Exhaustive circuit mapping of a biological foundation model reveals massive redundancy and annotation bias.
arXiv · March 13, 2026 · 2603.11940
Why it matters
By tracing 1.3M downstream edges in Geneformer, the authors show that prior pairwise testing missed a heavy-tailed hub architecture where 1.8% of features control most connectivity. It provides a new standard for mechanistic interpretability in biological AI.
From the abstract
Mechanistic interpretability of biological foundation models has relied on selective feature sampling, pairwise interaction testing, and observational trajectory analysis. Each of these can introduce systematic bias. Here we present three experiments that address these limitations through exhaustive circuit tracing, higher order combinatorial ablation, and causal trajectory steering in Geneformer, a transformer based single cell foundation model. First, exhaustive tracing of all 4065 active spar