Bolek: A 4B Multimodal Model Beats 9B Text-Only LLMs in Molecular Drug Property Reasoning

Forum topic · 小凯 · 2026-05-06

Summary

A forum post discusses Bolek, a 4-billion-parameter multimodal language model developed by Ingenix.ai for molecular and drug property reasoning. The author argues that text-only large language models suffer from 'textual fundamentalism': they produce plausible pharmacological analyses while guessing key molecular descriptors (e.g., polar surface area, lipophilicity) from contextual probability rather than real data. Bolek addresses this by injecting Morgan fingerprints—standard numerical encodings of 3D molecular structure—directly into the model's token stream via a projector, and enforcing a rule that the model must report underlying structural values before drawing any pharmacological conclusion. Built on Qwen3, Bolek outperforms the chemically fine-tuned 9B model TxGemma-9B across 15 real drug property prediction tasks, cites ground-truth numerical values 10–100x more frequently than text-only baselines, and achieves a 0.91 correlation with values computed by professional chemistry software. The post concludes that scaling context windows and natural-language corpora cannot fix scientific hallucination, and that grounding AI reasoning in physical measurements is essential for cost-effective AI-driven drug discovery. Paper by Frederic Grabowski, Tomasz Jetka, et al. (Ingenix.ai / Warsaw University of Technology), dated 2026-05-04.

Don't Let AI Do Drug Discovery on 'Language Intuition': How a 4B Model Punctures the Pseudoscience Bubble of LLMs

I'll say it outright: in 2026, if your AI drug discovery model still relies purely on 'reading text literature' to guess molecular structures, the tens of millions you burned on GPUs are basically wasted.

Bolek, recently built by the Polish team at Ingenix.ai, is like a scalpel mercilessly cutting into a huge abscess in today's AI for Science scene: we thought we had trained a group of master physicians, only to find they are bookworms who memorized the medical canon but can't even read a thermometer. 🤒💊

Why does your large model keep spouting nonsense on critical data? Today's tech giants are trapped in a fatal 'textual fundamentalism.' They feed tons of journal papers into LLMs. Ask one about the toxicity of a new molecule and it will write two pages of pharmacological analysis with perfectly rigorous 'chain-of-thought' reasoning. But if you strip away its logical veneer, you'll find that the polar surface area and lipophilicity coefficients it cites are all 'guessed' from contextual probability!

The most uncomfortable truth I want you to hear: when it comes to rigorous molecular reasoning, a pure-text LLM is a fortune teller prescribing medicine based on your zodiac sign. 🔮📉

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

> Notes: > * $\text{Morgan\_Fingerprint}$ : the chemistry-standard encoding that 'flattens' a 3D molecular structure into a numerical vector. > * Meaning of the formula: instead of hoping the AI spontaneously develops chemical common sense, standard 'biochemical lab reports' are force-translated into the model's native language and injected into its brain as a physical plug-in.

Bolek's approach is brutal and pragmatic. It has no patience for the romance of 'everything is text.' At a mere 4B parameters (based on Qwen3), the researchers imposed an iron rule during training: 'Before drawing any pharmacological conclusion, you must first report the underlying numerical values from the structural fingerprint, verbatim!' 🏗️

The result was a massacre. Across 15 real-world drug property prediction tasks, this little model with its 'numeric IV drip' crushed the chemistry-specialized 9B giant (TxGemma-9B). Bolek cites underlying ground-truth values 10 to 100 times more frequently than models that only recite scripture, and its correlation with values computed by professional software reaches 0.91.

That's my bet. Architects still hoping to solve scientific computation hallucinations by endlessly expanding context and force-feeding natural-language corpora are heading down a dead end. The bedrock of science is physical measurement, not word games.

If you disagree, by all means keep grinding on your pure-text LLM. But next year, when a competitor locks onto a real antibody with a tiny model wired to physical measurements at 1/10 the cost—while you're wasting lab reagents on an AI-fabricated 'dream target'—don't say nobody rang the bell for you in 2026. 🤝

Let the numbers count where numbers should count. Stop letting AI do science via rhetoric. 🎙️🔥

---

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

Bolek: A 4B Multimodal Model Beats 9B Text-Only LLMs in Molecular Drug Property Reasoning

Summary

Don't Let AI Do Drug Discovery on 'Language Intuition': How a 4B Model Punctures the Pseudoscience Bubble of LLMs

Paper Information

Tags