AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Same Question, Different Lies: Cross-Context Consistency (C³) for Black-Box Sandbagging Detection

Lin Yulong, Pablo Bernabeu-Perez, Benjamin Arnav, Lennie Wells, Mary Phuong

ICML 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Prototype and concept dictionary models

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
aggregate consistency over k=10 questions instead of k=1 (ablation)C3 · GPQA, GPT-4.1-nanorecall at 5% FPR: ~35% -> ~65%Improved
aggregate over k=20 instead of k=10 questions (ablation)C3 · GPQA, GPT-4.1-nanorecall at 5% FPR: ~65% -> ~80%Improved
ensemble C3 with trusted monitoring (mean of normalised scores) (ablation)C3 · GPQA, GPT-5-nano promptedrecall at 5% FPR: C3 67.9 / TM 39.5 -> 87.0Improved
ensemble C3 with trusted monitoring (ablation)TM · MATH, GPT-5-nano promptedrecall at 5% FPR: TM 99.6 -> 98.5No clear effect
control for model identity in lie detector (ablation)lie detector · USACOrecall: 89% -> 2%Hurt
add denial instruction against behavioral probe (ablation)behavioral probe · prompted sandbaggersadmission rate: 74% -> 3.4%Hurt
use deterministic error-recipe sandbagging promptsC3 · USACOrecall at 5% FPR: 4.7-18.0%Hurt
restrict to questions where both models answer wrong (honest-wrong control) (ablation)C3 · MATH GPT-4.1-nanorecall: 99.6% -> 99.4%No clear effect
honest-wrong control (ablation)C3 · GPT-5.4 GPQArecall: 48.0% -> 71.1%Improved
correctness-matched resampling of benign answers (ablation)C3 · GPQA GPT-4.1-nanorecall: 46.4% -> 48.6%No clear effect
use C3 instead of trusted monitoring on discovered-incentive sandbaggingC3 · GPT-5.4 MATHrecall at 5% FPR: 10.8% -> 61.8%Improved
use trusted monitoring instead of C3 on benign-prefill controlTM · Gemini 3.1 Pro ProtocolQAfalse positive rate: 6.1% vs 60.5%Hurt

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