Research · Nov 4, 2026, 00:00 UTC

ToolSpace proposes tool-centric environment synthesis for training long-horizon agents

The framework mines execution traces of real MCP services to build tool-composition environments, producing harder Pro benchmarks where seven frontier models drop by up to 38.2 points. It raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score and GLM-5.2 from 44.6 to 56.6 across seven benchmarks.

• ModelQwen3.6-35B-A3B
• ModelGLM-5.2
• ModelClaude Opus 5
• Topicresearch
• Topicagents
• Topicbenchmark