Tag
Cowork-Gym 通过 build-execute-revise 循环与有状态用户模拟器大规模合成协作式 agent 环境（64.9K 个、14 个应用场景），并用于后训练 9B/27B/35B-A3B Qwen 模型，在六个 agentic 基准上分别提升 8.7、6.4、4.1 分，其中 27B 版本在 PawBench 和 ClawEval 上取得最高平均分。
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.
Proposes a 'Truth Engine' architecture in which an environment synthesizer creates novel closed-loop micro-worlds, a frontier solver explores them without human demonstrations, and deterministic verifiers filter proven artifacts that get distilled into the base model — forming a self-compounding flywheel intended to push AI beyond the boundary of human knowledge.
Skill2Env is a capability-oriented framework that synthesizes executable environments from skills to train general AI agents, using task blueprints and iterative hardening to improve agent performance across benchmarks.
EnvCraft is an automated framework for synthesizing executable environments to address the scarcity in Agentic RL training, showing significant performance gains on claw-like and general tool-use benchmarks using Qwen models.
AgentMercury introduces a scalable framework for synthesizing verifiable environments from business scenarios, enabling reinforcement learning agents to improve performance on enterprise and out-of-domain benchmarks.
A new research paper introduces AgentMercury, a scalable framework for synthesizing executable environments from business scenarios, which improves agent training performance on various benchmarks.
Envs-FORGE is a prompting policy for synthesizing training environments for reinforcement learning agents, converting verifier rewards into per-seed actions to improve performance on benchmarks like SWE-bench and Terminal-Bench.
EnvFactory automates the creation of executable tool environments and natural multi-turn trajectories for training LLMs with agentic reinforcement learning, achieving superior performance on benchmarks like BFCLv3 and MCP-Atlas with fewer environments than prior work.
This paper proposes EvoEnv, a method where language models construct verifiable Python environments for self-improvement through reinforcement learning, achieving a 3.3% relative gain on Qwen3-4B-Thinking.
EnvScaler is an automated framework for scaling tool-interactive environments for LLM agents through programmatic synthesis, creating 191 diverse environments and 7K scenarios to improve agent performance on multi-turn, multi-tool interactions.
Agent-World introduces a self-evolving training framework for general agent intelligence that autonomously discovers real-world environments and tasks via the Model Context Protocol, enabling continuous learning. Agent-World-8B and 14B models outperform strong proprietary models across 23 challenging agent benchmarks.