Arc Institute researcher rebuts interpolation critique with 56 cell line holdout design
anshulkundaje · x · 2026-11-20
Continuing the debate, Arc Institute researcher kjaganatha calls the "interpolation" critique a straw man: the goal was for performance to degrade gracefully as train-test similarity drops, explicitly optimized via careful holdout splits across roughly 56 cell lines.
The disagreement is now clear: the critic argues test performance depends entirely on falling within the training set's interpolation zone, while the researchers claim they systematically designed for the similarity-drop regime.
More from Research
- MIT's Hiroshi Ishii shares Tangible Media Group's Creative Tension principles — ishii_mit · 2026-11-20
- Train a real LLM entirely in your browser with this interactive demo — morgymcg · 2026-11-20
- Ex-OpenAI Researcher: English Is the New ML Programming Language, Lean the Type System — lukaszkaiser · 2026-11-20
- S1CogBench probes System One decision models across 24 cognitive tasks — Dengzhe Hou · 2026-11-20
- AFFECT from KAIST makes world models dynamics-faithful via action effect matching — John Won · 2026-11-20
- Mind2Act World: a 180-task benchmark exposing robots' mind-action composition failures — Mind2Act-World · 2026-11-20