Tunable Tool-Call Rates in LLM Agents via Representation Steering

Yuqi Chen; Vincent Siu; Yang Liu; Dawn Song; Chenguang Wang

cs.AIAug 25, 2026

Tunable Tool-Call Rates in LLM Agents via Representation Steering

Authors: Yuqi Chen, Vincent Siu, Yang Liu, Dawn Song, Chenguang Wang

Organizations: Department of Computer Science and Engineering, UC Santa Cruz · Department of Computer Science, UC Berkeley

Abstract

Deciding whether to call a tool is a core competence of an LLM agent, and a costly one to get wrong: needless calls add latency, accrue cost, and may trigger irreversible side effects, while missing calls leave the model confidently wrong on questions it could only answer through tool-calls. Models manage this balance poorly, both over-using and under-using tools. Existing methods such as post-training and prompt engineering are expensive and difficult to modify at inference time. We show that whether an instruction-tuned model calls a tool can be controlled by a single linear direction in its residual stream, extracted without any training from the model's own tool-use preference signal and turned into an inference-time intervention with no prompt change. Adding the direction with strength $α$ moves the call rate monotonically from near $0\%$ to over $90\%$ while keeping calls well-formed. The steering works in both directions: dialing it down suppresses calls, and dialing it up induces new calls that land precisely on the questions the model cannot answer from its own knowledge. We also show that the direction generalizes to unseen tools with strength comparable to each tool's own direction and without favoring any specific tool choice. With live tool execution, a single sweep of the steering traces a cost/accuracy Pareto frontier and nearly doubles open-domain QA accuracy ( $0.29 \! \rightarrow \! 0.56$ ); the same recipe transfers across a diverse range of models spanning dense, MoE, and multimodal architectures, without any training. Our code is publicly available at https://github.com/YuqiChen4188/Steering-Tool-Use-Propensity.

Figures & tables

Figure 1: Bidirectional control of tool use through representation steering. For an uncommon factual question, the unsteered model answers incorrectly without using a tool. Adding the tool-use direction induces a search call and leads to the correct answer. For a math question that the base model can already solve, it calls the calculator unnecessarily, while subtracting the direction removes the needless call and the answer stays correct.

Figure 2: A single steering direction provides tool-general control and improves the cost–accuracy trade-off on Qwen3-4B. Left figure shows a single direction, extracted once in the multi-tool harness, reliably controls the tool-call rate when deployed on search, a calculator, or Python. Right figure shows that on PopQA dataset, sweeping

\alpha

traces a cost/accuracy Pareto frontier.

Figure 3: Cross-model reproducibility of whether-to-call steering. Across five models covering dense, mixture-of-experts (MoE), and multimodal architectures, increasing the steering coefficient

\alpha

consistently raises the tool-call rate from near zero to near one. The distribution of calls among search, calculator, and Python varies across models. An asterisk indicates that reasoning was bypassed.

Figure 4: Cost-accuracy Pareto frontiers under live search across five models from three model families. Each point corresponds to one combination of system prompt and steering coefficient

\alpha

on PopQA. For clarity, we show only the non-dominated configurations for each model, so the number of plotted points differs across models.

Figure 5: Tool-call behavior on PopQA for Qwen3-4B. Left: baseline log-probability of initiating a tool call across entity-popularity deciles. Right: tool-call rates under positive steering for different entity-popularity questions.

Figure 6: Layer $\times$ $\alpha$ steering map ( $\Delta\log p(t^{\star})$ ) on Qwen3-4B. Control is monotone and strongest in a mid-late band and the last layers are large but erratic.

Held-out tool	$v_{\mathrm{multi}}$	$v_{\mathrm{own}}$
translate	25.3	18.4
weather	20.1	18.8
unit converter	26.8	23.1
e-mail	22.4	16.7
SQL	20.2	22.9
stock price	23.9	16.6

Table 1: Suppression strength on six held-out tools. Each entry is the drop in the log-probability of the tool-call token when steering with

\alpha{=}{-}2

in that tool’s single-tool harness,

|\Delta\log P|

in nats averaged over its 100 template queries, so larger means stronger control.

v_{\mathrm{multi}}

is the direction extracted in the three-tool harness, which never saw these tools, and

v_{\mathrm{own}}

is the direction extracted from the tool’s own queries, rescaled to the same norm.

	baseline	steered ( $\alpha{=}{+}2$ )
Total call rate	0.07	0.82
PopQA calls $\to$ search	—	100%
GSM8K calls $\to$ calculator	100%	99%
code calls $\to$ Python	—	90%

Table 2: Increasing the steering strength raises the overall tool-call rate while largely preserving task-appropriate tool selection. A dash marks cells with no tool calls.

