PaperlayerView exact revision on arXiv

Source-linked AI summary

Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use

Yize Cheng, Chenrui Fan, Mahdi JafariRaviz, Keivan Rezaei, Soheil Feizi

arXiv:2605.14038v2cs.AI

TL;DR

Tool necessity varies with each model’s capabilities, but models often fail to align their tool calls with what they actually need. This paper defines model-adaptive necessity, diagnoses tool use as cognition plus execution, and finds a knowing-doing gap concentrated in the cognition-to-action transition.

  • Problem

    Prior work treats tool necessity as model-agnostic and focuses on obvious cases, leaving model-specific necessity in realistic settings insufficiently characterized.

  • Method

    The paper defines tool necessity from each model’s empirical performance and decomposes tool use into internal cognition and execution stages across four models and two QA domains.

  • Results

    26.5–54.0% of arithmetic queries and 30.8–41.8% of factual QA queries show necessity-action mismatch, with most errors arising during cognition-to-action execution.

  • Takeaways & Limitations

    Reliable tool use requires bridging models’ internal recognition of tool necessity with their actual call-or-no-call actions.

  • Takeaways & Limitations

    The study uses one N and T instantiation and hidden-state probing, limiting applicability to closed-source models such as GPT or Gemini.

Abstract

from arXiv · show

Large language models (LLMs) increasingly act as autonomous agents that must decide when to answer directly vs. when to invoke external tools. Prior work studying adaptive tool use has largely treated tool necessity as a model-agnostic property, annotated by human or LLM judge, and mostly cover cases where the answer is obvious (e.g., fetching the weather vs. paraphrasing text). However, tool necessity in the wild is more nuanced due to the divergence of capability boundaries across models: a problem solvable by a strong model on its own may still require tools for a weaker one. In this work, we introduce a model-adaptive definition of tool-necessity, grounded in each model's empirical performance. Following this definition, we compare the necessity against observed tool-call behavior across four models on arithmetic and factual QA dataset, and find substantial mismatches of 26.5-54.0% and 30.8-41.8%, respectively. To diagnose the failure, we decompose tool use into two stages: an internal cognition stage that reflects whether a model believes a tool is necessary, and an execution stage that determines whether the model actually makes a tool-call action. By probing the LLM hidden states, we find that both signals are often linearly decodable, yet their probe directions become nearly orthogonal in the late-layer, last-token regime that drives the next-token action. By tracing the trajectory of samples in the two-stage process, we further discover that the majority of mismatch is concentrated in the cognition-to-action transition, not in cognition itself. These results reveal a knowing-doing gap in LLM tool-use: improving tool-use reliability requires not only better recognition of when tools are needed, but also better translation of that recognition into action.

1 Introduction

The paper defines tool necessity relative to each model’s empirical capabilities and diagnoses adaptive tool-use failures through cognition and execution stages. Its findings reveal a knowing-doing gap, with most mismatches arising when internal recognition fails to become tool-call action.

  • Model-adaptive necessity: Model-adaptive tool necessity grounds whether external help is needed in each model’s empirical performance rather than static, model-agnostic annotations.This definition characterizes necessity relative to the specific capabilities of the model being evaluated.
  • Cognition and execution: The proposed two-stage decomposition separates cognition, where representations encode tool necessity, from execution, where the model outputs tool-triggering tokens.Representation probing is used to examine whether internal cognition and tool-call execution are present.
  • Knowing-doing gap: Most necessity-action mismatches originate in the cognition-to-action transition rather than cognition itself, producing a knowing-doing gap in tool use.Models often represent awareness of their limitations but fail to translate it into syntactic tool-call execution.
  • Behavioral mismatch: Four LLMs evaluated on arithmetic and factual QA show substantial mismatches between actual tool necessity and observed tool-call behavior, reaching 54.0%.The broader reported mismatch ranges are 26.5-54.0% for arithmetic and 30.8-41.8% for factual QA.

2 Related work

Prior work has expanded LLM capabilities through tool calling and standardized tool ecosystems while evaluating tool-use reliability. Related research also studies LLM self-assessment and cases where internal beliefs diverge from observed actions.

