LLM Agents Already Know When to Call Tools

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

Chung-En Sun¹, Linbo Liu², Ge Yan¹, Zimo Wang¹, Tsui-Wei Weng¹
¹University of California, San Diego ²Amazon AWS

Abstract

Tool-augmented LLM agents tend to call tools indiscriminately, even when the model can answer directly. Each unnecessary call wastes API fees and latency. The goal is to reduce these unnecessary tool calls without hurting accuracy on tasks that genuinely need tools. We design a benchmark to study this problem, show that prompt engineering and reasoning both fail to selectively reduce unnecessary calls, and propose a lightweight method that exploits the model's own hidden representations to achieve this goal.

Our contributions:

Benchmark: We design When2Tool, the first benchmark for studying tool-call decisions. It comprises 18 environments (15 single-hop, 3 multi-hop) across three categories of tool necessity with controlled difficulty levels, totaling 1,080 training and 2,700 test tasks.
Failure Analysis: We evaluate Prompt-only and Reason-then-Act baselines, revealing that both provide limited and coarse control over tool-call decisions, with hard tasks paying a disproportionate accuracy cost for each saved call.
Discovery & Mitigation: We probe pre-generation hidden states and find that tool necessity is linearly decodable (AUROC 0.89–0.96). We exploit this signal with Probe&Prefill, a lightweight method (<1ms overhead) that prefills the model's response based on the probe prediction. Probe&Prefill reduces tool calls by 48% with only 1.7% accuracy loss, while the best baselines either reduce only 6% at comparable accuracy (8× less efficient) or suffer 5× more accuracy loss. It also generalizes to real-world agentic search, reducing API calls by 20–56% on Search-o1.

48%

Tool-call reduction

1.7%

Accuracy loss

0.89–0.96

Probe AUROC

<1ms

Overhead

Contribution 1 — Benchmark

When2Tool: A Benchmark for Tool-Call Decisions

Existing benchmarks test whether models can use tools correctly, assuming every task requires a tool. When2Tool tests whether the model knows when a tool is needed: tasks range from trivially solvable without tools to impossible without them.

18 environments across 3 categories of self-assessment:

Category A — "Can I compute this?" (Computational scale)
Category B — "Do I know this?" (Knowledge boundary)
Category C — "Can I execute this reliably?" (Execution tracking)

Each with 3 difficulty levels (easy / medium / hard) creating a clear decision boundary.

Figure 1. Overview of When2Tool.

Contribution 2 — Failure Analysis

The Limits of Prompting and Explicit Reasoning

We test two natural training-free approaches:

Prompt-only — 5 modes spanning the full range of tool-use instructions: Force (F) "tool use is mandatory"; Default (★) no explicit requirements; Necessary (N) "only if necessary"; Sparse (S) "expensive, use sparingly"; No Tool (X) "do not use any tools."
Reason-then-Act — Before making a tool-call decision, the model is instructed to first reason about whether it can solve the task directly or needs a tool, then act on its own assessment. Applied on top of each prompt mode above.

Finding: Prompt engineering reduces tool calls indiscriminately — hard tasks lose more accuracy than easy tasks. Reasoning only partially helps but completely breaks on some models (e.g., Llama).

Figure 2. Accuracy vs. avg tool calls per difficulty for Qwen3-1.7B. ★=Default, F=Force, N=Necessary, S=Sparse, X=No Tool. Neither Prompt-only (gray) nor Reason-then-Act (red) can selectively reduce unnecessary calls without hurting hard tasks. Changing prompt mode provides only few operating points with no way to smoothly control the tradeoff, and many modes collapse to nearly identical behavior.

	Qwen3-4B-Inst.			Qwen3-14B			Llama-3.3-70B
Mode / Difficulty	ΔAcc	ΔTC	ΔAcc/−ΔTC	ΔAcc	ΔTC	ΔAcc/−ΔTC	ΔAcc	ΔTC	ΔAcc/−ΔTC
Prompt-only Easy	−14.5	−0.84	−17.3	−8.8	−0.59	−14.9	+1.6	−0.51	+3.2
Prompt-only Medium	−20.7	−0.86	−24.1	−12.9	−0.53	−24.3	+2.0	−0.41	+4.8
Prompt-only Hard	−20.3	−0.48	−42.4	−27.3	−0.47	−58.4	−0.2	−0.34	−0.5
Reason-then-Act Easy	−14.5	−0.86	−16.9	−4.4	−0.67	−6.6	−4.8	−1.98	−2.4
Reason-then-Act Medium	−22.4	−0.90	−24.8	−10.4	−0.62	−16.8	−18.9	−1.87	−10.1
Reason-then-Act Hard	−13.0	−0.35	−36.6	−9.7	−0.28	−34.7	−63.3	−1.99	−31.7

Table 1. Accuracy cost per saved call (ΔAcc / −ΔTC) when switching from Default (★) to Sparse (S). ΔAcc / −ΔTC measures how much accuracy is lost for each tool call eliminated; more negative means each saved call is more costly. Hard tasks pay a disproportionate price (bolded).

Contribution 3 — Discovery & Mitigation

Probing Analysis: Decoding Implicit Tool Necessity

We probe the model's hidden state at the last input token (before any generation) with a simple linear classifier:

Finding: A linear probe achieves AUROC 0.89–0.96 across all six models — substantially exceeding the model's own verbalized reasoning. Even for Llama models where reasoning completely fails, the probe still extracts a strong signal. The model already knows, it just can't show it.

