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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| aggregate consistency over k=10 questions instead of k=1 (ablation) | C3 · GPQA, GPT-4.1-nano | recall at 5% FPR: ~35% -> ~65% | Improved |
| aggregate over k=20 instead of k=10 questions (ablation) | C3 · GPQA, GPT-4.1-nano | recall at 5% FPR: ~65% -> ~80% | Improved |
| ensemble C3 with trusted monitoring (mean of normalised scores) (ablation) | C3 · GPQA, GPT-5-nano prompted | recall at 5% FPR: C3 67.9 / TM 39.5 -> 87.0 | Improved |
| ensemble C3 with trusted monitoring (ablation) | TM · MATH, GPT-5-nano prompted | recall at 5% FPR: TM 99.6 -> 98.5 | No clear effect |
| control for model identity in lie detector (ablation) | lie detector · USACO | recall: 89% -> 2% | Hurt |
| add denial instruction against behavioral probe (ablation) | behavioral probe · prompted sandbaggers | admission rate: 74% -> 3.4% | Hurt |
| use deterministic error-recipe sandbagging prompts | C3 · USACO | recall 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-nano | recall: 99.6% -> 99.4% | No clear effect |
| honest-wrong control (ablation) | C3 · GPT-5.4 GPQA | recall: 48.0% -> 71.1% | Improved |
| correctness-matched resampling of benign answers (ablation) | C3 · GPQA GPT-4.1-nano | recall: 46.4% -> 48.6% | No clear effect |
| use C3 instead of trusted monitoring on discovered-incentive sandbagging | C3 · GPT-5.4 MATH | recall at 5% FPR: 10.8% -> 61.8% | Improved |
| use trusted monitoring instead of C3 on benign-prefill control | TM · Gemini 3.1 Pro ProtocolQA | false positive rate: 6.1% vs 60.5% | Hurt |
Closest papers by meaning
- AI Sandbagging: Language Models can Strategically Underperform on Evaluations (ICLR 2025 · similarity 0.87)
- Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models (NeurIPS 2025 · similarity 0.87)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.85)
- CASE-Bench: Context-Aware SafEty Benchmark for Large Language Models (ICML 2025 · similarity 0.84)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.83)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.83)
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.
