AI Briefing#reinforcement-learning
The latest AI and developer news about #reinforcement-learning, with the original source and a short summary.
The latest AI and developer news about #reinforcement-learning, with the original source and a short summary.
LG AI Research has unveiled Multi-Phase RL, a reinforcement learning framework capable of complex stage-by-stage transitions, along with its platform RL-Studio.
Using GPT-Red, an automated security tester developed by OpenAI, vulnerabilities in AI models were found and the safety of GPT-5.6 was improved by 6x.
This is an open-source project that provides a browser-based MMO game as a Gymnasium environment for agent training.
A technology was unveiled that performs RL post-training using 14 Macs distributed across 4 countries worldwide.
The team unveiled Pan, a 4B-parameter Minecraft agent that leverages internet-scale video data to learn goal-directed behavior.
It proposes the SAO methodology, which improves agent training performance by enhancing the stability and efficiency of asynchronous reinforcement learning.
NVIDIA has unveiled Molt, a PyTorch-native RL framework designed for agentic training of large-scale MoE models.
Prime Intellect has launched verifiers v1, an environment stack for agentic RL and evals.
It maximizes LLM inference efficiency by dynamically adjusting the number of speculative decoding tokens according to the hardware's compute and memory bandwidth state.
SAO is a technique that improves the stability and efficiency of agentic reinforcement learning through single-rollout sampling and token-level clipping.
This study proposes the TGPO algorithm, which strengthens temporal reasoning to enhance MLLMs' understanding of egocentric video.
By combining reinforcement learning with Flow Matching, the approach solves the error propagation problem that occurs in multi-turn image editing.
Leanstral 1.5, an open-source model based on Lean 4 specialized for Formal Verification, has been released.
This introduces research that improves HIP kernel generation performance for AMD GPUs using synthetic data and reinforcement learning.
The study found that training only a specific intermediate layer during RL training can achieve performance similar to full-parameter training.
This work proposes a new policy learning method that maximizes the efficiency of token selection (unmasking) in Diffusion LLMs through reinforcement learning.
Apple researchers have proposed the Ctrl-R framework to help LLMs effectively learn the diverse reasoning patterns needed for complex problem solving.
RL-finetuned VLMs show improved visual reasoning performance, but exhibit vulnerability to textual perturbations.
RadixArk has released Miles, an open-source framework that integrates SGLang, Megatron-LM, and Ray for large-scale LLM reinforcement learning (RL) post-training.
Ornith-1.0, based on Gemma 4 and Qwen 3.5, has been released as a self-improving open-source model for agentic coding.
Meta proposed new measurement and clustering methods to address the Reward Hacking problem that arises when Reward Models overreact to similar responses.
DeepReinforce has open-sourced the Ornith-1.0 series, agentic coding models that learn by building their own scaffolds.
The Qwen team has released Qwen-AgentWorld, a native language world model that simulates 7 domains.
A new RL recipe and a large-scale dataset, TMax-15k, designed to maximize terminal agent performance have been released.