Clipping Bottleneck: Stabilizing RLVR via Stochastic Recovery of Near-Boundary Signals
Analysis identifies hard clipping as bottleneck in RLVR training; proposes stochastic recovery of near-boundary signals to stabilize GRPO optimization.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Analysis identifies hard clipping as bottleneck in RLVR training; proposes stochastic recovery of near-boundary signals to stabilize GRPO optimization.
Advantage Collapse Rate metric diagnoses GRPO training failure in RL preference optimization for LLM reasoning.
GRPO-based approach trains small Qwen3-1.7B model for zero-shot Text-to-SPARQL generation on DBLP using outcome-based RL rewards.
SAGE framework improves VLM spatial reasoning via geometric logic consistency and duality operations in GRPO training.
AlphaGRPO applies Group Relative Policy Optimization to unified multimodal models for reasoning-based text-to-image generation and self-reflective output refinement.
Sparse-to-dense reward allocation principle optimizes labeled data use by routing sparse rewards to exploratory models and dense rewards to distillation targets.
GRPO fails under binary rewards due to gradient starvation when all group responses are correct/wrong; group-mean centering fix demonstrated on Qwen3.5.
Positive-only policy optimization with implicit negative gradients improves RLVR for LLM reasoning by avoiding coarse failure discrimination in GRPO-style training.
EP-GRPO fixes credit assignment failures in GRPO-based LLM reasoning via token-level entropy, polarity-aware rewards, and zero-variance collapse mitigation.
Latent-GRPO stabilizes reinforcement learning in latent reasoning space by addressing probability density and sampling mechanism shifts.
GRPO-VPS enhances reasoning via verifiable process supervision and belief-probing for improved credit assignment.
EVPO addresses critic noise in sparse-reward LLM RL by casting baseline selection as Kalman filtering, challenging PPO/GRPO design tradeoffs.
M²GRPO integrates Mamba state-space policy with group-relative policy optimization for multi-agent cooperative pursuit in biomimetic underwater robots.
As LLMs transition from simple text generation to complex reasoning, reinforcement learning (RL) plays a central role. Algorithms like Group Relative Policy... As LLMs transition from simple text generation to complex reasoning, reinforcement learning (RL) plays a central role…
UDM-GRPO: first framework integrating Uniform Discrete Diffusion Models with GRPO RL; addresses training instability in discrete generative modeling.
CUTS decoding strategy prevents mode collapse in GRPO on saturated reasoning benchmarks by enforcing structure-preserving exploration.