AI Paper Atlas

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

Token-Efficient Change Detection in LLM APIs

Timothee Chauvin, Clément Lalanne, Erwan Le Merrer, Jean-Michel Loubes, Francois Taiani, Gilles Tredan

ICML 2026 · Large language models › Language, safety and interpretability · approach: Data and hallucination detection · 0 citations (OpenAlex)

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
use Border Inputs with only output tokens instead of black-box MET baselineB3IT · TinyChangeaverage ROC AUC: 0.61 -> 0.9Improved
sample 5 prompts 10 times instead of one prompt 50 times (ablation)B3IT · TinyChangeaverage ROC AUCImproved
increase number of prompts at fixed cost (ablation)B3IT · TinyChangeaverage ROC AUCImproved
reduce fine-tuning to a single step (ablation)B3IT · TinyChange fine-tuning scaleROC AUC: 0.87 ROC AUC at 1 stepMixed
add reasoning-budget query strategy for BI discovery (ablation)B3IT BI search on OpenRouter endpoints · 131 endpointscoverage with at least one BI (%): 76% -> 78%Improved
search 2k prompts instead of 1k (ablation)B3IT BI search · 131 endpointscoverage with at least one BI (%): 73% -> 76%Improved
raise sampling temperature above 0 for BI search (ablation)B3IT BI search · 93-endpoint candidate setBI coverage vs qualityMixed
compare logprob-free BI test with grey-box LTB3IT · TinyChangeaverage ROC AUCNo clear effect

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