AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

A Coin Flip for Safety: LLM Judges Fail to Reliably Measure Adversarial Robustness

Leo Schwinn, Moritz Ladenburger, Tim Beyer, Mehrnaz Mofakhami, Gauthier Gidel, Stephan Günnemann

ICML 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Adversarial prompts for language models · 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
restrict evaluation to 41 most consistent behaviors (ReliableBench)LLM safety judges · HarmBench subsetaverage judge accuracy: 53% -> 70%Improved
correct ASR by judge precisionattack success rate reporting · Gemma-3-1Bexpected ASR: BoN falls to lowest corrected ASRImproved
collect multiple judge-positive samples per behaviorattack evaluation · Gemma-3-1BP(at least one true positive)Improved
average scores over multiple judges (ensemble) (ablation)LLM safety judges · Llama-3.1-8Bcorrelation with human rating (R2): 0.02 -> 0.05No clear effect
adjust classification threshold of judge (ablation)JailJudge · Llama-3.1-8BAUROC: 0.48 to 0.64No clear effect
use high inter-judge agreement as proxy for correctnessJudge Concordance score · HarmBench behaviorsjudge accuracyNo clear effect
continue attack optimization (more GCG/GCG-R/PAIR steps)adversarial attacks · HarmBench subsethuman-rated harmfulnessNo clear effect
switch victim model from Llama-3.1-8B to Gemma-27BHarmBench judge on direct prompting · HarmBench subsetjudge accuracy: 45% -> 58%Improved
use judge-feedback in attack (GCG-REINFORCE)attack · HarmBench subsetJailJudge AUROC: 0.64 (GCG) -> 0.48Hurt
evaluate judges on samples where all other judges failLlamaGuard-3 · JudgeStressTestaccuracy: 0.54 -> 0.17Hurt

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