To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling

1. Problem Statement

Agentic AI architectures enable Large Language Models (LLMs) to extend their capabilities by invoking external tools (e.g., web search, calculators). However, current LLM agents lack principled mechanisms to determine when to call a tool.

• The Core Issue: Tool usage is not universally beneficial. Calls can be redundant (if the model already knows the answer) or harmful (if the tool introduces noise or latency).
• The Gap: Existing evaluations focus on aggregate task accuracy, overlooking the quality of individual tool-calling decisions. There is no framework to distinguish between true necessity/utility and the model's perceived necessity/utility.
• The Challenge: Specifically for web search, models struggle to estimate the "utility gain" of a tool call before making it, often leading to suboptimal decisions that violate cost constraints or fail to improve performance.

2. Methodology: A Principled Framework

The authors propose a framework inspired by Decision Theory and Rational Choice Theory to assess and optimize tool calling. The framework decomposes the decision into three dimensions and analyzes them through three distinct lenses:

A. Three Key Dimensions

1. Necessity: Does the model require external help to solve task $T$? (Can it solve it satisfactorily using only parametric knowledge?)
2. Utility: Will calling function $F$ improve performance on task $T$? (Does the tool provide a positive gain over no tool?)
3. Affordability: Is the utility gain worth the cost (latency, API fees)? (Can the model prioritize high-gain instances under a budget?)

B. Three Analytical Lenses

1. Normative Lens (The "Oracle"): Defines the ideal behavior.
   • True Need: Determined by comparing model performance in a NO TOOL setting against a quality threshold.
   • True Utility: Determined by comparing ALWAYS TOOL performance against NO TOOL.
   • True Affordability: Selecting the top-$K$ instances with the highest marginal utility gain ($\Delta = s_{AT} - s_{NT}$) under a budget.
2. Descriptive Lens (The "Self-Decision"): Observes actual model behavior.
   • Perceived Need: The model's self-reported need (via prompting) without tool access.
   • Perceived Utility: The model's actual decision to call a tool when given the option.
3. Prescriptive Lens (The "Controller"): Proposes a mechanism to bridge the gap between descriptive behavior and normative ideals using latent representations.

C. The Controller Framework (Latent Estimators)

To address the misalignment between perceived and true signals, the authors propose training lightweight Multi-Layer Perceptron (MLP) classifiers on the LLM's hidden states (specifically the final token representation) rather than relying on natural language outputs.

• Latent Need Estimator (LNE): Predicts True Need ($N^\star$) using hidden states from the NO TOOL setting.
• Latent Utility Estimator (LUE): Predicts True Utility ($U^\star$) using hidden states from NO TOOL or ALWAYS TOOL settings.
• Mechanism: These estimators act as external controllers, ranking instances by predicted utility/need and allocating a fixed budget to the top-$K$ instances, bypassing the model's flawed self-decision process.

3. Experimental Setup

• Models: Six open-source LLMs ranging from 3B to 120B parameters (including GPT-OSS, Qwen3, Gemma3, Mistral, Llama3.2), covering both instruction-tuned and reasoning variants.
• Tasks:
  1. Entity Task: Generating descriptive paragraphs about entities (evaluating long-form factuality).
  2. InVivoQuery Task: Answering factual questions derived from real-world user requests.
  3. BFCL Task: Berkeley Function Calling Leaderboard (atomic sub-questions).
• Tools: Web search via Google (SerpAPI), Perplexity, Brave, and Tavily.
• Metrics: Factuality scores (verified via LLM-as-a-judge and human annotation), number of tool calls, and utility gain under varying budget constraints.

4. Key Results & Findings

A. Misalignment of Perception vs. Reality

• Suboptimal Self-Decisions: Models consistently fail to align their tool-calling decisions with true utility.
  • Over-calling: Models often call tools for tasks they can solve internally (negative utility).
  • Under-calling: Models fail to call tools when they genuinely lack knowledge (missed positive utility).
• Self-Consistency vs. Truth: While a model's perceived need and perceived utility are often correlated (the model is internally consistent), these perceptions are misaligned with ground truth. The model's internal "belief" about its knowledge is unreliable.
• Cost Sensitivity: Models struggle to adhere to budget constraints. Even when explicitly told the cost per call and remaining budget, models frequently exceed limits or fail to prioritize high-utility queries effectively (low NDCG scores in budget-constrained scenarios).

B. Performance of Latent Estimators

• Superior Prediction: The Latent Need Estimator (LNE) and Latent Utility Estimator (LUE) significantly outperform the models' explicit self-decisions in predicting true need and utility.
• Improved Task Performance: Using the estimators to guide tool calling (the "Optimized" setup) leads to:
  • Higher factuality scores compared to "Self-Decision."
  • Fewer unnecessary tool calls.
  • Better performance under strict budget constraints, often matching or exceeding the performance of much larger models.
• Small Model Gains: The improvement is most pronounced in smaller models (e.g., Llama3.2-3B), which benefit most from the external guidance of latent signals.

C. The "Oracle" Gap

Even with the estimators, performance does not fully reach the "Oracle" (Optimal) level. The authors note that perfect tool calling requires modeling the tool's behavior itself, which is inherently difficult due to the stochastic nature of web search results.

5. Significance and Contributions

1. Theoretical Framework: Introduces the first principled framework to evaluate tool calling based on Necessity, Utility, and Affordability, moving beyond simple aggregate accuracy metrics.
2. Empirical Insight: Demonstrates that contemporary LLMs are unreliable judges of their own knowledge and the value of external tools, leading to systematic inefficiencies.
3. Practical Solution: Proposes a non-invasive, lightweight control mechanism (Latent Estimators) that improves agent decision-making without fine-tuning the base LLM. This suggests that effective agentic behavior may require explicit external controllers rather than relying solely on the model's native reasoning.
4. Cost-Awareness: Highlights the critical gap in current models regarding cost-benefit analysis, showing that even with explicit budget prompts, models fail to optimize tool usage, necessitating external intervention.

Conclusion

The paper argues that the "call or not to call" decision is a fundamental bottleneck in agentic AI. Current models rely on flawed self-perceptions that lead to redundant or harmful tool usage. By leveraging latent representations to estimate true need and utility, researchers can build controllers that significantly enhance task performance and cost-efficiency, paving the way for more rational and reliable AI agents.