CoKL regularizes only the correctness-conditioned response distribution, decoupling total correctness from mode preservation and improving retention-adaptation balance in LLM RL.
Advances in Neural Information Processing Systems , volume=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Toward Plasticity-Preserving KL Regularization for Capability Retention in LLM Reinforcement Learning
CoKL regularizes only the correctness-conditioned response distribution, decoupling total correctness from mode preservation and improving retention-adaptation balance in LLM RL.