Tool calls varied widely
The central finding is a wide spread in how often the models called a tool. Across the reported steering sweep, the tool-call rate ranged monotonically from near 0% to more than 90%, while calls remained well-formed. The proposed method uses a single linear direction added to a model’s residual stream during inference. It is training-free and does not change the prompt.
The pattern was not simply a one-way rise in every setting. Lower and higher steering were associated with opposite call-use patterns, and higher steering was associated with new calls on questions the model could not answer from its own knowledge. The method concerns the decision of whether to call a tool, not the quality of the tool’s execution.
The models did not start alike
The cross-model evaluation covered five language models spanning dense, mixture-of-experts and multimodal architectures. Baseline call rates ranged from 0.07 to 0.83. At alpha = -2, every model was at 0.00. At alpha = +2, four of the five models reached 1.00, while Qwen3-4B was at 0.82.
The call-rate measurements used sampled generation at a temperature of 0.7, top-p of 0.8, at most 384 new tokens and a fixed seed for the sweeps. The reported rates belong to that sampling setup.
At alpha = +2, one routing check recorded an overall call rate of 0.82, compared with 0.07 at baseline. The task-to-tool pattern in the same check remained largely appropriate: PopQA questions routed to search reached 100%; GSM8K questions routed to the calculator were at 100% at baseline and 99% at the steered setting; and code questions routed to Python reached 90%.
The signal traveled across tools
Transfer was tested on six held-out tools. The multi-tool direction was stronger than each tool-specific direction for five of the six, and it was within 12% of the tool-specific direction for SQL. The comparison was a test of whether the reported control extended across the tools included in that evaluation.
The knowledge test looked at whether extra calls were linked to questions outside the model’s own knowledge. The popularity-decile study drew 40 questions per decile. A separate projection comparison used 100 rare-tail questions and 100 head questions, terms for the less and more popular ends of that sample. In the over-user’s single-tool search setting, baseline call rates were 0.95 for the tail and 0.69 for the head, a gap of +0.26. After the reported ablation, the figures were 0.99 and 0.97, and the gap was +0.02.
Accuracy came with search cost
The live-search PopQA evaluation paired answer accuracy with searches per question. With zero searches, the reported model accuracies ranged from 0.18 to 0.34. The frontiers peaked at 0.44 to 0.52 when models made 0.75 to 1.2 searches per question, a reported gain of 1.5 to 2.5 times. The numbers describe different points on an accuracy-and-search frontier, rather than an accuracy score detached from search use.
In a separate open-domain question-answering evaluation with live tool execution, reported accuracy was given as a near doubling from 0.29 to 0.56. Taken together with the live-search result, the study reports a trade-off: the higher accuracy points were paired with more searches per question.
The boundary of the result
The authors draw a boundary around the method’s reach. It addresses whether a tool should be called, not how well the tool is executed. For reasoning-first models, the reasoning span must be bypassed before the call decision becomes readable. Strong positive steering may also corrupt tool-call formatting.
The manuscript is an arXiv preprint, version 1, dated 25 Aug 2026. Its abstract states that the code is publicly available. The reported evidence therefore centers on call frequency, the tested task-to-tool mappings, transfer to held-out tools and live-search evaluations, with execution quality left as a separate question.
Paper data and sources
Original title: Tunable Tool-Call Rates in LLM Agents via Representation Steering
Authors: Yuqi Chen, Vincent Siu, Yang Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text