Researchers have developed Guidance Contrastive Policy Optimization (GCPO), a new algorithm that assigns per-token credit by contrasting model predictions under positive and negative prompts, addressi…
Researchers introduced Frost Training, a method that improves Monte Carlo-based policy optimization for Cross-Entropy Games by exploiting the gradient of the reward function in embedding space. The te…
NVIDIA researchers released Polar, a rollout framework that trains language agents through reinforcement learning without altering their agent harnesses. Using GRPO on a Qwen3.5-4B base model, Polar i…
Researchers have developed EchoDistill, a self-distillation framework that improves Audio Large Language Models' resistance to real-world noise by aligning noisy student responses with clean semantic …
Researchers introduced Iterative Group Relative Policy Optimization (iGRPO), a two-stage reinforcement learning method that improves large language model reasoning by having the model generate and ref…
Here is a 2-3 sentence factual summary of the article: The article describes the process of migrating an online reinforcement learning (RL) training system from the vLLM V0 engine to the V1 rewrite, …
The Granite 4.1 family consists of dense, decoder-only LLMs (3B, 8B, and 30B parameters) trained from scratch on approximately 15 trillion tokens through a five-phase pre-training pipeline that progre…