The abstract is read at the publisher; the summary is JClub's.
Across three ESMC protein language models (300M–6B parameters), concept detectability depends strongly on readout depth and unit choice, with 2.5–5.4× more annotated-residue concepts from dual-signed neuron poles than random rotations.
- Why it matters: Interpretability claims for protein language models are often untested against model-independent, residue-level ground truth, leaving uncertainty about which internal representations truly correspond to biological concepts.
- What they did: They analyze how concepts are decoded and scored across model layers and neuron units, compare linear versus nonlinear probes, test feed-forward neuron metrics under saturation, and evaluate annotation shuffles and leakage-controlled benchmarks.
- The result: Concepts become linearly readable along depth before erasure harms masked-token prediction; shuffled annotations label nearly every latent without held-out validation, while causal reliance appears in neuron groups, and a selected-state readout improves MCC by 5% on a leakage-controlled interaction benchmark.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.