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

Key points

• Tool calling, invoking external tools on demand, is central to agentic LLMs, yet the mechanism that decides whether a model calls a tool or responds directly remains poorly understood.1
• Agentic prompts are long and heavily scaffolded, combining role instructions, tool schemas, format templates, and the user's request across hundreds of tokens, creating a noisy, highly entangled context in which no single controllable variable for mechanistic analysis is obvious.1
• We trace the decision to a vector, $μΔ$, that is both causally necessary and sufficient and generalizes beyond the discovery prompts to native multi-turn $τ^2$-Bench trajectories and verb-free requests.1
• Behavioral ablations show that the scaffold establishes a tool-call prior; Transcoder decomposition then reveals that analysis verbs suppress this prior through features signaling that tool use is unnecessary, whereas execution verbs largely leave it intact.1

Sources (1)

• [1]How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 08:01 AM

  To obtain such a variable, we propose a method that converts complex agentic prompts into minimal contrastive pairs in which a single request verb determines the tool-call decision: replacing an execution-verb (e.g., write) with an analysis-verb (e.g., discuss) reliably flips the decision, suggesting it is mediated by a compact internal state.

  Tool calling, invoking external tools on demand, is central to agentic LLMs, yet the mechanism that decides whether a model calls a tool or responds directly remains poorly understood.

Extractive summary: sentences quoted from the sources.