Google DeepMind Introduces 'Evo2': Co-Evolving Agent Harnesses and Model Weights
Google DeepMind has introduced 'Evo2', a co-evolution framework that trains models using harness edit logs. Coupling proposer-reflector loops with SFT and inter
#ReinforcementLearning
TAU-HOME.COM stories tagged ReinforcementLearning.
Alibaba Qwen Unveils QwenGyre: An Elastic RL Framework for Extreme-Long-Horizon Agent Training
Alibaba's Qwen team has unveiled QwenGyre, an elastic reinforcement learning framework designed for extreme-long-horizon agent training, cutting end-to-end trai
Hugging Face Turns OpenAI Math Problems into Open-Source RL Environments
Hugging Face's CEO announced open-source reinforcement-learning environments built from OpenAI's math problems. He also disclosed the early limitation: the veri
Agent Lightning: RL Training for Existing Agents Without Rebuilding Them
A look at Agent Lightning, the Microsoft Research agentic reinforcement-learning framework: how it connects existing agents to RL training almost unchanged, wha
Why a 15M-Parameter Agent Is Claimed to Beat the Craftax Final Boss 44% of the Time
A verified-scope rundown of Dan Kondratyuk's October 7, 2026 agent meta-harness claim: a 15M-parameter Craftax agent with a claimed 44% final-boss win rate, abo
Laminar Releases 'flow-1': An RL Model for Agent Trace Error Detection at 23x Lower Cost
Laminar has unveiled flow-1, an RL model specialized in AI agent trace analysis and error detection. It matches GPT-6-sol in diagnostic quality while processing
OpenEnv Released: Hugging Face Turns Claude Code and Codex into RL Environments
Hugging Face released OpenEnv, an open-source capture proxy turning coding harnesses like Claude Code and Codex into RL environments without modifying harness o
Training 1T-Param Kimi K2.7 with RL in SkyRL: Two-Node (16x B300) 262K Full-Context Reference
SkyRL has documented a Kimi K2 INT4 LoRA recipe enabling full-context reinforcement learning for the 1T-parameter MoE model Kimi-K2.7-Code across just two 8-GPU
mjbatch: Open-Source C++ Thread Pool for Parallel MuJoCo Physics Simulations on CPU
Kevin Zakka's open-source mjbatch bypasses Python GIL via a C++ thread pool, running thousands of parallel MuJoCo simulations on standard CPUs with batched stat