acceptodds
Under review as a conference paper at ICLR 2027

Tool Calling is Linearly Readable and Steerable in Language Models

Abstract

Language-model agents can take real actions by calling tools, so choosing the wrong tool can cause errors that are difficult to undo. Most evaluations only observe the tool choice after the model generates a call. We read this choice from the model before generation, and we steer it. For each tool, we average the model's hidden states from a few example requests to obtain a tool vector. Comparing a new request with these tool vectors predicts which tool it needs. The difference between two tool vectors gives a steering direction that can move the model toward a chosen tool without retraining. Across eight instruction-tuned models from 4B to 27B parameters, steering changes the generated call to a chosen target tool in 56–78% of held-out tool pairs, depending on the model. We also observe steering on real APIs from -bench and ToolBench, though less reliably. At the final layer, steering could work simply by raising the score of the target tool name. To test this, we compare the steering direction with a direction that only raises that score, layer by layer. In the middle layers, on the three models we examine in depth, the steering direction switches more calls than the score-raising direction. So the tool vectors capture part of the tool choice before the final layer. The same tool vectors also help identify likely tool-selection errors before generation. Calls are more likely to be wrong when a request is similarly close to two tool vectors. Across four models, this signal achieves an AUROC of 0.61–0.78 and outperforms a first-token confidence baseline on three of them. Together, these results suggest that the model's hidden state provides a way to read, steer, and check tool choice before a call is made.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, activation steering, tool calling, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_k1y5OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Tool Vector, an inference-time activation steering method to predict and intervene an LLM’s tool choice before tool-call generation. It constructs tool prototype vectors by averaging final-layer hidden states from a small set of demonstrations, then steers activations using the difference between two tool vectors without fine-tuning. The method achieves 56–78% steering success on clean tool pairs but degrades substantially on ToolBench and T-Bench. A central open question is whether the effect arises from semantic representation shifts or is merely equivalent to logit bias.

Critical strengths
  • Novel perspective: it predicts and intervenes tool choice before decoding, different from most post-hoc correction methods.
  • Lightweight: training-free, only vector addition at inference time with low overhead.
  • Multi-model evaluation: tested across a range of instruction-tuned models from 4B to 27B.
Critical weaknesses
  • Weak mechanistic evidence. Head-to-head comparison against logit bias is missing, so semantic steering cannot be distinguished from token-level bias.
  • Flawed evaluation: only measuring whether the model switches to the target tool, without checking whether the resulting tool call matches the query, enabling invalid forced steering.
  • Missing ablations and validation: the core clustering assumption is untested; ablations over demo count, steering strength α\alpha and token positions are absent; failure modes on real benchmarks are not analyzed.
Questions for authors

Same as weakness

Reviewer_Ni7wOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how tool-choice information is represented in language models before a tool call is generated. The authors construct a vector for each tool by averaging hidden states from example requests. The authors use these vectors to read tool labels, steer generation toward a chosen tool, and rank potential selection errors. Through experiments across models and layers, the authors find that tool-related information is readable and that intervening on it can change the generated tool name. The authors also identify limitations involving shared name prefixes and similar tools.

Critical strengths
  1. The paper connects readable representations with causal interventions. It goes beyond showing that tool labels can be decoded, changing the tool-vector direction also changes generated tool names. This provides evidence that tool-related representations can influence model behavior.
  2. The layer-wise controls clarify what the intervention changes. The comparison with output-weight directions and the removal of output-aligned components provide evidence that middle-layer effects cannot be fully explained by a direct change to the final tool-name score. The shared-prefix experiments also identify a concrete boundary of last-layer steering.
Critical weaknesses
  1. The representation may partly reflect request semantics rather than a distinct tool-selection decision. Readout is often evaluated against the dataset’s tool label, and a request-only sentence encoder performs competitively in some settings. The paper includes relevant controls, but the interpretation of the decoded information remains somewhat uncertain.
  2. Generalization to realistic tool-use settings remains limited. The main real-API experiments reuse test requests when constructing tool vectors; the authors provide additional held-out checks for three models, but these use few request wordings. It remains unclear how robust the mechanism is across varied tool schemas and multi-turn contexts.
Questions for authors
  1. How much of the tool readout remains after controlling more directly for the semantic similarity between requests and tool descriptions?
  2. Can the steering effect be reproduced across more diverse held-out requests, tool schemas, and multi-turn contexts?
  3. Can intervention at earlier layers or generation positions distinguish tools that share the same initial token?
Reviewer_a69oOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks inside an LLM to see which tool it's about to call, before any execution happens. This is done by inspecting hidden states. The paper shows you can steer the model to a different tool than the one favored by the model. There's some ability to predict incorrect calls.

Critical strengths

The methodology seems sound, especially in the area testing whether steering is doing something real.

Critical weaknesses

The main thing for me is the utility of the technology. Is it helping to make models more accurate or efficient or better in some way. I'm not getting enough of that.

Questions for authors

none

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.