Tag
提出 Recursive Environment Scaling (RES) 框架，在保持任务目标不变的前提下递归变换基准环境以提升难度，既改善长程智能体评估，也使 SFT 在九组模型-基准对比中全部优于原始轨迹微调。
StoreBench introduces a live-commerce RL environment where autonomous agents run an online apparel store with 29 merchant tools, testing long-horizon planning under uncertainty. No frontier LLM matches a scripted smart-triage baseline (best, DeepSeek-V4-Pro, passes 49% vs 97%), while GRPO post-training on Qwen3.5-27B substantially improves held-out performance.
AutoEnvScaling treats RL environment design as a terminal task, letting a proposer agent automatically build training environments from solver rollouts to automate the data flywheel. The method raises the valid-environment rate up to 25x at 75% lower cost, improves Qwen terminal-agent scores by up to 12.2 points, and enables a single model to co-evolve as both solver and generator, a step toward recursive self-improvement.
EnvHarness is a programmable wrapper framework that dynamically adapts static environments to enhance LLM agent training, outperforming existing methods by providing superior optimization signals for reinforcement learning across multiple benchmarks.
This paper proposes the LLM-as-Environment-Engineer framework, where a policy model analyzes failures to automatically redesign the training environment for reinforcement learning, and introduces MAPF-FrozenLake as a controllable testbed. The framework, using Qwen3-4B, outperforms larger models like GPT and Gemini, showing that policy learning improves the model's ability to diagnose weaknesses.
This paper introduces LLM-as-Environment-Engineer, a framework where LLMs design their own training environments for reinforcement learning in multi-agent reasoning tasks, enabling self-improving training that surpasses larger proprietary models.
A technical blog post that explains how to build agent training systems from first principles using a text-to-diagram agent as an example, covering environment definition, teacher trajectory generation, student fine-tuning, and reinforcement learning.