大規模言語モデルの臨床推論が失敗する理由：説明可能な深層学習からの洞察

Medical large language models (LLMs) achieving high benchmark accuracy exhibit unexplained variability in clinical tasks, producing errors that clinicians cannot safeguard against. Sparse autoencoders offer a mechanistic interpretability approach to reveal how models represent medical knowledge and why they fail.

We evaluated clinical reasoning stability in GPT-5, MedGemma-27B-Text-IT, and OpenBioLLM-Llama3-70B using 355 systematic perturbations in physician-validated oncology cases, comparing staging and treatment against National Comprehensive Cancer Network (NCCN) and American Joint Committee on Cancer (AJCC) 8/9th edition guidelines. We trained sparse autoencoders on 1 billion tokens from 50,000 Medical Information Mart for Intensive Care (MIMIC-IV) clinical notes to analyze how each architecture encodes polysemous medical terms, performed 850 ablation experiments, and tested a two-stage retrieval intervention for sense disambiguation.

Models exhibited dramatic reasoning instability. OpenBioLLM staging accuracy shifted from 45.9% to 99.1% based solely on prompt format. For cases with intentionally insufficient information, MedGemma generated definitive staging 100% of the time while GPT-5 appropriately indicated uncertainty, yet GPT-5 was only 7% accurate in staging IVB tumors. Overall, AJCC concordance was 97.5% for GPT-5 versus 26.8% for MedGemma and 31.0% for OpenBioLLM. Sparse autoencoder analysis revealed MedGemma's top-10 features showed 77.8% overlap across word senses (e.g., "cardiac arrest" versus "respiratory arrest") versus 13.6% for OpenBioLLM, indicating hierarchical versus distributed encoding. Ablating context-encoding features degraded prediction by 16.4 nats in MedGemma versus 6.7 nats in OpenBioLLM. A retrieval intervention improved MedGemma disambiguation by 10.2% (p=0.022, paired t-test) but harmed OpenBioLLM by 2.0% (p=0.021).

Medical AI systems exhibit clinical reasoning fragility not captured by benchmark performance. We find that architecturally distinct models encode medical concepts differently, and interventions effective for one architecture may harm another. We recommend that safety validation must be architecture-specific, as benchmark equivalence does not imply functional equivalence.