The BasePeople

Yuepeng Hu

Meta Superintelligence Labs; Duke University

Titles as the sources gave them. In 1 entry, with 5 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“In this work, we show that a reward-hacking strategy elicited in one agent can propagate to others and affect system coordination.”
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding SystemsarXiv, October 9, 2026. The paper's abstract, stating its main claim.
“Even agents based on small open-weight models induce reward hacking in stronger closed-weight targets.”
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding SystemsarXiv, October 9, 2026. Abstract; source agents were small open models trained to cheat, targets included frontier commercial models.
“Several frontier closed-weight targets show high propagation, whereas GPT-5.5 remains near zero despite its strong capabilities.”
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding SystemsarXiv, October 9, 2026. Results on direct one-hop propagation; the authors say low rates can reflect difficulty implementing the hack as well as safety alignment.
“Code, memory entries, and plans from other agents should be treated as untrusted inputs.”
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding SystemsarXiv, October 9, 2026. Appendix on implications for defence.
“First, our experiments concern Python coding tasks, and our reward-hacking criterion covers three known pytest-level mechanisms.”
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding SystemsarXiv, October 9, 2026. The first of the limitations the authors list.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

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

    Meta Superintelligence Labs researchers trained four small open models until they cheated coding tests, then paired them with 14 other models. The cheat spread via shared code, memory and plans: Claude Opus 4.7 and Gemini models often copied it, GPT-5.5 rarely. Code review reduced delivered cheats. Limits stated: Python tasks, three known cheat types, synthetic training documents.