MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery

Training starts before the first update - with the tokenizer 🧪

Each chemical entity is represented in multiple ways: SMILES, SELFIES, an IUPAC name where available, and non-canonical traversals of the same molecular graph.

This helps the model learn the molecule - not memorize one syntax - while format isolation ensures it answers in the requested format.

We also extend the LM vocabulary with chemistry-native SMILES, SELFIES, and FASTA tokens, so a chemical symbol is never fragmented into a handful of text tokens.

The gym spans 400+ tasks across six groups:

• 2D molecules
• 3D molecules
• 2D proteins
• 3D proteins
• Drug–gene interactions
• Cross-domain tasks

Each batch samples a group, then a task, then an example - so no modality gets starved.

The rest of the recipe:

• Domain-faithful reasoning traces curated with specialist predictive models, physics-based simulators, and expert chemists - not generic chain-of-thought

• Task-specific RL rewards mapped to a shared scale, so regression, classification, and generation can coexist during multi-task GRPO

• Automated benchmarking with train/test decontamination for every task

We applied the recipe to LFM2, Liquid AI efficient hybrid architecture ⚙️

Two medicinal-chemistry reasoning benchmarks provide a strong first test:

MuMO-Instruct asks the model to edit a molecule to improve 3–4 conflicting properties while preserving similarity to the original.

FGBench asks how a functional-group change affects a measured property - the kind of question a chemist answers at the bench.

After one SFT + RFT run, LFM2-2.6B-MMAI edges out GPT-5.5 and Claude 4.6 Opus on both benchmarks. It also beats GeLLM³O-3, which was trained on MuMO-Instruct’s own task distribution.

But success rate alone is misleading ⚠️

Several baselines report large property improvements while falling below 0.2 Tanimoto similarity - quietly violating the instruction to preserve the original molecule.

Retrosynthesis is where the gap becomes widest 🔁

Exact match is the wrong metric here.

A single target may have several credible precursor sets, even if only one appears in the recorded answer. ChemCensor may judge both plausible.

So we evaluate plausibility and diversity, not agreement with one reference route.

Base LFM2-2.6B scores zero on both - not merely “low.” It cannot produce a parseable reactant set.

After MMAI Gym:

• 94% of its proposals are unique and valid
• The 24B MoE variant achieves the best average ChemCensor plausibility of every model compared, at every measured top-K

Both properties matter.

Several baselines find one highly plausible prediction but generate few unique candidates - a sign that they may be leaning on memorized common patterns.

Real synthesis planning needs a diverse set of credible disconnections to choose from 🧩

And because URSA-expert-2026 contains out-of-distribution targets with undisclosed answers, these results cannot be explained by benchmark contamination.

Next: the core of drug discovery - ADMET, PK, and toxicity 💊

Many molecules. Many properties. Many questions.

LFM2-2.6B-MMAI beats TxGemma-27B on 13 of 22 endpoints, despite having roughly 1/10 the number of parameters.

Among foundation models, it is best-in-class on:

• Lipophilicity
• Drug-induced liver injury
• Blood–brain barrier penetration

On lipophilicity, it even edges past the top TDC leaderboard entry.

Where does it lose? On chemistry that appears to need more capacity: half-life, CYP2D6, and hERG.

Specialist non-LLM models on the TDC leaderboard still lead on most endpoints.

The same model can also write 3D molecular structures as text 🧬

Nothing has to condition it: asked for a molecule and nothing else, the model writes the graph and every atom’s coordinates.

Given a protein pocket, it does the same thing on demand - the ligand’s SMILES graph first, then coordinates one atom at a time. No diffusion module.

We measure this on Bench3DFit, our own benchmark for pocket-conditioned generation: both LFM2-MMAI variants achieve the strongest UniDock binding affinity of any language model we tested - beating several specialist diffusion baselines and approaching the strongest models overall.

The contrast is striking:

GPT-5.5 generates molecules that are almost perfectly valid and geometrically tidy - but bind poorly. Our models do the opposite.

MMAI Gym pushes the LM toward the objective that matters for structure-based drug design, rather than rewarding it for simply playing safe 🎯