RESEARCH / LANGUAGE MODEL AGENTSA VISUAL GUIDE · ≈8–10 MIN
A framework to guide better tool decisions

To Call
or Not to Call

Assessing tool-use decisions through necessity, utility, and affordability.

QuestionCall atool?Answer directlyUse a toolthen answerNoYesThe tool-use decision reflects the model’s agency.

Qinyuan Wu Seungeon Lee Soumi Das Mahsa Amani Arijit Nag Krishna Gummadi Abhilasha Ravichander Muhammad Bilal Zafar

Max Planck Institute for Software Systems · Ruhr University Bochum · UAR RC Trust

A framework for evaluating the necessity, utility, and affordability of language-model tool calls—examining intermediate tool-calling decisions beyond final task outcomes.

More tool calls do not guarantee better performance

For GPT-OSS-120B on Entity, Self-Decision scores 0.72, compared with 0.76 for Always Tool and 0.81 for the oracle. These final scores reveal a performance gap, but do not explain what happened: did the model miss useful calls, call when no help was needed, or use a tool that made the answer worse? To understand the gap, we need to examine the quality of individual tool-calling decisions—whether each call was needed and whether it actually helped.

GPT-OSS-120B · Entity task · score from 0 to 1
No Tool
0.61
Always Tool
0.76
Self-Decision
0.72
Oracle
0.81
00.51
Figure A. Policy comparison. Values reproduced from Table 1. The Entity score is factuality, not exact-answer accuracy. The oracle selects using both realized outcomes; it is a hindsight reference, not a deployable policy. §3
Compare all seven models
Entity task · score (tool-call rate)
ModelNo ToolAlways ToolSelf-DecisionOracle
GPT-OSS-120B0.61 (0%)0.76 (100%)0.72 (30%)0.81 (61%)
Qwen3-30B-A3B0.70 (0%)0.81 (100%)0.80 (56%)0.88 (51%)
Qwen-3-30B-IT0.68 (0%)0.82 (100%)0.82 (95%)0.87 (60%)
Llama3.2-3B-IT0.58 (0%)0.70 (100%)0.70 (100%)0.83 (57%)
Mistral3.1-24B-IT0.70 (0%)0.83 (100%)0.70 (0%)0.88 (51%)
Gemma3-27B-IT0.60 (0%)0.80 (100%)0.80 (92%)0.85 (59%)
GPT-5.50.85 (0%)0.86 (100%)0.85 (32%)0.94 (39%)
Table A. Entity policy results. Reproduced from Table 1. Scores are normalized factuality scores, not exact-answer accuracy. “Oracle” is the paper’s hindsight OPTIMAL policy. Models use different fixed harnesses, so these rows should not be read as a controlled ranking of intrinsic model capability.

Definitions and evaluation protocol

Drawing on rational choice theory (Peterson, 2017; Tversky & Kahneman, 1974), we investigate tool-calling decisions along three dimensions: necessity—does the model need external help? Utility—does the available tool improve the answer? And affordability—does the expected benefit justify the cost, and where should limited calls be allocated? These dimensions let us compare what a model should do with what it perceives and chooses, then investigate how to improve its decisions.

No Tool Reference outcome

Task xModelAnswer
sNT(x)

Always Tool Reference outcome

Task xModel + tool
Required call
Answer
sAT(x)

Model generates tool input; tool response informs the answer.

Self-Decision Observed behavior

Task xModel decides
Call the tool?
No · π(x) = 0Answer without tool
Yes · π(x) = 1Call tool → use response → answer

Decision context: tool description, instructions, and budget information when provided.

The tool-use decision reflects the model’s agency.

NORMATIVE · OUTCOME-BASED REFERENCE

What should be called?

True necessity
Is the No Tool answer below acceptable quality?
True utility
Does Always Tool improve the score?
Δ*(x) = sAT(x) − sNT(x)
True affordability
Allocate up to K calls to the largest positive gains.
Compare reference and behaviorIdentify unnecessary calls, missed benefits, and allocation gaps.
DESCRIPTIVE · MODEL PERCEPTIONS & ACTIONS

What does the model choose?

Perceived need
Does the model report needing help?
Perceived utility
Does the model choose to invoke the tool?
Perceived affordability
Which instances receive the first K chosen calls?
Framework overview. Adapted from Figure 1 and §3. The reference compares realized No Tool and Always Tool outcomes; its optimal allocation uses hindsight. Perceived need is elicited separately, while calling behavior serves as a proxy for perceived utility. The later controller section develops the prescriptive lens: how these distinctions guide improved allocation.

