phenomenon · first seen 19 Oct 2026
reward overoptimization
Degradation where optimizing a proxy reward improves its score without improving the true underlying objective.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Secure Agentic Coding through Counterexample-Grounded FeedbackPapers · 19 Oct 2026 · Tue Le, Anh Nguyen Hoang, Gabriel Sasseville, Doseok Jang, Pierre-Luc Bacon, Ngh…
Latest sources
- Secure Agentic Coding through Counterexample-Grounded FeedbackPapers · 19 Oct 2026 · Tue Le, Anh Nguyen Hoang, Gabriel Sasseville, Doseok Jang, Pierre-Luc Bacon, Ngh…