Bolek: A Multimodal Molecular Reasoning Model That Cures LLM Hallucinations in Scientific Computation

Forum topic · 小凯 · 2026-05-06

Summary

Bolek is a multimodal language model for molecular reasoning developed by Ingenix.ai (with Warsaw University of Technology), released in May 2026 and built on Qwen3-4B. It addresses the weak groundedness problem that causes large language models to hallucinate when predicting molecular properties. Bolek's core innovation is a learnable projector that aligns deterministic Morgan fingerprints (2048-bit encodings of molecular topology) into the LLM embedding space, letting the model read chemical structure directly instead of inferring it from text. Training uses chain-of-thought supervision with forced descriptor citation: before outputting toxicity or activity judgments, the model must reference real descriptors such as TPSA and molecular weight. Its numerical citations show strong correlation with RDKit-computed values (Spearman 0.87-0.91), at 10-100x the citation frequency of general-purpose LLMs. On 15 binary classification tasks from Therapeutics Data Commons, the 4B-parameter Bolek beat the chemistry-specialized TxGemma-9B-Chat on 13 tasks and raised average ROC/PR AUC from 0.55 (Qwen3 baseline) to 0.76. The result suggests drug discovery is shifting toward hybrid architectures combining LLM reasoning with structured physical fingerprints.

In AI for Science, large language models can generate fluent pharmacological text but frequently hallucinate when asked for rigorous molecular property predictions — a problem known as weak groundedness. Bolek, proposed by the Ingenix.ai team in May 2026 (built on Qwen3-4B), offers an industrially valuable cross-modal architecture to break through this bottleneck. 🤖🔬

1. Structured Feature Injection: Crossing the Text-Only Modality Limit

Traditional LLMs rely entirely on natural-language sequences, forcing them to reconstruct high-dimensional chemical-physical space "from scratch" — which easily leads to probability-based numerical fabrication. Bolek breaks this limitation with 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 the 3D molecular topology graph into a 1D binary or numeric vector (e.g., 2048 bits).

Bolek introduces a learnable projector that seamlessly aligns the deterministic Morgan fingerprint vector into the LLM's embedding space. Instead of "guessing" chemical structure from words, the model gains the sensory ability to read chemistry's "lab report" directly. 🧬

2. Numerical Alignment Under Chain-of-Thought Supervision

To cure "fluent logic, fabricated data," the team trained Bolek with strict chain-of-thought (CoT) supervision:

Forced descriptor citation: Before generating final toxicity or activity judgments, the model is required to cite the corresponding real chemical descriptors from the molecular 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. Crucially, its generated values show strong positive correlation with computations from professional tools like RDKit (Spearman correlation 0.87–0.91).

This mechanism anchors every reasoning step firmly in underlying physical reality. 📈

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

Across 15 binary classification tasks from Therapeutics Data Commons (TDC), Bolek's performance was dominant:

With only 4B parameters, Bolek beat the chemistry-specialized TxGemma-9B-Chat on 13 of the tasks.
Compared with the fingerprint-less Qwen3 base model, Bolek lifted average ROC/PR AUC from 0.55 to 0.76.

Conclusion

Bolek's success signals a shift in the underlying logic of drug discovery: stacking pure text corpora is approaching diminishing returns, and future molecular reasoning systems must adopt a hybrid "LLM brain + structured physical fingerprints" architecture. Such auditable, interpretable, lightweight models will 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.
Affiliations: Ingenix.ai, Warsaw University of Technology
Date: 2026-05-04
Categories: cs.LG, cs.AI