Clamp target $c$	Qwen3-4B	Gemma-4-E4B
no intervention	0.05	0.37
$-2\sigma$	0.01	0.01
$-1\sigma$	0.02	0.17
$0\sigma$ (baseline mean)	0.03	0.61
$+1\sigma$	0.06	0.87
$+2\sigma$	0.11	1.00

Table 3: Tool-call rates under projection clamping and directional ablation for an under-user, Qwen3-4B

L22

, and an over-user, Gemma-4-E4B

L35

, evaluated on 150 mixed questions. Clamp targets are standardized relative to each model’s baseline projection distribution, with

0\sigma

denoting its mean. The no-intervention row reports the model’s natural call rate. Directional ablation sets the raw projection onto

\hat{v}

to zero, corresponding to

+2.1\sigma

for Qwen3-4B and

+1.3\sigma

for Gemma-4-E4B.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Model family	Opener
Qwen3	<tool_call>
Gemma-4	<\|tool_call>
gpt-oss	channel-route readout (no single opener)

Appendix

Table 4: Tool-call opener

t^{\star}

per model family.

Tool	Arguments	Description
search	query	Search the web and return the top results. Useful for up-to-date or long-tail factual information.
calculator	expression	Evaluate an arithmetic expression and return the numeric result.
python	code	Execute a short Python snippet and return its stdout.

Appendix

Table 5: Tools of the multi-tool harness.

Tool	Arguments	Description
translate	text , target_language	Translate text between languages.
get_weather	location , when	Get the current or forecast weather for a location.
unit_converter	value , from_unit , to_unit	Convert a quantity between units of measurement.
send_email	to , subject , body	Send an email on the user’s behalf. This action is irreversible.
run_sql	query	Run a read-only SQL query against the company database (tables: customers, orders, products).
stock_price	ticker	Look up the latest stock price for a ticker symbol.

Appendix

Table 6: Held-out tools, listed under the function identifiers the model sees. The main paper refers to the same six tools, in this order, by task name: translation, weather, unit conversion, e-mail, SQL, and stock lookup.

Figure 7: Projection interventions. Left: clamping the projection onto

\hat{v}

moves the call rate monotonically for both an under-user (Qwen3-4B) and an over-user (Gemma-4-E4B), and the ablation point (

c{=}0

, stars) lies on each model’s clamp curve. Because the baseline mean projection is negative, raw zero sits above the baseline operating point on the axis. Right: ablation erases knowledge selectivity, collapsing the tail/head call-rate gap from

+0.26

+0.02

LLM Tool Use

Large language model tool use studies how models select, call, and coordinate external tools to extend their reasoning and take actions.

See all 289 papers

LLM Agents Already Know When to Call Tools -- Even Without Reasoning
May 10, 2026Chung-En Sun, Linbo Liu, Ge Yan +2Tool-Augmented Language Model Agents AI Agent Benchmarks
University of California, San Diego
Tool Calling is Linearly Readable and Steerable in Language Models
May 8, 2026Zekun Wu, Ze Wang, Seonglae Cho +4AI Agent Reliability Tool-Augmented Language Model Agents
University College London · Holistic AI · Imperial College London
How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression
Oct 7, 2026Xijie Gong, Tingxu Han, Jiahao Zhang +5LLM Tool Use LLM Decision-Making
Mohamed bin Zayed University of Artificial Intelligence · University of Electronic Science and Technology of China · Nanjing University +3

May 10, 2026J/K move · Enter open · S save

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

Chung-En Sun, Linbo Liu, Ge Yan, Zimo Wang+1

University of California, San Diego

Tool-Augmented Language Model Agents AI Agent Benchmarks LLM Tool Use Tool Selection Tool Selection for Language Model Agents LLM Agent Efficiency

Open paper AskPDF arXiv

Tool-augmented LLM agents tend to call tools indiscriminately, even when the model can answer directly. Each unnecessary call wastes API fees and latency, yet no existing benchmark systematically studies when a tool call is actually needed. We propose When2Tool, a benchmark of 18 environments (15 single-hop, 3 multi-hop) spanning three categories of tool necessity -- computational scale, knowledge boundaries, and execution reliability -- each with controlled difficulty levels that create a clear decision boundary between tool-necessary and tool-unnecessary tasks. We evaluate two families of training-free baselines: Prompt-only (varying the prompt to discourage unnecessary calls) and Reason-then-Act (requiring the model to reason about tool necessity before acting). Both provide limited control: Prompt-only suppresses necessary calls alongside unnecessary ones, and Reason-then-Act still incurs a disproportionate accuracy cost on hard tasks. To understand why these baselines fail, we probe the models' hidden states and find that tool necessity is linearly decodable from the pre-generation representation with AUROC 0.89--0.96 across six models, substantially exceeding the model's own verbalized reasoning. This reveals that models already know when tools are needed, but fail to act on this knowledge during generation. Building on this finding, we propose Probe&Prefill, which uses a lightweight linear probe to read the hidden-state signal and prefills the model's response with a steering sentence. Across all models tested, Probe&Prefill reduces tool calls by 48% with only 1.7% accuracy loss, while the best baseline at comparable accuracy only reduces 6% of tool calls, or achieves a similar tool call reduction but incurs a 5

\times

higher accuracy loss. Our code is available at https://github.com/Trustworthy-ML-Lab/when2tool

Abstract

LLM Tool Use

Explore similar work

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

Tool Calling is Linearly Readable and Steerable in Language Models

How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression