ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
ToolSpace training lifted Qwen3.6-35B-A3B’s average score from 23.0 to 41.0 across seven long-horizon tool-use benchmarks.
The framework builds environments from dependencies observed in real tool-execution traces, then generates tasks that must replay successfully from their initial state. It uses deterministic rules to identify those dependencies rather than LLM judgment. Reinforcement learning pushed Qwen’s average to 43.2, while GLM-5.2 rose from 44.6 to 56.6. The pipeline also produced harder versions of two benchmarks, where seven frontier models lost up to 38.2 points.
HF Daily Papers' note
The framework builds environments from dependencies observed in real tool-execution traces, then generates tasks that must replay successfully from their initial state. It uses deterministic rules to identify those dependencies rather than LLM judgment. Reinforcement learning pushed Qwen’s average to 43.2, while GLM-5.2 rose from 44.6 to 56.6. The pipeline also produced harder versions of two benchmarks, where seven frontier models lost up to 38.2 points.
HF Daily Papers' note
score 5