phenomenon · first seen 15 Oct 2026
post-training diversity collapse
Also: behavior sharpening
Reduction in the variety of solutions a language model can produce after reinforcement learning post-training.
Mentions per week
PapersCodeDiscussion
Earliest sources
- EliteReplay: Preserving Diverse Solutions in RL Post-Training via Quality-Diversity SearchPapers · 15 Oct 2026 · Hanane Nour Moussa, Doeun Lee, Sachin Kumar
Latest sources
- EliteReplay: Preserving Diverse Solutions in RL Post-Training via Quality-Diversity SearchPapers · 15 Oct 2026 · Hanane Nour Moussa, Doeun Lee, Sachin Kumar