Frozen NLA explanation caches and result files for an audit of demographic representational shift in clinical language models

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Frozen Natural Language Autoencoder (NLA) explanation caches, constructed demographic pair texts, and computed result files underlying the study. Contains the NLA explanations generated for every case-variant (Gemma-3-27B-IT at layer 41, and Llama-3.3-70B at layer 53, both via the Neuronpedia API), model completions, all identical-input noise-floor draws, the constructed minimal-pair texts, and the JSON result files from which every number reported in the paper is regenerated. Why a frozen cache. Served NLA pipelines drift at month scale: cached-versus-fresh distances of identical inputs measured roughly double the within-regime decoding-noise floor. Re-running the same inputs months later therefore does not reproduce the original explanations, and these frozen caches — not fresh API calls — are the reproduction substrate for the reported statistics. Contents. Explanation caches (Gemma primary, Llama replication, Sevgi held-out); model completions and output-concordance results; noise-floor draws across multiple estimates and draw-regimes; constructed ethnicity, age and sex pair texts; glaucoma bookend draws; character-delta and power analyses; computed statistics under two embedding models (BGE-large-en-v1.5, e5-large-v2); the pre-registration; and the sex-rewrite correction log. Attribution and modification notice (CC BY 4.0). sevgi_pair_texts.parquet and sex_pairs.parquet contain clinical vignettes from Sevgi et al., Complementary Human-AI Clinical Reasoning in Ophthalmology (arXiv:2510.22414), used under CC BY 4.0. These texts have been modified: demographic descriptors were inserted or substituted and pronouns rewritten, to construct the minimal-pair variants analysed here. The text_a/text_b columns are derivative works, not the original vignettes, which remain available from the authors. The vignettes are authored teaching cases mapped to the American Board of Ophthalmology exam blueprint, not real patient records. Not included. MedCaseReasoning case texts (MIT licensed, available as zou-lab/MedCaseReasoning on Hugging Face); records here reference them by case identifier only.

Links and sources

Comments