AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Assistive Prompt Mediation: Evaluating Language Models Under Accessibility Constraints

Priyaranjan Pattnayak, Ishan Banerjee

ICML 2026 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
increase impairment severity from alpha 0.2 to alpha >= 0.6 (ablation)10 frontier LLMs · APM intent preservationIPR: mean IPR >= 4 at alpha=0.2, drops sharply at alpha >= 0.6Hurt
increase severity of N2 telegraphic noise from alpha 0.2 to 0.8 (ablation)LLM mediators · APMBRS (burden robustness): -5.81 -> -7.34Hurt
increase severity of N1 orthographic noise from alpha 0.2 to 0.8 (ablation)LLM mediators · APMBRS: 0.04 -> -0.39Hurt
mid-severity noise N2/N3 versus N1 orthographic noiseLLM mediators · APMfalse robustness (high IPR with burden inflation)Hurt
switch from English to Hindi prompts (ablation)LLM mediators · APMhallucination rate: 0.272 -> 0.396Hurt
switch from English to Chinese prompts (ablation)LLM mediators · APMhallucination rate: 0.272 -> 0.193Improved
use larger frontier models vs small modelsLLM mediators · APMhallucination rate: 0.452 (llama31-8b), 0.513 (ministral-3-3b) -> 0.183 (gpt-4o)Improved
small model ministral-3-3b protocol complianceministral-3-3b · APMPCR: PCR 0.826 but IPR 2.563 and HIR 0.513Mixed
judge with Gemini 3 Pro instead of GPT-5.2 (ablation)APM evaluation · judge robustnessranking correlation: rho approx 0.81-0.86No clear effect
measure usability (ABI, LCI) instead of intent preservationAPM severity curves · APMmedian failure alpha: ABI and LCI fail at alpha 0.4; IPR at 0.6Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy