# Where Do Apparent LLM Clinical Triage Failures Arise? Localizing the Multiple-Choice Format Effect

arXiv: https://arxiv.org/abs/2605.29889 · HTML: https://mlsanity.com/paper/2605.29889/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Builds on 13 of its 68 references that are in the map.
- First posted on arXiv 28 May 2026.

## Builds on (most cited first, 13 of 13)

- [Sparse Autoencoders Find Highly Interpretable Features in Language Models](https://mlsanity.com/paper/2309.08600/) (2023)
- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://mlsanity.com/paper/2305.04388/) (2023)
- [Steering Language Models With Activation Engineering](https://mlsanity.com/paper/2308.10248/) (2023)
- [Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting](https://mlsanity.com/paper/2310.11324/) (2023)
- [Discovering Latent Knowledge in Language Models Without Supervision](https://mlsanity.com/paper/2212.03827/) (2022)
- [Large Language Models Are Not Robust Multiple Choice Selectors](https://mlsanity.com/paper/2309.03882/) (2023)
- [Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models](https://mlsanity.com/paper/2403.19647/) (2024)
- [Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2](https://mlsanity.com/paper/2408.05147/) (2024)
- [Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders](https://mlsanity.com/paper/2407.14435/) (2024)
- [k-Sparse Autoencoders](https://mlsanity.com/paper/1312.5663/) (2013)
- [Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control](https://mlsanity.com/paper/2405.08366/) (2024)
- [Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations](https://mlsanity.com/paper/2603.18353/) (2026)
- [Pattern Recognition or Medical Knowledge? The Problem with Multiple-Choice Questions in Medicine](https://mlsanity.com/paper/2406.02394/) (2024)

## Built on by (most cited first, 4 of 4)

- [High-Stakes Decisions with Language Models: Insights from Emergency Triage](https://mlsanity.com/paper/2608.01361/) (2026)
- [What Does Activation Steering Control? Attribution Across Answer Encodings and Output-Sensitive Subspaces](https://mlsanity.com/paper/2608.22985/) (2026)
- [Decodable but Misrouted: Sparse Features Uncover a Readout Gap in Vision-Language Models for Harmful Meme Detection](https://mlsanity.com/paper/2609.18860/) (2026)
- [The Persona Is Still There, but Who Is Speaking? Latent Identity Reversion in Persistent AI Agents](https://mlsanity.com/paper/2610.01490/) (2026)

## Abstract (arXiv)

LLM evaluations using clinician-authored triage vignettes have reported substantial under-triage under constrained multiple-choice testing. Yet model performance on the same clinical cases can change when responses are generated in free text. We test whether this format effect appears while the case is processed or when clinical information is mapped to the final answer. Using sparse-autoencoder (SAE) features in Gemma 3 4B/12B IT and Qwen3-8B, we find that medical features fire on the shared clinical narrative under both formats but are inactive at the multiple-choice decision token. Emergency-tier information is linearly decodable from vignette representations with ROC-AUC 0.95--1.00 under both formats, with no significant format difference, but is attenuated at the decision token. Natural-language autoencoder verbalization and top-feature characterization associate that token with the multiple-choice scaffold. In a direct linear projection, the identified medical features contribute zero, whereas scaffold-peaking features account for over 91% of unsigned attribution in both Gemma models. Behaviorally, whether multiple choice improves or worsens performance depends on the model. Option-order shuffles rule out simple positional bias, and cases that differ between formats are usually one severity tier apart. Together, these findings place the strongest correlates of the format effect at answer selection while leaving open whether unmeasured clinical representations also differ. Code and data to reproduce experiments are available in the study repository. https://github.com/dafraile/SAEmad