Causal Circuit Tracing Reveals Distinct Computational Architectures in Single-Cell Foundation Models: Inhibitory Dominance, Biological Coherence, and Cross-Model Convergence
Ihor Kendiukhov · 2026
Read Causal Circuit Tracing Reveals Distinct Computational Architectures in Single-Cell Foundation Models: Inhibitory Dominance, Biological Coherence, and Cross-Model Convergence by Ihor Kendiukhov (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-01752.txt.
Description
Ihor Kendiukhov traces causal links inside single-cell foundation models in this 2026 study. Such models read gene activity in single cells, yet how their internal features affect each other was unknown. Sparse autoencoders split model activity into readable features, which gives the study its handles.
Kendiukhov ablates, or switches off, single features and watches how downstream features respond. Applied to two large models across four conditions, the method maps tens of thousands of causal edges. Both models show about half their links as biologically coherent, plus a strong tilt toward inhibitory links. Shared domain pairs across models are far more common than chance, and disease-linked domains join that shared set often. Gene-level lab checks, though, suggest the links encode co-expression more than true causal control.
Catalog record
| CREATOR | Ihor Kendiukhov |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 78727 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.01752