Kundaje: fine-tuned scFMs perform abysmally, observational pretraining can't learn causality
anshulkundaje · x · 2026-10-06
In posts 4-5 of his thread, Anshul Kundaje delivers his core verdict: scFMs trained on observational data cannot reliably learn causal effects and cannot be used directly for perturbation prediction or causal discovery. Moreover, the absolute performance of scFMs fine-tuned on perturbation data is abysmal — in no way vindicating the often-made claims that scFMs were the path to perturb prediction or causal transcription regulation models.
More from Research
- User Sim Index is broken: trivial bot scores 95% across behavioral dims — ericzelikman · 2026-10-06
- Used OpenAI Dots as a Free Agent Swarm to Break a 47-Year-Old Math Record — jaxchang · 2026-10-06
- Stanford prof says current experiment designs can't train causal virtual cell models — anshulkundaje · 2026-10-06
- TIDES dataset on multi-party and multi-agent collaboration to debut at COLM 2026 — josephseering · 2026-10-06
- LakeQuest QA benchmark, testing RAG on messy enterprise tables and docs, hits COLM — hllo_wrld · 2026-10-06
- Frontier Data Summit lineup: Chollet, Dawn Song headline batch of new agent benchmarks — StanfordAILab · 2026-10-06