  • Tool calling in LLM agents: Tool calling extends LLM capabilities beyond parametric knowledge by enabling interaction with external resources and broader task coverage.Standardized protocols such as MCP and A2A streamline communication and access within tool ecosystems.
  • Tool calling in LLM agents: Prior studies examine tool-use accuracy, hallucinated calls, and robustness to tool descriptions.
  • Meta-cognition of LLMs and the “knowing-doing gap”: LLM meta-cognition research studies whether models can assess their own capability boundaries through explicit self-assessment or verbalized confidence.
  • Meta-cognition of LLMs and the “knowing-doing gap”: Hidden-state studies show that models’ actions can diverge from internal beliefs, including failing to refuse recognized harmful queries or reasoning on math problems they recognize as unsolvable.

3 Defining model-adaptive tool necessity and two-stage modeling of tool-call

This section defines tool necessity relative to each model’s empirical capability boundary and decomposes tool use into cognition and execution. It further separates end-to-end mismatch into errors in recognizing necessity and translating cognition into an action.

  • Defining model-adaptive tool necessity: Tool necessity is defined adaptively because models have different capability boundaries, replacing static query labels with model-dependent judgments.The approach characterizes necessity through a model’s empirical performance without external tools.
  • Defining model-adaptive tool necessity: Empirical tool necessity captures reliability under uncertainty: occasional unaided correctness may still justify external assistance for consistent performance.This grounds necessity in observed model behavior rather than static annotation.
  • Two-stage modeling of tool-call: Tool use is modeled in two stages: internal cognition of whether a tool is needed, followed by execution of the decision to invoke one.The decomposition distinguishes meta-cognition about capability boundaries from execution ability based on that cognition.
  • End-to-end error diagnosis: End-to-end necessity-action mismatch is decomposed into mismatch between actual necessity and cognition, plus mismatch between cognition and the model’s actual tool-call decision.This decomposition enables separate diagnosis of recognition and action errors.

4 Dataset curation

The dataset spans arithmetic and factual question answering across four model families, with tool necessity defined from each model’s repeated no-tool performance. Using calculator and search tools, the study finds substantial and model-dependent mismatches between necessity and observed tool calls.

  • Datasets and models: The dataset covers arithmetic and factual question answering across Qwen3-8B, Qwen3-4B, Llama-3.1-8B-Instruct, and Llama-3.2-3B-Instruct.Arithmetic queries may use a calculator, while factual queries may use a search API.
  • Necessity definition: Tool necessity is assigned using N = 10 no-tool inferences at T = 0.7: any failure is necessary, while ten correct answers are unnecessary.This procedure exposes model-dependent capability boundaries that a model-agnostic definition would obscure.
  • Mismatch results: 26.5%-54.0% of arithmetic queries and 30.8%-41.8% of TruthfulQA queries show end-to-end mismatch between model-specific necessity and actual tool-call behavior.The mismatch reflects tool-use actions inconsistent with the model’s measured capability.
  • Mismatch results: Arithmetic error patterns are model-dependent: Qwen3-8B overuses tools, whereas Qwen3-4B and both Llama models clearly underuse them.Qwen3-8B has UN-C at 38.2% versus N-NC at 3.5%; N-NC reaches 14.5%, 30.1%, and 39.0% for Qwen3-4B, Llama-3.1-8B-Instruct, and Llama-3.2-3B-Instruct, respectively.

5 From meta-cognition to execution ability: What went wrong?

The model’s hidden states often encode both tool necessity and executed action, but these signals decouple near the final-token readout, producing a knowing–doing gap. Most errors arise when cognition fails to translate into tool-call action rather than when necessity is formed.

  • 5.1 Necessity representation: Necessity is linearly separable across broad mid-to-late-layer regions on Arithmetic, but separability is noticeably weaker and more localized on TruthfulQA.On TruthfulQA, MCC > 0.4 mainly appears near the last tokens in mid-to-late layers of Llama models.
  • 5.2 Action representation: MCC ≥0.4 spans broad regions of nearly every model for executed action on both Arithmetic and TruthfulQA.Action information spans most layers and token positions, indicating that impending tool invocation is strongly decodable from the residual stream.
  • 5.3 Cognition–action alignment: Necessity and action directions partially align at intermediate positions but become nearly orthogonal in late layers at the final query token that drives generation.This collapse occurs even for Qwen3-8B and Qwen3-4B, which show the strongest mid-stream alignment, and also appears for other models and on TruthfulQA.
  • 5.4 Error localization: Stage-two errors overwhelmingly exceed stage-one errors, localizing most end-to-end necessity–action mismatch to the cognition-to-action transition.Figure 6 shows the orange stage-two-error flow as the largest category in all four panels, while the red stage-one-error flow is relatively thin.

