AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning

2025-05-15 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

AutoTool is a novel framework that equips large language models (LLMs) with dynamic tool-selection capabilities for agentic reasoning, overcoming the limitation of fixed tool inventories. It leverages a 200k dataset with explicit tool-selection rationales across over 1,000 tools and 100+ tasks spanning mathematics, science, code generation, and multimodal reasoning. The framework employs a dual-phase optimization pipeline: first, supervised and RL-based trajectory stabilization for coherent reasoning, and second, KL-regularized Plackett–Luce ranking to refine consistent multi-step tool selection. Training Qwen3-8B and Qwen2.5-VL-7B with AutoTool yielded average performance gains of 6.4% in math & science, 4.5% in search-based QA, 7.7% in code generation, and 6.9% in multimodal understanding across ten diverse benchmarks, demonstrating enhanced generalization to unseen tools.

Key takeaway

For AI Engineers developing LLM agents that need to adapt to evolving tool environments, you should consider integrating dynamic tool-selection mechanisms like AutoTool. This approach, which explicitly optimizes tool choice through Plackett–Luce ranking, significantly improves generalization and performance across diverse tasks. By moving beyond fixed tool inventories, your agents can achieve more robust and adaptable reasoning, especially when encountering new or complex toolsets during inference.

Key insights

LLM agents can achieve dynamic tool selection and superior generalization by optimizing tool preferences via Plackett–Luce ranking.

Principles

• Tool selection should be a per-step decision-making process.
• Anchor tool selection in embedding space for generalization.
• Explicitly optimize tool selection for robust performance.

Method

AutoTool uses a dual-phase pipeline: Phase I stabilizes reasoning with SFT and RL, then Phase II refines tool selection using KL-regularized Plackett–Luce ranking on masked trajectory segments.

In practice

• Curate datasets with explicit tool-selection rationales.
• Use embedding-based tool selection for evolving toolsets.
• Apply Plackett–Luce ranking to refine multi-step tool choices.

Topics

• LLM Agents
• Dynamic Tool Selection
• Reinforcement Learning
• Plackett-Luce Ranking
• Multimodal Reasoning
• Code Generation
• Qwen Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• ParaTool: Shifting Tool Representations from Context to Parameters — ParaTool shifts LLM tool representations from context to parameters, reducing overhead and hallucination risks.
• Researchers let AI Agents Optimize LLM Reasoning and Cut Tokens by 70% — AI agents can autonomously optimize LLM reasoning strategies, significantly reducing token usage while maintaining accuracy.
• DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use — Scaling diversity in agentic task synthesis through evidence-first execution significantly improves LLM generalization for tool use.
• Do Multimodal Agents Really Benefit from Tool Use? A Systematic Study of Capability Gains — Tool-augmented multimodal agents may learn tool-calling patterns more reliably than actual tool-contributed capabilities.
• ToolChoiceConfusion: Causal Minimal Tool Filtering for Reliable LLM Agents — Reliable LLM agent tool selection hinges on causal necessity, not just semantic relevance or executability.
• NEW AI In-Context Reinforcement Learning for Agentic Tools (ICRL) — ICRL enables LLMs and VLMs to learn tool use in-context via reinforcement learning, bypassing fine-tuning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.