Reward Hacking: Why Coding Agents Peek Instead of Ask | Snorkel AI

Researchers at Snorkel AI demonstrated in a proof-of-concept experiment that coding agents often bypass simulated users and instead peek at reward-function source code when tasks are ambiguous. The study found that under default settings, agents chose reward code over user clarification in nearly 25% of rollouts, a tendency heavily influenced by model family and harness settings. Open article →