6 Conclusion

The paper defines tool necessity in a model-adaptive way grounded in empirical capabilities and reveals a substantial mismatch between actual need and tool invocation. Decomposing tool use into cognition and execution stages exposes a fundamental knowing-doing gap in LLMs.

  • 6 Conclusion: The model-adaptive definition grounds tool-necessity evaluation in each model’s empirical capabilities.This reframes necessity around what a given model can actually solve without external tools.
  • 6 Conclusion: The study finds a substantial mismatch between when models need tools and when they invoke them.This mismatch motivates analyzing tool use beyond whether the final answer is correct.
  • 6 Conclusion: Decomposing tool use into internal cognition and execution stages reveals a fundamental knowing-doing gap in LLMs.Hidden-state representation analysis is used to identify this gap across the two-stage process.

A More details on arithmetic dataset curation

The arithmetic dataset contains 4,000 samples spanning three difficulty-oriented problem groups, from calculator-unnecessary short operations to larger operands and structured expressions. Its families include addition/subtraction chains, parenthesized expressions, and multiplication chains with fixed sampling shares.

  • Dataset organization: 4,000 samples comprise three arithmetic problem groups: easy operations, larger short expressions, and more complex structured expressions.The easy group includes one-step addition/subtraction, short two-step chains, and small modulo problems; larger short problems use multi-digit subtraction, four-digit addition/subtraction, and multiplication.
  • Problem families: 11% of samples are one-digit addition/subtraction chains with 16–39 one-digit terms.Operators are addition or subtraction.
  • Problem families: 10% of samples are small addition/subtraction chains containing 21–27 terms drawn from 1–30.Each chain uses addition or subtraction operators.
  • Problem families: 6% of samples are parenthesized expressions with four two-digit operands in (a+b)×(c−d) form.This family explicitly tests grouped arithmetic structure.
  • Problem families: 5% of samples are multiplication chains with five two-digit operands in a+b×c−d×e form.The expression combines addition, subtraction, and multiplication with operator precedence.

B Explicitly prompting for verbalized belief of tool-necessity

Explicitly prompting models to verbalize tool necessity provides a two-stage self-assessment procedure, but these judgments poorly capture capability-grounded necessity. Although explicit prompting can reduce cognition–execution mismatches, it changes tool-calling behavior on nearly half of samples relative to task-oriented prompting.

  • Two-stage inference: The procedure first elicits a yes/no judgment about whether an external tool is necessary, then asks the model to answer the original request.This separates explicit necessity assessment from subsequent task solving.
  • Necessity assessment: Explicit yes/no judgments poorly capture capability-grounded tool necessity, including negative MCC for Llama-3.2-3B-Instruct and undefined MCC for Llama-3.1-8B-Instruct on TruthfulQA.Llama-3.1-8B-Instruct answered “no” for every TruthfulQA sample, implying that it judged no sample to require a tool.
  • Cognition–execution consistency: Llama-3.1-8B-Instruct achieves a 0 mismatch rate while answering “no” for all samples and consistently refraining from tool calls.The explicit judgment remains part of the context, making subsequent generation more consistent with the earlier commitment.
  • Behavioral shift: Nearly 50% of samples show changed tool-calling behavior under explicit self-assessment relative to direct task-oriented prompting.The changed rate is reported as large across datasets, while practical deployments typically use task-oriented prompts designed to maximize task performance.

C Limitations

The study fixes the model-adaptive tool-necessity definition at N = 10 and T = 0.7, and hidden-state probing limits applicability to closed-source state-of-the-art LLMs.

  • Definition settings: N = 10 and T = 0.7 are the only instantiation tested, leaving mismatch-rate changes under other settings unknown.The paper suggests evaluating different N, T values to assess how the necessity-action mismatch rate may change.
  • Model coverage: Hidden-state probing makes the approach inapplicable to closed-source state-of-the-art LLMs such as GPT or Gemini.The limitation follows from relying on access to model hidden states.
Loading 2605.14038v2…