Researchers call for simplest-possible baselines in ML×Bio papers over flashy ablations
anshulkundaje · x · 2026-10-03
Yasa Baig argues ML×Bio papers favor cool architectural ablations while lacking basic controls: always baseline against the simplest possible models and explanations. Computational biologist Anshul Kundaje added that computational researchers need the same drilling, and that it matters even more for large models given their many moving parts and confounders.
More from Research
- Math PhD Trains AI to Crack Open Research Problems, Reports Progress in Arithmetic Physics — shuchaobi · 2026-10-03
- childes-db 2026.1 Released: Child Language Database Grows to 24.2M Utterances with New Annotations — najoungkim · 2026-10-03
- SETA Terminal Agent Training Suite Accepted at NeurIPS, Open-Sources 4,500+ Verifiable RL Environments — Thom_Wolf · 2026-10-03
- MedARC journal club: pretraining foundation models for intracranial EEG — iScienceLuvr · 2026-10-03
- Mosaic: exact constrained decoding for diffusion LLMs via finite automata, NeurIPS paper — StefanoErmon · 2026-10-03
- NYU Researchers Challenge Anthropic's Claim That LLMs Can Introspect — tallinzen · 2026-10-03