Tag
The Athanor paper from ByteDance and other institutions shows 4 Claude Opus 5.5 agents, each occupying an 8-GPU machine, autonomously completing post-training in 104 hours with no human intervention, raising Qwen3-4B-Base to official Instruct level (dev set 15.9→67.3, held-out 9.2→48.2).
WorkRover is a framework for training long-horizon cowork agents through recursive environment evolution, introducing Factorial Compatibility Replay and Verifier-to-Turn Credit Assignment to improve training signals. Post-training Qwen3.5-35B-A3B into WorkRover-35B-A3B yields large gains on JobBench, OfficeQAPro, and GDPvalv2 benchmarks.
Maximilien Labonne highlights an open-sourced guide to training models with RL inside real agent harnesses like Claude Code, Codex, and OpenCode, enabling training of any model (e.g., LFM2.5-2.6B) on any task set across harnesses.
DeepSeek Elastic Compute (DSec) is a sandbox infrastructure for effective agentic training of large language models at scale, featuring elastic execution with unified SDK and integration with reinforcement learning frameworks.
TRACE is a new self-improvement approach where an AI agent identifies the missing capabilities behind its own failures and trains itself to address them. TRACE-trained Qwen3.6-27B achieves 73.2% on SWE-bench Verified, outperforming much larger models with fewer training rollouts.