Researchers introduced the Relative Surprisal Index (RSI), an information-theoretic metric that balances token probability and entropy to improve reinforcement learning for large language models. The …
Researchers developed a closed-form model of Group Relative Policy Optimization (GRPO) training dynamics, predicting reward trajectories and stability thresholds from first principles. The model subsu…
Researchers introduced BV-Blend, a critic-free reinforcement learning framework that stabilizes advantage estimation for aligning large language models by blending prompt-local on-policy statistics wi…
Researchers propose ATOD, a hybrid online distillation algorithm that combines on-policy distillation and reinforcement learning to train small language-model agents for multi-turn tasks. ATOD uses an…
Researchers introduced Retroactive Advantage Correction (RAC), a method for reinforcement learning from human feedback (RLHF) that handles delayed reward signals. RAC reduces policy bias by up to 47.9…
Researchers propose Tandem Reinforcement Learning (TRL), extending the tandem training paradigm to reinforcement learning with verifiable rewards (RLVR). Training Qwen3-4B-Instruct on competition math…
Researchers at arXiv identified a memory-update gap in LLM agents, where accuracy drops from 92% to 77% when using bounded memory instead of full context, even with frontier models like gpt-5.4. The g…
Weibo's AI group released VibeThinker-3B, a 3-billion-parameter model that achieves 94.3 on AIME and outperforms Claude Opus 4.5 on competition reasoning, but collapses on general knowledge tasks. The…
Researchers at ACL 2026 found that entropy collapse in LLM reasoning, which undermines test-time scaling, is driven by premature overconfidence at a small set of critical tokens. They proposed SCOPE, …
Nvidia researchers introduced Zone of Proximal Policy Optimization (ZPPO), a method that uses a replay buffer to repeatedly expose student models to hard questions, improving rollout accuracy without …
AI progress is driven primarily by massive amounts of domain-specific human expert data and reinforcement learning, not by architectural breakthroughs, creating a data black hole that powers model cap…
Decagon deployed self-evolving AI agents that automatically improve their own prompts without human intervention, using a feedback loop that tests and promotes winning versions. The team found that th…
Researchers propose RODS (Reward-driven Online Data Synthesis) to address the depletion of informative samples in multi-turn tool-use reinforcement learning. RODS uses progress reward variance as a bo…
A developer post-trained a language model to roll a die, revealing that models default to outputting 4 due to training data bias. Standard reinforcement learning fails to encourage exploration, but ad…
Researchers introduced CoRA, a GRPO-based reinforcement learning framework that aligns LLM confidence with rationale quality in chain-of-thought reasoning. Across MedQA, MathQA, and OpenBookQA, CoRA r…
A new study finds that post-training techniques, including supervised fine-tuning and reinforcement learning, can transform large language models into effective medical coders for ICD coding, challeng…
A developer implemented the SDAR gate, a gated distillation mechanism for reinforcement learning with language models, in PyTorch. The gate uses a sigmoid function to weight a per-token KL divergence …
Rudrite Research published a comprehensive comparison of AI and ML approaches, covering 14 side-by-side analyses of techniques such as Transformers vs Mamba, FlashAttention vs PagedAttention, and PPO …
Researchers developed a variance-aware reward framework using Group Relative Policy Optimization (GRPO) to improve heart-focused medical question answering in large language models. The approach, whic…
Researchers introduced VideoKR, a large-scale training corpus of 315,000 video reasoning examples over 145,000 expert-domain videos designed to strengthen knowledge- and reasoning-intensive video unde…