Search papers, labs, and topics across Lattice.
This paper investigates why LLMs perform poorly on clinical triage benchmarks when constrained to multiple-choice outputs, despite performing better with free-text generation. Using sparse autoencoders on Gemma and Qwen models, the authors find that medical features are activated during the clinical narrative regardless of output format, but are suppressed at the multiple-choice decision token. Further analysis reveals that scaffold and format features, rather than medical knowledge, drive the final decision logits, suggesting the failure stems from output format constraints.
LLM triage failures in multiple-choice settings aren't due to a lack of medical knowledge, but rather a disconnect between internal representations and the constrained output format.
Patient-voiced clinical-triage benchmarks report high under-triage rates for consumer LLMs for constrained multiple-choice output, yet the same cases score differently with free-text. We ask whether output format changes the model's \emph{clinical representation} or only the mapping from a preserved representation to an answer. Using sparse-autoencoder (SAE) features in Gemma 3 4B/12B IT and Qwen3-8B, we find the same medical features fire on the shared clinical narrative under both formats but go {silent} at the multiple-choice decision token in all the cases at every model. Three independent methods (natural-language autoencoder verbalization, decision-token logit attribution, and top-feature characterization) agree that scaffold and format features, but not medical features, drive the decision logits. Behaviorally, the multiple-choice penalty inverts under both structured and natural-language input, option-order shuffle rules out positional bias, and the gap is dominated by off-by-one decision (the model picks an adjacent acuity letter to the gold answer) rather than knowledge failure. Thus, the failure originates in the output format and not in the clinical representation.
