Action Value Gradient enables deep policy-gradient reinforcement learning with fully incremental updates, without replay buffers, target networks, or batch updates.

Stream-X introduces Stream TD, Stream Q, and Stream AC, enabling deep reinforcement learning directly from a stream of experience without replay buffers or batch updates.

Detecting recurring and novel MDP contexts online with MCUSUM, while maintaining context-specific dynamics models and policies.

BAPR combines Bayesian online change-point detection, context-conditioned reinforcement learning, and robust ensemble SAC to adapt conservatism after abrupt regime changes.

DARLING is a model-free detection-restart framework for piecewise-stationary RL that periodically probes the environment for reward and transition changes and restarts the stationary learner when a change is detected.

SeCBAD is a deep reinforcement learning method for non-stationary environments with piecewise-stable latent contexts. It jointly infers the posterior over the current context and the current segment length, then conditions the policy on the inferred belief to adapt to abrupt within-episode context changes.

This paper studies why on-policy distillation is more parameter-efficient than RL and argues that OPD develops foresight by identifying useful modules and stable low-rank update directions early in training.

Why policy entropy collapses during RL for reasoning LLMs, how the entropy change reduces to a covariance between log-probability and logit change, and two covariance-based fixes, Clip-Cov and KL-Cov.

This paper uses large-k pass@k evaluation to argue that RLVR improves sampling efficiency without expanding the base model's reasoning boundary, while distillation from a stronger teacher does expand it.

This paper traces RLVR's pass@k shrinkage to negative interference and on-policy winner-take-all dynamics, and proposes SELF, a greedy-failure filter plus forward KL that improves large-k coverage over GRPO.

This paper shows RL forgets less than SFT at matched new-task accuracy, finds that KL from the base model on the new task predicts forgetting, and attributes RL's advantage to on-policy sampling.

This paper shows that some hard training examples stay unlearnable under GRPO despite receiving correct rollouts, rules out optimization-side explanations, and attributes the failure to flawed representations that mid-training may fix.

Reducing catastrophic forgetting in reinforcement learning with Elephant activation functions that produce sparse representations and sparse gradients.

During RLVR, prompts the model has already mastered quietly regress. ReMind keeps a review queue of mastered prompts and re-tests them with fresh rollouts, trying to catch regression while repair is still cheap.

This paper argues entropy regularization helps policy gradient methods by smoothing the objective rather than only by encouraging exploration, using exact gradients in a gridworld and random-perturbation landscape probes in MuJoCo.