Yuepeng Hu
Meta Superintelligence Labs; Duke University
Titles as the sources gave them. In 1 entry, with 5 checked quotations.
Who they are
Roles and parties from Wikidata and from the sources in the Base, as precise as each gives them. When we compare what groups say, each statement counts in the group the person belonged to that day.
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.”
“Even agents based on small open-weight models induce reward hacking in stronger closed-weight targets.”
“Several frontier closed-weight targets show high propagation, whereas GPT-5.5 remains near zero despite its strong capabilities.”
“Code, memory entries, and plans from other agents should be treated as untrusted inputs.”
“First, our experiments concern Python coding tasks, and our reward-hacking criterion covers three known pytest-level mechanisms.”
In the Base
Every entry they appear in.
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.