Research map / Large language models / Agents, code and math

LLM Agents Already Know When to Call Tools - Even Without Reasoning

NeurIPS 2026 · Large language models › Agents, code and math · approach: Software engineering agents · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces When2Tool, an 18-environment benchmark for tool-call necessity, shows tool necessity is linearly decodable from pre-generation hidden states (AUROC 0.89-0.96), and proposes Probe&Prefill, which steers tool-call decisions by prefilling a sentence based on a linear probe.

Problem

LLM agents call tools indiscriminately even when they can answer directly, wasting API cost and latency; existing benchmarks do not study when a tool call is actually necessary, and prompt or reasoning-based controls are coarse and costly.

Approach

Tool necessity is linearly decodable from the last-input-token hidden state; a lightweight linear probe trained on binary necessity labels drives a prefilled steering sentence that controls whether the model calls a tool, giving a tunable accuracy-efficiency tradeoff without reasoning tokens.

Key results

• Probe&Prefill reduces tool calls by 48% with 1.7% accuracy loss across models, while best baseline at comparable accuracy reduces only 6% of tool calls.
• Linear probe on pre-generation hidden states achieves AUROC 0.89-0.96 across six models (Qwen3 and Llama families).
• On Search-o1 agentic benchmark, Probe&Prefill reduces API calls by 20-56% without accuracy degradation.

Surprising findings

• Reason-then-Act collapses tool calling on Llama models (e.g., Llama-3.1-8B accuracy drops from 79.5% to 31.2%) because the model narrates tool intent without producing valid invocations.
• The linear probe's signal remains strong (AUROC above 0.9) even for models whose generation-time tool decisions fail completely.

Limitations

• Soft prefill is partially ignored by Llama models due to weak instruction following, requiring hard prefill for control.
• Labels depend on the model's no-tool accuracy, so tool necessity is defined empirically per model rather than intrinsically.

Experimental setup

Datasets: When2Tool

Benchmarks: When2Tool Search-o1

Metrics: accuracy total tool calls AUROC accuracy cost per saved call

Baselines: Prompt-only Reason-then-Act SFT

Builds on: ReAct Reflexion Toolformer Search-o1

Closest papers by meaning

• ContractBench: Can LLM Agents Preserve Observation Contracts? (NeurIPS 2026 · similarity 0.88)
• Are Tools Always Beneficial? Learning to Invoke Tools Adaptively for Dual-Mode Multimodal LLM Reasoning (ICML 2026 · similarity 0.88)
• Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.88)
• ToolACE: Winning the Points of LLM Function Calling (ICLR 2025 · similarity 0.88)
• Agentic Abstention: Do Agents Know When to Stop Instead of Act? (NeurIPS 2026 · similarity 0.87)
• Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning (ICLR 2026 · similarity 0.87)