🤖 AI

Controlling Tool Use with Heading-Specific Activation Steering

This paper demonstrates that while steering vectors extracted from heading-anchors can effectively control tool invocation in large language models, geometric analysis reveals that these vectors lack the clean linear structure typical of parametric concepts, exhibiting instead diffuse, bimodal alignments and distinct signatures for different tool types that reflect the non-parametric nature of external tools.

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

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a large language model (LLM) as a brilliant, knowledgeable chef who has memorized millions of recipes (parametric knowledge). However, this chef also has a drawer full of external tools: a search engine, a calculator, and a phone to call a customer for clarification.

The problem is that this chef sometimes gets a bit anxious. Even when they know the answer from memory, they might panic and grab the calculator or call the customer unnecessarily. This wastes time, costs money, and slows things down.

This paper investigates a way to "steer" the chef's brain to stop these unnecessary actions without retraining the chef from scratch. Here is the breakdown of their discovery:

1. The "Traffic Light" in the Brain

The researchers found that when the model is about to decide, "Should I use a tool?" or "Should I just think about this?", there is a specific moment in its internal processing (right before it writes a heading like ### Code or ### Search) where a decision is made.

They discovered a "steering vector"—think of it as a magic nudge or a traffic light signal—that lives inside the model's brain at that exact moment.

The Nudge: If you add this specific signal to the model's brain, it acts like a "Stop" sign, effectively telling the model, "Don't use the tool; just think."
The Reverse Nudge: If you remove this signal (by projecting the brain's activity in the opposite direction), it acts like a "Go" sign, making the model use tools even more than usual.

This is surprising because tools aren't "hard-coded" into the model's memory like facts are; they exist only in the conversation context. Yet, the model still has a stable internal signal for when it thinks it needs them.

2. The "Goldilocks" Effect (It Works Best for Math)

The researchers tested this on three types of tasks:

Math: These are problems the chef can usually solve with their own brain (memory).
Time: These require looking up current data (like "What time is it in Tokyo right now?").
Intention: These require asking the user for missing details (like "What color do you want?").

The Result:

In Math, the "Stop" signal worked beautifully. The chef stopped grabbing the calculator for simple addition and just solved it in their head. The answers were still correct, but the tool usage dropped dramatically.
In Time and Intention, the "Stop" signal was too strong. When they told the chef "Don't use tools," the chef stopped looking up the time or asking for details, and the answers became wrong.

The Lesson: The steering works great when the model shouldn't be using tools, but it's dangerous to use it blindly when the model needs tools to get the right answer.

3. The "Messy Map" (Geometry is Weird)

Usually, when scientists find a "concept" inside a computer brain (like "honesty" or "refusal"), it looks like a clean, straight line on a map. If you move along that line, the behavior changes predictably.

However, the researchers found that the "Tool Use" signal is messy.

The Analogy: Imagine trying to find the "North" direction on a map. For "Honesty," North is a straight arrow. For "Tool Use," North looks like a cloud of fog with two different clusters. The model's brain doesn't have a single, clean "I need a tool" button. Instead, it has a scattered, fuzzy collection of signals that only look like a decision when you zoom out.
Different Tools, Different Maps: The internal signal for "Search" is quite different from the signal for "Ask the User." They don't overlap much. It's like the chef has a completely different set of nerves for calling a customer than for using a calculator.

4. Why Does It Work If the Map is Messy?

This is the biggest mystery the paper leaves open. Even though the internal map of "Tool Use" is fuzzy and scattered, the "magic nudge" still works perfectly to stop the model from using tools.

The authors suggest it's because the model makes its decision at a very specific "bottleneck"—the moment it has to write the heading (like ### Code). Even if the path to that decision is messy, hitting that specific bottleneck with the right nudge is enough to change the outcome.

Summary

The paper proves that you can control whether an AI uses external tools by tweaking a specific signal in its brain at the moment it decides to act.

Good news: You can stop it from wasting time on tools it doesn't need (like in math).
Bad news: If you turn the signal off too much, it stops using tools it actually needs (like checking the time).
The Twist: The internal mechanism for this decision is messy and scattered, not a clean straight line, which makes it a unique and tricky part of how AI thinks.

Technical Summary: Controlling Tool Use with Heading-Specific Activation Steering

Problem Statement

Tool-augmented Large Language Models (LLMs) extend capabilities beyond parametric knowledge by integrating external tools (e.g., web search, code execution). However, a persistent failure mode is tool overuse, where models invoke tools unnecessarily even when internal reasoning suffices. This leads to increased latency, cost, and potential error propagation.

Existing mitigation strategies typically require expensive retraining or operate at the output interface, failing to address the latent decision-making process within the model's internal activations. A fundamental challenge arises because, unlike concepts such as sentiment or factuality which are encoded in model weights during training, tools exist entirely in context at inference time. Consequently, standard assumptions from the linear representation hypothesis and representation engineering—where concepts have stable parametric grounding—do not straightforwardly apply. The paper investigates whether tool-use decisions possess a stable internal representation that can be extracted and manipulated despite this non-parametric nature.

