The ICML 2026-accepted paper AutoTool (arXiv:2512.13278) proposes a "dynamic tool selection" training framework, directly hitting the pain point of "the tool set must be fixed" in current Agentic RL. The method is in two phases: Phase I uses SFT + RL to stabilize the trajectory of "inserting tool calls in long CoT"; Phase II uses KL-constrained Plackett-Luce Ranking for multi-step tool-selection refinement, extending the "stabilize then refine" post-training paradigm to the tool dimension. In dataset scale, they built a 200K explicitly annotated tool-call trajectory dataset, with each step labeled with which tool was selected and why, covering 1346 tools and 120 task categories (math, science, search QA, code, multimodal all included). Results on 10 benchmarks: Qwen3-8B averages +6.4% (math/science), +4.5% (search QA), +7.7% (code), +6.9% (multimodal), fully open-sourced. Most worth elaborating on is the "unseen-tool generalization" experiment: at training time the model is exposed to only 460 tools, yet at inference it can stably perform in a 1346-tool pool (including 886 unseen tools) — this pushes tool-calling from "closed-set selection" to "open-vocabulary retrieval", with major implications for the Agent ecosystem. Another note worth flagging: Qwen2.5-VL-7B walking the same training pipeline also gets +6.9% multimodal gain, showing that "which visual tool to pick" is also a learnable thing. Code, model, and data are all open-sourced on GitHub (Gen-Verse/Open-AgentRL), and any team building an Agent training framework should read it through.