True need indicates that No Tool performance falls below an acceptable-quality threshold. True utility is the difference between Always Tool and No Tool scores. Need does not imply positive utility: assistance may fail to improve the answer.

Δ*(x) = sAT(x) − sNT(x)

sAT and sNT are normalized scores under Always Tool and No Tool. Positive, zero, and negative Δ* denote improvement, no change, and deterioration.

WHEN WEB SEARCH HURTS · GPT-OSS-120B

Query: “In a paragraph, could you tell me what you know about Sky Blue?”

Without web search

Describes the color sky blue

The model gives a description of the color. The paper’s evaluation assigns the answer a factuality score of 1.00.

Factuality score1.00
With web search

Shifts to football clubs

Search returns results about Coventry City and Sky Blue FC. The model’s answer adopts this interpretation, and its factuality score falls to 0.58.

Factuality score0.58
Figure B. Web search reduces factuality in a reported example. The tool-assisted answer scores 0.42 lower than the no tool answer: Δ* = 0.58 − 1.00 = −0.42. Responses and scores are reported in Appendix D, pp. 22–23. This is one observed negative-utility case; the shift in interpretation alone does not establish why the factuality score declined.
How the paper measures the decision

No Tool establishes no tool performance. Always Tool supplies the comparison outcome. Self-Decision reveals the model’s choices.

True need means no tool quality falls below a threshold. Perceived need comes from asking whether help is needed; perceived utility is inferred from calling behavior. The normative, descriptive, and prescriptive lenses ask what should happen, what does happen, and how to improve it.

The paper measures open-ended need using low/mid no tool score bands and exact-answer need using correctness. Calling behavior is an operational proxy for perceived utility, not a calibrated probability of benefit.

Observed score transitions and decision errors

Web search · noisy external evidence

A stochastic, paid tool that retrieves potentially noisy evidence. Successful use requires retrieving relevant information and interpreting it to answer the question.

Entity · InVivoQuery · BFCL
Calculator · deterministic computation

A local tool with no per-call fee that returns an unambiguous result for a valid expression. The model must still identify the right computation, formulate valid input, and integrate the result.

GSM-Hard · multiplication · large-digit squaring

Together, these tools cover two complementary settings: acquiring external knowledge through uncertain evidence, and delegating precise computation to a deterministic tool. They test different sources of difficulty—retrieval and evidence use versus input formulation and arithmetic—so the framework is evaluated beyond a single type of tool. §4, Models and Tools

Web-search samples: Entity n = 500; InVivoQuery n = 500; BFCL n = 314 atomic questions. Calculator samples: GSM-Hard n = 1,319; multiplication n = 1,000; large-digit squaring n = 1,000. Open-ended tasks use model-based assessment; calculator tasks use exact match.

Each cell below counts Entity instances by factuality band without and with search. Rose denotes a higher band; blue denotes a lower band.

Performance WITH search →

Higher bandSame bandLower band
Figure C. Factuality-band transitions. GPT-OSS-120B, Entity (n = 500). Redrawn from Figure 2. Low: 0–0.1; mid: 0.1–0.9; high: 0.9–1, as labeled in the source.
44%of low/mid no tool answers improve to a higher band (154/350).
32%of high no tool answers drop to a lower band (48/150).

Calculated from Figure C: (57 + 15 + 82)/350 = 44%; (3 + 45)/150 = 32%. The diagonal contains 290/500 = 58% of instances. These are band-transition statistics for one model–task pairing, not universal tool-use rates.

Models misjudge both their need for help and the benefit of calling

Perceived need is a self-report; perceived utility is inferred from calling. Neither reliably matches the outcome-based reference.

GPT-OSS-120B · Entity · 500 instances per matrix

Does reported need match true need?

True ↓ / Reported →NeedNo need
Need208142
No need11139

40.6% of true need goes unrecognized (142/350).

Do calls match positive utility?

True utility ↓ / Action →CallNo call
Positive8272
Other70276

46.8% of positive-utility cases are missed (72/154); 46.1% of calls fall outside positive utility (70/152).

