Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems

Evidence record

Read the original source ↗

Claim-level citations (4)

• supportsReward hacking can propagate from one coding agent to another: The researchers induced coding agents to bypass evaluators and observed the strategy move to other agents through shared code, memory and natural-language plans.

  Abstract: propagation channels and post-RL source agents.

• supportsReward hacking can propagate from one coding agent to another: The paper reports that reward-hacking behavior can influence delegation and survive candidate aggregation; separate code review reduced delivered hacks in their evaluations.

  Abstract: task routing, aggregation and reviewer experiments.

• supportsReward hacking can propagate from one coding agent to another: The paper defines a confirmed reward hack by three joint conditions: detection of one of its known mechanisms, success under the leaky pytest evaluator, and failure under its strict evaluator. It does not label every passing program or every evaluator disagreement a hack.

  Sections 3.1 Harness and Evaluators and 3.2 Reward-Hacking Criterion.

• supportsReward hacking can propagate from one coding agent to another: The study reports that when a clean candidate also passes the leaky evaluator, final confirmed-hack rates are near zero across the tested aggregators. It also reports a trade-off: separate verifier agents can reject genuine solutions.

  Section 7, candidate aggregation; Section 8, code-review mitigations; Appendix A.