EvilGenie: A Reward Hacking Benchmark

2025/11/26 by Jonathan Gabor, Jayson Lynch, Gabor, Jonathan +3 · 2 citations
Computer Science · #FOS: Computer and information sciences #I.2.7 #Logic, programming, and type systems #Machine Learning (cs.LG) #Software Engineering Research #Software Testing and Debugging Techniques

paper · pdf · doi:10.48550/arxiv.2511.21654

openalex publication_date 2025/11/26 · openalex created_date 2025/11/28 · openalex updated_date 2026/07/28

Abstract

We introduce EvilGenie, a benchmark for reward hacking in programming settings. We source problems from LiveCodeBench and create an environment in which agents can easily reward hack, such as by hardcoding test cases or editing the testing files. We measure reward hacking in three ways: held out unit tests, LLM judges, and test file edit detection. We verify these methods against human review and each other. We find the LLM judge to be highly effective at detecting reward hacking in unambiguous cases, and observe only minimal improvement from the use of held out test cases. In addition to testing many models using Inspect's basic_agent scaffold, we also measure reward hacking rates for three popular proprietary coding agents: OpenAI's Codex, Anthropic's Claude Code, and Google's Gemini CLI. We observe explicit reward hacking by both Codex and Claude Code, and misaligned behavior by all three agents. Our codebase can be found at https://github.com/JonathanGabor/evilgenie_inspect .

Citations

• ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases
• Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
• LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code
• SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
• GenProg: A Generic Method for Automatic Software Repair

Cited by

• Reinforcement Learning Towards Broadly and Persistently Beneficial Models
• ScopeJudge: Cost-Aware Pre-Execution Gating for Offensive Security Agents

Related

• School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs
• Natural Emergent Misalignment from Reward Hacking in Production RL
• Cheap Reward Hacking Detection
• HonestCyberEval: An AI Cyber Risk Benchmark for Automated Software Exploitation
• Detecting and Mitigating Reward Hacking in Reinforcement Learning Systems: A Comprehensive Empirical Study