Liquid AI × Insilico Medicine drug discovery foundation model accepted at EMNLP 2026

JosephJacks_ · x · 2026-09-12

• Liquid AI and Insilico Medicine's paper "MMAI Gym for Science" has been accepted to the EMNLP 2026 Industry Track.
• They combined Insilico's MMAI Gym molecular data platform with a Liquid Foundation Model across 400+ drug-discovery tasks (molecular optimization, ADMET prediction, retrosynthesis, drug-target activity, functional group reasoning), using chemistry-native tokens with SFT then RL.
• Core claim: general LLMs relying on in-context learning fall short, and simply scaling up or adding reasoning tokens doesn't help—the key is a better gym (data formats, task-specific training and benchmarking recipes), not a bigger model. The resulting efficient LFM matches or beats much larger general and specialist models on molecular benchmarks.
• Author team includes Liquid AI's Ramin Hasani and Mathias Lechner.

Original post →

More from Research

• AI math proofs: verifiable rigor matters, not who ran the AI — abc conjecture as cautionary tale — RexDouglass · 2026-09-12
• Terence Tao's real point: AI firms should finish the job and automate math entirely — RexDouglass · 2026-09-12
• Investigating the Randomness of Prime Numbers — tak3sh8 · 2026-09-12
• Dev Announces 'Fly Language Model' Built on Generative Pre-trained Fly Architecture — CatAstro_Piyush · 2026-09-12
• Study: 69% of 31,000+ agent runs contained at least one reward-hacking episode — dair_ai · 2026-09-12
• kalomaze: Poor model transfer reflects narrow human problem selection, not failed generalization — kalomaze · 2026-09-12

Research channel →