Methodology

The authors propose a lightweight, inference-time intervention method called Heading-Specific Activation Steering. The approach relies on structured generation trajectories where the model emits intermediate reasoning using explicit section headers (Heading-Anchors) such as ### Reasoning, ### Code, ### Search, and ### AskUser.

1. Steering Vector Construction

The method constructs tool-specific steering vectors by analyzing hidden states at the token positions corresponding to these Heading-Anchors.

Data Collection: The authors collect labeled multi-turn trajectories containing specific tool headings.
State Extraction: Hidden states are extracted from a specific layer $l^*$ at the token corresponding to the leading ### anchor marker for both the ### Reasoning state and the target tool-heading state (e.g., ### Code).
Delta Calculation: For each trajectory, a steering delta is computed as the difference between the average hidden state of the reasoning anchors and the average hidden state of the tool anchors:
$\Delta^{(l)}_{k,i} = \bar{h}^{(l)}_{R,i} - \bar{h}^{(l)}_{k,i}$
Vector Aggregation: These deltas are averaged across multiple trajectories to form the final steering vector $v^{(l)}_k$ . This sample-level averaging prevents bias toward trajectories with a higher frequency of anchor occurrences.

2. Intervention Mechanisms

The extracted vectors are applied at the Heading-Anchor formation steps during inference using two distinct mechanisms:

Activation Addition (Suppression): The steering vector is added to the hidden state to shift it away from the tool-invocation representation:
$\tilde{h}^{(l^*)}_t = h^{(l^*)}_t + \alpha v^{(l^*)}_k$
(where $\alpha=1$ ).
Orthogonalization (Amplification): The hidden state is projected to be orthogonal to the steering vector, removing the suppressive component to encourage tool use:
$\tilde{h}^{(l^*)}_t = h^{(l^*)}_t - \text{proj}_{v^{(l^*)}_k}(h^{(l^*)}_t)$

Key Results

The study evaluates five open-source instruction-tuned models (Mistral-7B, Llama-3.1-8B, Mistral-Nemo, Mistral-Small, Llama-3.1-70B) across three domains: Math (parametric reasoning sufficient), Time (requires up-to-date info), and Intention (requires user clarification).

1. Bidirectional Causal Control

Suppression: Activation addition significantly reduces tool invocation rates. In the Math domain, tool usage was suppressed to near-zero levels (e.g., from 3.90 to 0.56 calls per prompt for Mistral-7B) with only modest drops in task accuracy, indicating that many tool calls in this domain were redundant.
Amplification: Orthogonalization consistently increased tool usage above the baseline across all models and domains. However, this artificial increase did not improve task performance and often degraded accuracy, suggesting the base model already maintains a partial internal balance and that removing the suppression vector leads to overuse.
Domain Specificity: Suppression was most effective in Math but degraded in Time and Intention domains, where external retrieval or clarification is genuinely required. In these domains, suppressing tool use led to significant performance drops, confirming the intervention respects the necessity of tools when parametric knowledge is insufficient.

2. Geometric Irregularity

Contrary to the "clean linear structure" observed in parametrically grounded concepts (like refusal or truthfulness), the geometric analysis of tool-use vectors revealed:

Diffuse Bimodal Alignment: Tool-invocation steps exhibited a diffuse, bimodal cosine similarity distribution with respect to the suppression vector, rather than the consistent negative alignment predicted by linear encoding accounts.
Low Cross-Tool Overlap: Different tool types (Code, Search, AskUser) recruited largely distinct internal feature dimensions with low Jaccard overlap (e.g., ~0.04–0.06 for three-way overlap).
Decoupling of Decodability and Control: In the Llama-3.1-70B model, a supervised linear probe could perfectly separate ### Code from ### Reasoning states (AUROC 1.000), yet the direction identified by the probe was far less effective at causally suppressing tool use than the mean-difference vector. This suggests that the direction most causally effective for control is geometrically distinct from the direction that is most linearly decodable.

3. Robustness

The intervention was found to be sensitive to the schema but not solely to the literal surface string. When tested with cross-format transfer (Markdown to JSON) and heading renaming, the suppression effect persisted even when the literal word "Code" was absent, though the effect was asymmetric and schema-dependent.

Significance and Claims

The paper claims to be among the first to apply activation steering to a behavior that is explicitly non-parametric by construction. The primary significance lies in demonstrating that:

Causal Control Exists: Despite the lack of parametric encoding for tools in model weights, heading-anchored steering vectors can exert bidirectional causal control over tool-invocation behavior.
Non-Parametric Geometry: The geometric properties of tool-use steering vectors differ fundamentally from those of parametrically grounded concepts. The observed geometric irregularity (diffuse alignment, low overlap) is hypothesized to be a predicted consequence of the non-parametric nature of tools, distinguishing them from concepts like sentiment or refusal.
Practical Utility: The method offers a parameter-free, inference-time mechanism to mitigate tool overuse, particularly in domains where parametric reasoning is sufficient, without the cost of retraining.

The authors conclude that while the causal effectiveness of the steering is clear, the relationship between this effectiveness and the observed geometric irregularity remains an open question, potentially linked to the multi-turn dependency structure of tool-augmented trajectories.