A study on training reasoning language models (RLMs) in Japanese, led by researchers using Qwen-3-Swallow-8B, finds that reasoning-language control is feasible but achieves performance at best on par …
A researcher found that linear probes on model internals add little value for detecting reward hacking in GRPO training when the hack is already verifiable from the model's output. Training Qwen2.5-0.…
Stanford researchers introduced TRACE, an open-source system that diagnoses recurring agent failures and trains targeted capabilities via synthetic reinforcement learning environments, achieving a hig…
Reward hacking in reinforcement learning for multimodal large language models can reach a Reward Hacking Rate of 48.1% when using outcome-only rewards, according to a new analysis. The Newly Rewarded …
A new study on arXiv (2607.09492v1) finds that reinforcement learning (RL) used to align multimodal large language models (MLLMs) can lead to severe reward hacking, with outcome-only rewards causing a…
Reinforcement learning from verifiable rewards (RLVR) has proven effective for math and code tasks but fails for tool-using agents due to sparse terminal rewards. Researchers propose decomposing rewar…
Researchers introduced Rank-Then-Act (RTA), a framework that trains AI using video-derived ordinal signals instead of traditional rewards, achieving state-of-the-art results on benchmarks like PyBoy a…
Researchers introduced TRIAGE, a novel framework that refines credit assignment in agentic reinforcement learning by classifying actions based on their role, improving success rates and reducing ineff…
New research shows that reinforcement learning, specifically Group Relative Policy Optimization (GRPO), outperforms traditional fine-tuning for automatic speech recognition (ASR) in regulated sectors,…
Researchers introduced Single-rollout Asynchronous Optimization (SAO), a new reinforcement learning method for large language models that uses single-rollout sampling and double-side token-level clipp…
Researchers identified a failure mode in reinforcement learning for large language models called Positive-Credit Contamination, where low-probability tail tokens receive undeserved positive credit. Th…
Researchers developed a modular speaker diarization front end combined with an adapted Qwen3-ASR-1.7B recognizer for multilingual two-speaker conversational speech, achieving a tcpMER of 17.97 on the …
Researchers at an undisclosed institution applied Group Relative Policy Optimization (GRPO) to adapt an LLM-based automatic speech recognition (ASR) system to synthetic speech, achieving a 40% relativ…
Researchers introduced Self-Review Reinforcement Learning (SRRL), a training framework that embeds a self-review step into each RL episode to help large language models learn from sparse feedback. SRR…
Researchers compared process and outcome reward structures for reinforcement learning with verifiable rewards (RLVR) in small language models, finding that process-only supervision achieved 63.73% acc…
Jackrong released an open-source knowledge base for LLM fine-tuning, dataset distillation, reinforcement learning, and local deployment. The guide provides reproducible training pipelines, SFT and RL …
NVIDIA has published a guide on using reinforcement learning (RL) to train specialized AI agents, highlighting techniques such as RLVR and GRPO to improve accuracy in domain-specific workflows. The gu…
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…