Research map / Large language models / Language, safety and interpretability
Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis
Darshan Girish Deshpande, Anand Kannappan, Rebecca Qian
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 |
|---|---|---|---|
| evaluate in contrastive cluster setting instead of isolated classification (ablation) | GPT-5.2 · TRACE | Detection Rate: 45% -> 63% | Improved |
| increase cluster size from N=1 to N=10 (ablation) | Claude-4.5-Opus · TRACE | Detection Rate: 34% absolute difference | Improved |
| increase cluster size from N=1 to N=5 (ablation) | all evaluated LLMs · TRACE | Match Rate: approximately 35% Match Rate improvement | Improved |
| increase cluster size from N=5 to N=10 (ablation) | all evaluated LLMs · TRACE | Match Rate | No clear effect |
| increase benign ratio from 25% to 90% (ablation) | all evaluated LLMs · TRACE | Match Rate | Improved |
| semantic rather than syntactic reward hack categories | evaluated LLMs · TRACE | Match Rate: 0.6-0.95 -> 0.0-0.4 | Hurt |
| provide ground truth to the LLM judge instead of comparative matching (ablation) | LLM judge · TRACE | Match Rate (judge) | Improved |
Closest papers by meaning
- Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.90)
- Is it Thinking or Cheating? Detecting Implicit Reward Hacking by Measuring Reasoning Effort (ICLR 2026 · similarity 0.89)
- Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking (ICLR 2025 · similarity 0.87)
- SecCodePRM: A Process Reward Model for Code Security (ICML 2026 · similarity 0.87)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.86)
- IFDECORATOR: Wrapping Instruction Following Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.86)
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.
