization to teach LLMs to perform high-fidelity, interpretable reasoning on molecular tasks?

2. Methodology: The MolReasoner Framework

The authors propose MolReasoner, a two-stage training framework designed to transition LLMs from memorization to high-fidelity chemical reasoning. The framework utilizes the ChEBI-20 dataset and focuses on two tasks: Molecule Captioning (text generation from structure) and Text-based De Novo Molecule Generation (structure generation from text).

Stage 1: Mol-SFT (Knowledge-Guided Supervised Fine-Tuning)

Goal: To provide a "warm-up" that establishes a foundation for structured reasoning and familiarizes the model with interpretable reasoning processes.
Data Construction: The authors construct a high-quality Chain-of-Thought (CoT) dataset (~42,000 samples) using GPT-4o.
- Knowledge Injection: Prompts are enriched with detailed chemical structural information (e.g., rings, aromaticity, molecular weight) and functional group identification (using EFGs).
- Representation: They use SELFIES (Self-Referencing Embedded Strings) instead of SMILES. SELFIES guarantees chemical validity by construction, eliminating syntax errors like mismatched parentheses that plague SMILES.
- Process: The model is trained to generate step-by-step reasoning (identifying core structures, functional groups, and assembly logic) before producing the final answer in a strict <answer> format.

Stage 2: Mol-RL (Task-Adaptive Reinforcement Learning)

Goal: To refine reasoning capabilities, mitigate hallucinations, and move beyond simple imitation using Group Relative Policy Optimization (GRPO).
Reward Mechanism: Unlike standard binary rewards (correct/incorrect), MolReasoner employs a multi-dimensional, task-adaptive reward system:
- For Molecule Captioning: Uses a Language Similarity Reward ( $R_{language}$ ) based on BLEU, ROUGE, and METEOR scores, combined with a Format Accuracy Reward (ensuring <answer> tags are used).
- For De Novo Generation: Uses a Structural Similarity Reward ( $R_{structural}$ $R_{s t r u c t u r a l}$ ) composed of four key components:
  1. Fingerprint Similarity: Mean of Morgan, MACCS, and RDKit fingerprints.
  2. SELFIES Similarity: Character-level BLEU score.
  3. Fragment Similarity: Jaccard overlap and recall of structural sub-components.
  4. Functional Group Matching: Exponential decay based on the difference in functional group counts.
Mechanism: This composite reward guides the model to align global molecular semantics with local structural details, preventing "reward hacking" and ensuring chemically coherent outputs.

3. Key Contributions

Two-Stage Synergistic Framework: MolReasoner successfully bridges the gap between knowledge injection (SFT) and policy optimization (RL), demonstrating that a "cold-start" with knowledge-guided CoT is essential for effective RL in scientific domains.
Novel Reward Design: The introduction of a multi-level reward system (incorporating fingerprints, fragments, and functional groups) specifically addresses the unique challenges of molecular generation, such as structural hallucinations and functional group mismatches.
Interpretability and Error Analysis: The framework produces outputs with explicit reasoning chains. Error analysis reveals that MolReasoner fails in "graceful" ways (localized, diagnosable errors) rather than the catastrophic hallucinations seen in baseline models.
Representation Choice: The paper validates the superiority of SELFIES over SMILES for LLM-based molecular tasks, showing significant improvements in validity and metric scores.

4. Experimental Results

The model was evaluated on the ChEBI-20 test set against strong baselines, including closed-source models (GPT-4o, DeepSeek-R1), open-source LLMs (Qwen, LLaMA), and specialized molecular models (ether0, Mol-Instructions).

Molecule Captioning: MolReasoner (7B parameters) achieved state-of-the-art performance, significantly outperforming all baselines across BLEU, ROUGE, and METEOR metrics. It surpassed even much larger models (e.g., LLaMA3-70B) and specialized reasoning models.
Text-based De Novo Generation: MolReasoner achieved the highest scores in chemical validity (Val.), exact match rates, and structural similarity (Tanimoto scores, Fragment metrics). It notably reduced structural hallucinations compared to fine-tuned models without explicit reasoning.
Ablation Studies:
- Removing the Warm-up (SFT) stage caused a drastic performance drop, confirming the necessity of knowledge-guided CoT for RL initialization.
- Using SELFIES outperformed SMILES by a large margin (e.g., >4.5x improvement in BLEU-4).
- Composite Rewards were shown to be superior to single-metric optimization, which often led to reward hacking.
Out-of-Distribution (OOD) Generalization: MolReasoner maintained its performance advantage on unseen molecular datasets (TOMG) and general reasoning benchmarks, indicating it learned transferable chemical reasoning principles rather than just memorizing the training set.

5. Significance

Scientific Reliability: By shifting from memorization to reasoning, MolReasoner offers a more reliable tool for high-stakes applications like drug discovery and materials design, where hallucinations can be costly.
Interpretability: The explicit Chain-of-Thought outputs allow researchers to audit the model's logic, identifying exactly where a reasoning step failed (e.g., misidentifying a functional group vs. incorrect assembly), which is crucial for scientific adoption.
Methodological Blueprint: The paper provides a principled approach for adapting LLMs to other scientific domains, emphasizing the combination of domain-knowledge injection (via CoT) and granular, multi-dimensional reward shaping (via RL) to overcome the limitations of standard fine-tuning.

In conclusion, MolReasoner demonstrates that equipping LLMs with structured reasoning capabilities and domain-specific reward signals can effectively solve complex molecular translation tasks, achieving high fidelity and interpretability previously unattainable with standard LLM approaches.