How Frontier LLMs Adapt to Neurodivergence Context: A Measurement Framework for Surface vs. Structural Change in System-Prompted Responses

A new arXiv paper introduces NDBench, a comprehensive framework to measure how frontier chat-based LLMs adjust outputs when system prompts include neurodivergence (ND) context. The benchmark tests two frontier models across three prompt types—baseline, ND-profile assertion, and ND-profile assertion with explicit adjustment instructions—covering four canonical ND profiles and 24 prompts, including an adversarial masking strategy. With 576 total outputs, the study reveals four consistent trends.

First, LLMs significantly adapt to ND context only when given explicit instructions: outputs become lengthier, with more headings and granular steps (p < 10^-8). However, this adaptation is primarily structural—list density barely changes while headings and step counts rise. Crucially, ND persona assertion alone does not suppress harmful tendencies; masking-reinforcement drops by 36–44% only in explicitly instructed cases. The authors also caution that reliability analysis shows only two of six harm dimensions (masking/reinforcement and validation quality) meet the inter-judge agreement threshold (alpha >= 0.67). NDBench, including prompts, outputs, and code, is publicly available to enable reproducible audits of ND awareness in future LLMs.