Figure D. Perception–outcome misalignment. Counts from Figure 5; percentages calculated from those counts. Peach cells mark mismatches. “Other” combines neutral and negative utility under the figure’s categorization; it does not mean every call is harmful.
Different models, different failure patterns
Entity · percentages with explicit denominators · Figure 5
ModelTrue need unrecognizedCalls outside positive utilityPositive-utility cases missed
GPT-OSS-120B40.6%142 / 350 needing help46.1%70 / 152 calls46.8%72 / 154 positive cases
Llama3.2-3B-IT89.9%347 / 386 needing help54.9%274 / 499 calls0%0 / 225 positive cases
GPT-5.546.1%88 / 191 needing help72.3%115 / 159 calls61.1%69 / 113 positive cases

Llama3.2 misses little benefit by calling almost everywhere, despite underreporting need. GPT-5.5 also reports need on 83/309 no-need cases. These are within-model diagnostics under each model’s fixed harness.

Stated need and action can also diverge. Llama3.2 calls on all 454 cases where it reports no need; GPT-OSS calls on 3/281 such cases (Figure 3). Agreement with a self-report is therefore distinct from making a beneficial call.

Affordability under a call-count constraint

With a maximum of K calls and uniform per-call cost, the oracle selects up to K largest positive realized gains. The descriptive comparison retains the first K instances on which the model elects to call.

Try a call budget · illustrative gains, not experimental data
Question A+0.40Selected
Question B+0.05Not selected
Question C−0.20Not selected
Question D+0.30Selected
Question E0.00Not selected

2 calls used · total gain +0.70

Figure E. Conceptual allocation demonstration. The five gains are synthetic, not measurements from the paper. The interaction illustrates the policy in §3. It uses at most K calls and leaves non-beneficial calls unused. Total gain is a sum across instances, not an average task score.

For Entity, the affordability experiments vary permitted call fractions from 1% to 100% under a nominal $10,000 budget and uniform call cost B/K. The paper reports an oracle–model utility gap (Figure 14), excess uncapped calls under explicit budgets (Figure 15), imperfect prioritization (Figure 16), and failures under implicit budgets (Figure 17). These are experimental budget conditions, not measured deployment prices.

From uncertain model decisions to deterministic tool-call control

The necessity/utility/affordability framework guides controller design: predict need, allocate calls under a budget, and evaluate their realized utility. A lightweight Latent Need Estimator (LNE) predicts need from the fixed model’s hidden states, outperforming self-reports in most tested conditions.

NecessityPredict need
AffordabilityAllocate at most K calls
UtilityEvaluate realized gains
Figure F. Framework-guided controller design. Need-prediction evidence: Figure 6; allocation method: §4.3.
Gemma3-27B-IT · GSM-Hard · natural setting

Higher final performance

0.56 → 0.59

Exact-match task score · Self-Decision → LNE

Fewer tool calls

94% → 29%

A 65-percentage-point reduction in the actual tool-call rate.

Higher performance with fewer calls. Reported in Table 2. This natural setting imposes no external call cap. LNE improves the rounded task score by 0.03 while selecting fewer calls.
Gemma3-27B-IT · Entity · 40% tool-call budget

Higher score at the same budget

0.67 → 0.71
Self
0.67
LNE
0.71

Factuality score; bars use a 0–1 scale. Both policies use the reported 40% call rate.

Utility over No Tool

+0.07 → +0.11

Self-Decision → LNE. The score improves by 0.04 at the same call rate, illustrating better allocation of the available calls.

Figure G. Allocation at a fixed budget. Gemma3-27B-IT, Entity; both policies use a 40% call rate. Scores from Table 2; No Tool baseline (0.60) from Table 1. Gains are differences of rounded mean scores.

The controllers demonstrate allocation benefits in the tested settings, not universal improvement. Need is a proxy for benefit; utility prediction remains harder, and cross-task transfer is a separate empirical question.

Interpretation and limitations

  • Deployment: LNE requires hidden-state access and supervised labels. Reported controller results concern open models; some model–task entries are unavailable because need labels are degenerate.
  • Open problem: utility estimation remains unresolved in the reported experiments. Interpretive boundary: offline ranking of available instances does not itself establish an online policy for unknown future arrivals.

The results motivate separate evaluation of need, realized utility, and budget allocation; they do not establish a universally optimal tool-calling policy.

Reference

[1] Qinyuan Wu, Seungeon Lee, Soumi Das, Mahsa Amani, Arijit Nag, Krishna Gummadi, Abhilasha Ravichander, and Muhammad Bilal Zafar. 2026. To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling. arXiv preprint arXiv:2605.00737v3, version 3, 6 August 2026.

Versioned manuscript (PDF) · Authors’ code and datasets