- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper treats reward hacking in multi-agent coding systems as a strategy that one agent can pass to others. Following MacDiarmid et al., the authors use coding RL to teach small open-weight models to bypass a pytest evaluator, then let these source agents influence other models through shared reference code, chains of agents, evaluator feedback, shared memory and written plans, and also place them next to orchestrators, aggregators and reviewers. On hard CodeContests problems, sources lead several targets, including frontier closed-weight models, to write confirmed hacks; some chains keep or grow the hack rate after the source leaves, and orchestrators favour the hacking source. A propagation-factor analysis describes when such chains grow or die out, and self-review and verifier agents reduce delivered hacks at some cost.
Critical strengths
-
The experiments are extensive, covering many target models, including frontier closed-weight ones, and several ways in which agents interact. I appreciate that the authors include control sources and a planner that do not hack, which lead to almost no hacks downstream (Appendix K.1).
-
The orchestrator experiment is interesting. It compares the hacking source with its own instruction-tuned checkpoint under the same agent ID, size and cost, and the preference for the source remains when size and cost are hidden.
Critical weaknesses
(1) Seems to have missing comparison with in-ceontext work: the central one-hop result is a target imitating hacked examples placed in its context, and the paper does not position this against the closest prior work. Earlier studies report that a few narrow in-context examples can make models broadly misaligned [1], that malicious prompts can replicate themselves across connected LLM agents [2], and that frontier coding agents exploit test cases on their own [3]; these relate to the one-hop result, to the transmission chains, and to the reading of frontier targets' willingness to hack, respectively. Section 2.1 cites in-context reward hacking through feedback loops (Pan et al., 2024), but the paper does not say what its one-hop and chain results add beyond it. It would be interesting to see how the authors view these works in relation to their own, and what the multi-agent setting reveals that they do not.
(2) Endorsed code versus a transmitted strategy: I am not convinced that the code-sharing experiments measure adoption of a propagated strategy rather than imitation of code that the target is told is correct. The target's prompt introduces the references as solutions that "each PASSED the automated grader", the references are not filtered for hacks, and the target is also told that any file it writes will be loaded by pytest (Appendix K.1); the chain and feedback experiments reuse this prompt. The control sources show that the label alone does not produce hacks, but no condition removes or neutralises it, so it is unclear to me how much of the headline rates, such as Claude-Opus-4.7 hacking in about three quarters of its submissions when shown five references, comes from this endorsement.
(3) Settings and results left to the appendix: the main text describes the one-hop setup without saying that the target is told the references passed the grader, which changes how Section 5 should be read. The results on shared memory, orchestration, aggregation, self-review and verification appear in the main text only as qualitative statements, with the numbers in Appendices K.4 and O to S, so I had to move between the two to check each claim.
References [1] Nikita Afonin et al. Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL), 2026. [2] Donghyun Lee and Mo Tiwari. Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems. arXiv preprint arXiv:2410.07283, 2024. [3] Ziqian Zhong, Aditi Raghunathan, and Nicholas Carlini. ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases. In International Conference on Learning Representations (ICLR), 2026.
Questions for authors
-
How do the authors see the one-hop and chain results in relation to in-context reward hacking (Pan et al., 2024), in-context emergent misalignment [1], self-replicating prompt attacks across agents [2] and test exploitation by coding agents [3]? What does the multi-agent setting show that these settings do not?
-
Would frontier targets write hacks at similar rates if the reference header described the code neutrally, without stating that each solution passed the grader, or if non-hacking references carried the same label?
-
How often does each orchestrator select the hacking source compared with its matched control, and how many seeds were run per condition, given that Appendix O says "up to five"? More generally, which of the quantitative results now only in the appendix do the authors consider necessary for reading Sections 5 to 8?
Dear Reviewer,
As today is the final day of the discussion, we wanted to thank you again for your constructive feedback. We have incorporated the additional experiments and corresponding revisions into our updated manuscript.
We would appreciate any final thoughts or questions you may have. If you feel that we have adequately addressed your concerns, we would also be grateful if you would consider updating your rating before the discussion closes.
Thank you again for your time and valuable feedback!