phenomenon · first seen 20 Oct 2026
reward misspecification
Training rewards that differ from the true task intent, causing models to optimize unintended behaviors.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Reward Misspecification Degrades Chain-of-Thought MonitorabilityPapers · 20 Oct 2026 · Tianyi Lorena Yan, John Hewitt
Latest sources
- Reward Misspecification Degrades Chain-of-Thought MonitorabilityPapers · 20 Oct 2026 · Tianyi Lorena Yan, John Hewitt