Two bottlenecks holding back perturbation models: system representation and signal identification
yusufroohani · x · 2026-10-06
Researcher yusufroohani argues that perturbation models struggle to generalize across conditions due to two core issues:
- System representation: without capturing experimental, cellular, temporal and organismal context, perturbation effects may not be identifiable or generalizable.
- Signal identification: even with context, disentangling true perturbation effects from measured ones is hard; shifting task resolution (e.g. pseudobulking) may amplify signal.
He frames both as problems of the data we generate, and his team attacked them algorithmically to build intuition, identify what's learnable, and guide which experiments to run next.
More from Research
- NeurIPS paper: self-model measurements predict emergent misalignment intervention success — jiaxinwen22 · 2026-10-06
- BRIE: a living benchmark for evaluating LLM information retrieval from electronic health records — BraydonDymm · 2026-10-06
- New paper probes Olmo 3 puzzle: near-identical short-context models diverge after long-context extension — kylelostat · 2026-10-06
- Researcher heads to COLM 2026 with two Ai2 papers on hybrid models and long context — kylelostat · 2026-10-06
- Verifier Agents and Placebo Tests: Catching an AI Causal Analysis That Reasons Itself Wrong — hugobowne · 2026-10-06
- Stanford's DITTO-X does forward and reverse teleoperation for dexterous robot hands — chris_j_paxton · 2026-10-06