Bolek: A Multimodal Molecular Reasoning Model Tackling LLM Hallucinations in Scientific Computing

Forum topic · 小凯 · 2026-05-06

Summary

Bolek is a multimodal language model for molecular reasoning developed by Ingenix.ai and Warsaw University of Technology (paper dated May 2026). Built on Qwen3-4B, it addresses the hallucination problem that arises when pure-text LLMs attempt quantitative molecular property prediction. Its core innovation is a learnable projector that aligns deterministic Morgan fingerprint vectors into the LLM embedding space, letting the model read chemical structure directly rather than reconstructing it from text. Training uses chain-of-thought supervision with mandatory citation of real chemical descriptors (e.g., TPSA, molecular weight), yielding 10-100x higher numeric citation frequency and strong correlation (Spearman 0.87-0.91) with RDKit-computed values. On 15 binary classification tasks from Therapeutics Data Commons, the 4B model beat the chemistry-specialized TxGemma-9B-Chat on 13 tasks, and raised average ROC/PR AUC from 0.55 to 0.76 versus the base Qwen3 model. The result suggests drug discovery systems should combine LLM reasoning with structured physical fingerprints for auditable, interpretable inference.

In AI for Science, large language models can generate fluent pharmacological text but commonly hallucinate when doing rigorous molecular property prediction—a problem of weak groundedness. Bolek, a model proposed by the Ingenix.ai team (paper dated May 4, 2026), offers an industrially valuable cross-modal architecture to address this bottleneck.

1. Structured Feature Injection: Beyond the Text Modality

Traditional LLMs rely solely on natural language sequences, forcing them to reconstruct high-dimensional chemical-physical space from scratch—leading to probability-based fabrication of numbers. Bolek's core innovation is a multimodal fusion architecture:

\text{Embedding} = \text{Projector}(\text{Morgan\_Fingerprint}) \oplus \text{Text\_Tokens}

> What is a Morgan Fingerprint? > A standard cheminformatics algorithm that captures the local neighborhood structure of atoms in a molecule, encoding a 3D molecular topology graph into a 1D binary or numerical vector (e.g., 2048 bits).

Bolek introduces a learnable projection layer that seamlessly aligns deterministic Morgan fingerprint vectors into the LLM embedding space. The model no longer has to "guess" chemical structure from text—it gains sensory access to a chemical "lab report."

2. Numerical Alignment Under Chain-of-Thought Supervision

To fix the "fluent logic, fabricated data" problem, Bolek was trained with strict CoT supervision:

Mandatory descriptor citation: Before producing a final toxicity or activity judgment, the model must cite the real chemical descriptors corresponding to the fingerprint (e.g., polar surface area TPSA, molecular weight).
Numerical auditability: Bolek cites numerical values 10 to 100 times more frequently than general-purpose LLMs. Its cited values correlate strongly with computations from tools like RDKit (Spearman correlation 0.87–0.91).

This mechanism anchors every reasoning step in underlying physical reality.

3. Performance: A 4B Model Outperforms a 9B Competitor

Across 15 binary classification tasks in Therapeutics Data Commons (TDC):

With only 4B parameters, Bolek beat the chemistry-specialized TxGemma-9B-Chat on 13 tasks.
Compared with the base Qwen3 model without fingerprints, average ROC/PR AUC jumped from 0.55 to 0.76.

Conclusion

Bolek signals a shift in the underlying logic of drug discovery: stacking pure text corpora is hitting diminishing returns, and future molecular reasoning systems must adopt a hybrid architecture of "LLM brain + structured physical fingerprints." Such auditable, interpretable lightweight models can substantially reduce hallucination-driven trial-and-error costs in pharmaceutical pipelines.

---

Paper Information

Title: Bolek: A Multimodal Language Model for Molecular Reasoning
Authors: Frederic Grabowski, Tomasz Jetka, et al.
Institutions: Ingenix.ai, Warsaw University of Technology
Published: 2026-05-04
Categories: cs.LG, cs.AI