Skip to content

AutoTool: Efficient Tool Selection for Large Language Model Agents

🕒 Published (v1): 2025-11-18 16:41 UTC · Source: Arxiv · Venue: AAAI 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AutoTool is a graph-based framework that reduces LLM inference cost in tool-using agents by exploiting "tool usage inertia"—the empirically observed tendency of tool invocations to follow predictable sequential patterns. It constructs a Tool Inertia Graph (TIG) from historical trajectories and uses it to bypass LLM calls for predictable tool selections and parameter filling. Across three benchmarks it cuts LLM calls by 15–25% and token consumption by 10–40% with negligible task performance loss.

Problem

ReAct-style agents invoke the LLM at every step to decide which tool to call next, creating high inference latency and cost. Existing efficiency work either requires fine-tuning or focuses on retrieval/search strategies that still rely on LLM calls for each decision. No prior tuning-free method explicitly exploits sequential structure in tool-call patterns to offload decisions from the LLM.

Method

AutoTool builds a Tool Inertia Graph (TIG): a directed, dynamically updated graph where nodes are tools (with embedded parameter subgraphs) and edges encode (1) tool-sequence transition frequencies and (2) parameter dependency flows across tool boundaries. Edge weights are updated online—reinforced on successful inertial calls, decremented on failure.

Before each agent step, AutoTool attempts an "inertial invocation" in two stages: 1. Inertia Sensing: searches the TIG for candidate successor tools matching the last k tool calls; scores each with a CIPS score combining a frequency score (from TIG edge weights) and a contextual relevance score (SimCSE embeddings of agent "intuition" vs. tool description) via CIPS = (1−α)·Scorefreq + α·Scorectx. If max CIPS > threshold θ_inertial, proceed. 2. Hierarchical Parameter Filling: populates tool arguments via (a) dependency backtracking through TIG parameter edges, (b) environmental state matching, (c) heuristic filling. Only if all parameters are resolved is the inertial call executed; otherwise falls back to full LLM inference.

Constraints cap inertial calls at ≤30% of total operations and prohibit consecutive inertial calls to preserve task quality.

Key Contributions

  • Empirical identification and Markov-chain quantification of tool usage inertia: conditional entropy drops from 3.50 bits (0-order) to 2.52 bits (1st-order) to 1.93 bits (2nd-order) on 6,014 ScienceWorld tool invocations (G²-test, p < .001).
  • Design of the Tool Inertia Graph (TIG): hierarchical, online-updated graph capturing both tool-sequence patterns and parameter data flows.
  • A CIPS scoring mechanism blending historical frequency with semantic contextual relevance for tool candidate ranking.
  • A hierarchical parameter filling algorithm that backfills tool inputs from TIG edges without LLM calls.
  • Demonstration of model-agnostic efficiency gains across Llama4-Scout-17B, Llama-3.3-70B, Qwen2.5-72B, and DeepSeekV3.

Results

Evaluated on AlfWorld, ScienceWorld, and ToolQuery-Academic with ReAct and Reflexion baselines (Llama4-Scout-17B, cold-start graph):

  • ReAct+AutoTool vs. ReAct:
  • AlfWorld: 1.18× fewer LLM calls, 1.60×/2.87× reduction in input/output tokens; PR improves 0.394→0.531
  • ScienceWorld: 1.31× fewer LLM calls, 1.30×/1.41× token reduction; PR 0.716→0.708 (−1.1%)

  • ToolQuery-Academic: 1.20× fewer LLM calls, 1.15×/0.92× token reduction; PR 0.901→0.895

  • Reflexion+AutoTool vs. Reflexion:

  • AlfWorld: 1.29× fewer LLM calls; PR 0.481→0.453 (slight drop)
  • ScienceWorld: 1.28× fewer LLM calls; PR 0.730→0.712

  • ToolQuery-Academic: 1.26× fewer LLM calls; PR 0.917→0.923

  • AutoTool overhead: semantic relevance computation accounts for only 2.7% ± 1.5% of total task time; non-semantic graph operations are on the order of seconds vs. thousands of seconds of LLM inference.

Limitations

  • The 30% inertia-call cap and prohibition on consecutive inertial calls are manually tuned constraints that limit the maximum achievable speedup.
  • Cold-start performance depends on accumulating sufficient historical trajectories; early task runs have sparse TIG edges.
  • Hyperparameter sensitivity: θ_inertial and α interact; the paper reports that the 30% cap is the operative constraint in most configurations, somewhat neutralizing the effect of threshold tuning.
  • Evaluated only on text-based and structured-API benchmarks (AlfWorld, ScienceWorld, ToolQuery-Academic); generalization to real-world multi-modal or open-ended tool ecosystems is unverified.
  • Slight progress-rate degradation observed in some settings (e.g., Reflexion+AutoTool on AlfWorld: 0.481→0.453).
  • Detailed limitations are deferred to Appendix G (not provided in the text).

Relevance to Agentic AI / LLM Agents

AutoTool directly targets one of the core cost bottlenecks in production LLM agents—the per-step LLM inference for tool selection in ReAct-style loops—making it highly relevant to anyone designing or deploying tool-augmented agents at scale. The "tool usage inertia" observation, rigorously supported with Markov entropy analysis, provides a principled basis for offloading routine decisions from the LLM to lightweight statistical mechanisms, a direction complementary to work on workflow memory (Agent Workflow Memory) and tool retrieval (ToolNet, AnyTool). The plug-in nature of AutoTool—applicable on top of ReAct or Reflexion without retraining—reflects a broader trend toward modular efficiency layers in agentic architectures.