Bolek: A 4B Multimodal Model That Exposes the Pseudoscience of Text-Only LLMs in Drug Discovery

Forum topic · 小凯 · 2026-05-06

Summary

In a sharp critique of text-only AI-for-drug-discovery, researchers at Ingenix.ai and Warsaw University of Technology introduced Bolek, a 4-billion-parameter multimodal language model (built on Qwen3) that forces explicit numerical grounding before any pharmacological claim. Rather than letting large language models infer molecular descriptors such as polar surface area or logP from linguistic context, Bolek encodes Morgan fingerprints as numeric embeddings and concatenates them with text tokens. Across 15 real drug-property prediction benchmarks, this compact model outperformed the 9B chemistry-tuned TxGemma, cited underlying structural values 10 to 100 times more often, and reached a 0.91 correlation with professional cheminformatics software. The work argues that hallucination in scientific reasoning stems from relying on language priors instead of physical measurements, and that grounding small models in deterministic molecular features is more reliable and cost-effective than scaling text corpora.

Key points

Core problem: Text-only LLMs used in AI-driven drug discovery hallucinate molecular descriptors such as polar surface area and lipophilicity (logP) by extrapolating from journal-text context rather than computing them, producing plausible but unreliable pharmacological reasoning.
Proposed architecture: Bolek is a 4B-parameter multimodal model built on Qwen3 that injects deterministic molecular features into the language model.
Grounding mechanism: Morgan fingerprints, the chemistry community's standard encoding of molecular structure as numeric vectors, are translated into embeddings and concatenated with text tokens via a projector:

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

Training rule: Before producing any pharmacological conclusion, the model must explicitly recite the underlying structural fingerprint values, making numeric grounding a hard constraint rather than a learned afterthought.
Benchmark results: On 15 real-world drug-property prediction tasks, Bolek outperformed the 9B chemistry-tuned TxGemma despite being less than half its size.
Grounding frequency: Bolek cited real underlying structural values 10 to 100 times more often than text-only baselines.
Correlation with professional tools: Bolek's predictions correlated at r = 0.91 with values produced by established cheminformatics software.
Strategic argument: Scaling text corpora and context length cannot fix scientific hallucination. Physical measurements, not rhetorical fluency, are the foundation of valid scientific inference, and small grounded models can outperform larger text-only ones at a fraction of the cost.

Source details

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

Bolek: A 4B Multimodal Model That Exposes the Pseudoscience of Text-Only LLMs in Drug Discovery

Summary

Key points

Source details

Tags