BioSimBench: RL-Trained 27B Model Beats Frontier LLMs at Simulating Cell Responses
david_van_dijk · x · 2026-11-04
- CellType launches BioSimBench (built on NVIDIA): models must predict how real cells respond to drugs, gene knockdowns, and immune signals — a test of whether LLMs can simulate biology.
- Even the best frontier models get only part of the real lab responses right, leaving large room for improvement.
- The team also built BioSimGym, an RL training environment; a 27B model trained there outscores every frontier model on all tasks, suggesting domain-specific RL beats general-purpose models for biological simulation.
More from Research
- TLAPS-Bench and agentic AI surface a read-consistency bug in SlateDB's WAL spec — tianyin_xu · 2026-11-05
- CellType claims 27B biology model beats all 12 frontier models at predicting cell behavior — alifcoder · 2026-11-05
- No-Prerequisite Information Geometry Tutorial Released With Video and PDF Notes — FrnkNlsn · 2026-11-05
- structRFM: structure-guided RNA foundation model trained on 21M pairs lands in Nature — EricTopol · 2026-11-05
- Yandex open-sources Yambda-5B, a 4.79B-interaction music recommendation dataset — teortaxesTex · 2026-11-05
- Adversarial speculative decoding boosts long-horizon agent success with 0.06% params — kastnerkyle · 2026-11-05