Model	AUROC	Accuracy	AUROC by difficulty
			Easy	Med	Hard
Qwen3-1.7B	0.894	0.847	0.864	0.831	0.904
Qwen3-4B-Inst.	0.948	0.877	0.933	0.906	0.948
Llama-3.1-8B-Inst.	0.927	0.849	0.892	0.867	0.884
Qwen3-14B	0.957	0.892	0.955	0.907	0.941
Qwen3-32B	0.952	0.885	0.951	0.903	0.939
Llama-3.3-70B-Inst.	0.936	0.872	0.906	0.849	0.956

Table 2. AUROC and accuracy for predicting tool necessity from pre-generation hidden states. All probes achieve high AUROC (0.89–0.96), confirming that tool necessity is consistently encoded in hidden representations—even for Llama models where verbalized reasoning completely fails to express this knowledge during generation.

Contribution 3 — Discovery & Mitigation

Probe&Prefill: Turning Hidden Knowledge into Better Decisions

A lightweight inference-time method that translates the probe signal into better tool-call decisions, requiring no fine-tuning and no reasoning overhead:

Figure 3. Overview of Probe&Prefill.

Step 1: Extract last hidden states

Run a single forward pass over the input tokens and extract hidden states at the last token position. This is the standard prompt-encoding step for KV cache, so extraction adds no additional forward passes.

Step 2: Linear probe prediction

Apply the trained linear probe to all-layer hidden states, producing probability \(p\). Threshold \(\tau\) converts this to a binary decision: if \(p < \tau\), the task is solvable without tools; otherwise a tool call is necessary. Sweeping \(\tau\) provides smooth control over the accuracy–efficiency tradeoff.

Step 3: Prefill before generation

Prepend a steering sentence to the model's response based on the probe decision:

If \(p < \tau\) (tool unnecessary): "I can solve this directly without using a tool."
If \(p \geq \tau\) (tool necessary): "I need to use a tool for this question."

The model continues generating from this prefill.

Finding: Probe&Prefill achieves a strictly better accuracy–tool-call tradeoff than all baselines. By sweeping \(\tau\), it traces a smooth Pareto curve—unlike prompt engineering which offers only a few discrete operating points. At \(\tau\)=0.5, it reduces tool calls by 48% with only 1.7% accuracy loss. The probe selectively skips easy calls while preserving hard ones, yielding 8× better efficiency than the best baseline.

Figure 4. Accuracy vs. total tool calls for Qwen models. Probe&Prefill (green) sweeps threshold \(\tau\) from 0.1 to 0.9, tracing a smooth Pareto curve that strictly dominates both Prompt-only baselines (gray markers) and Reason-then-Act (red). Each gray marker is a fixed prompt mode providing no interpolation; Probe&Prefill enables fine-grained control over the operating point.

Figure 5. Accuracy vs. total tool calls for Llama models. Soft prefill (green) is partially ignored due to weak instruction following; hard prefill (purple) forces the output format and restores full tradeoff control. Reason-then-Act (red) catastrophically collapses tool calling on both Llama models, yet the probe still reads the correct signal from hidden states.

	Easy			Medium			Hard			Overall
Method	ΔAcc	ΔTC	ΔAcc/−ΔTC	ΔAcc	ΔTC	ΔAcc/−ΔTC	ΔAcc	ΔTC	ΔAcc/−ΔTC	ΔAcc	ΔTC	ΔAcc/−ΔTC
Necessary (N)	−0.3	−0.09	−3.5	−0.7	−0.04	−17.9	−1.9	−0.04	−44.4	−1.0	−0.06	−16.8
Necessary + Reason-then-Act	−8.1	−0.95	−8.5	−16.3	−0.82	−19.7	−23.2	−0.70	−32.9	−15.8	−0.82	−19.2
Sparse (S)	−6.3	−0.55	−11.3	−7.9	−0.47	−16.9	−11.1	−0.35	−31.5	−8.4	−0.46	−18.4
Sparse + Reason-then-Act	−9.9	−1.13	−8.8	−19.9	−1.04	−19.1	−29.3	−0.84	−34.7	−19.7	−1.00	−19.6
No Tool (X)	−18.1	−0.72	−25.2	−26.7	−0.72	−36.9	−41.4	−0.69	−60.4	−28.7	−0.71	−40.5
No Tool + Reason-then-Act	−12.4	−1.18	−10.5	−23.5	−1.12	−20.9	−36.6	−0.94	−39.1	−24.2	−1.08	−22.4
Probe&Prefill (Ours)	−1.1	−0.66	−1.6	−3.4	−0.54	−6.2	−0.8	−0.24	−3.4	−1.7	−0.48	−3.6

Table 3. Accuracy cost per saved call (ΔAcc / −ΔTC), averaged across six models (τ=0.5 for Probe&Prefill). All changes relative to Default (★). More negative ΔAcc/−ΔTC means each saved call costs more accuracy. Every baseline shows strongly negative costs especially on hard tasks, indicating indiscriminate reduction. Probe&Prefill achieves the lowest cost across all difficulty levels (−1.6 easy, −3.4 hard), confirming it selectively skips easy calls while preserving hard ones.

BibTeX

@article{sun2026when2tool,
  title={LLM Agents Already Know When to Call Tools -- Even Without Reasoning},
  author={Sun, Chung-En and Liu, Linbo and Yan, Ge and Wang, Zimo and Weng, Tsui-Wei},
  journal={arXiv preprint arXiv:2605.09252},
  year={2026},
  url={https://arxiv.org/abs/2605.09252}
}