Reliable function calling requires Large Language Models (LLMs) to select appropriate functions and generate arguments consistent with user intent. However, syntactically valid calls may still be semantically incorrect due to inappropriate function selection or mismatched arguments. In this paper, we propose Decision-Model-Guided Function Calling (DM-Call), a framework that combines local schema validation with separate decision-model verification of function selection and argument correctness. For rejected candidates, DM-Call accumulates verification probabilities, binary judgments, and schema errors across attempts. A hybrid feedback representation further combines numerical probabilities with verbal correctness labels to guide iterative refinement without executing external tools. Extensive evaluation across BFCL and API-Bank (first-call) demonstrates DM-Call boosts function-calling accuracy versus direct generation, at modest additional cost.
Figures & tables
Figure 1. Comparison of conventional function-calling approaches and DM-Call . (a) Conventional approaches rely on schema validation or self-review, which may fail to detect semantically incorrect calls. (b) DM-Call separately verifies function selection and argument correctness through probabilistic judgments, accumulating rejected candidates and their hybrid feedback to guide subsequent revisions.
gpt-5-nano
gpt-5.4-mini
claude-sonnet-5.5
gemma-4-31b-it
Method
BFCL ↑
API-Bank †↑
USD ↓
BFCL ↑
API-Bank †↑
USD ↓
BFCL ↑
API-Bank †↑
USD ↓
BFCL ↑
API-Bank †↑
USD ↓
Direct Generation
Direct
64.12
61.32
0.167
81.59
80.13
2.146
88.25
86.97
12.598
86.67
81.41
–
Self-Refine
69.45
58.55
0.762
81.27
79.91
9.591
88.43
87.39
56.184
86.69
81.41
–
Multi-Agent
MAD-FC
63.53
66.03
1.493
81.10
80.98
16.501
85.22
81.62
95.244
84.91
82.48
–
Table 1. Comparison of function-calling methods across four LLM generators on BFCL and API-Bank. We report three-run mean accuracy (%) at r⋆=3 (s.d. 0.00–2.89) and total inference cost (USD). Decision. Generation refers to function-call generation guided by separate decision-model verification. Bold and underline indicate the best and second-best results per generator and metric, respectively; – denotes unpriced local inference. † API-Bank uses first-call exact match without no-call negatives.
Method
Call (%)
Func. (%)
Args. (%)
Direct
64.11±0.31
75.11±0.29
81.27±0.14
(i) DM-Call-Function
73.12±0.46
83.84±0.52
81.78±0.07
(iii) DM-Call-Argument
73.61±0.33
83.73±0.52
82.64±0.38
(iii) DM-Call-Latest
73.05±0.24
83.61±0.22
82.05±0.16
DM-Call (Default)
74.32±0.22
84.80±0.26
82.18±0.13
Table 2. Ablation study of DM-Call on BFCL using gpt-5-nano and Jev ( r=3 ), reporting three-run mean accuracy (%) ± standard deviation for overall function calling, function selection, and argument correctness. Bold and underline indicate the best and second-best results per metric, respectively.
Jev
PPLX
Generator
Numeric
Text
Hybrid
Numeric
Text
Hybrid
gpt-5-nano
74.20
74.14
74.63
73.95
73.68
74.24
gpt-5.4-mini
82.45
82.47
82.54
82.22
82.26
82.20
claude-sonnet-5.5
88.53
88.57
88.56
88.45
88.46
88.46
gemma-4-31b-it
86.93
86.91
86.91
86.89
86.91
86.88
Avg. Score
83.03
83.02
83.16
82.88
82.83
82.95
Table 3. Comparison of Numeric, Text, and Hybrid feedback representations. Results show three-run mean accuracy (%) (s.d. 0.02–0.54), with overall-averages across generators. Bold and underline indicate the highest and second-highest average scores per decision model, respectively.
Figure 2. Effect of repair budget r on DM-Call using gpt-5-nano , BFCL, and Jev . (a) Call accuracy, (b) mean and p95 latency, and (c) latency comparison with baselines.
t
Proposed call ct
pf
pa
Correct
(a) Function. claude-sonnet-5.5 , BFCL: “how to cook sikhae”
0
HNA_WQA_search(keyword="how to make sikhae Korean sweet rice drink recipe")
0.37
0.94
✗
1
cookbook_search_recipe(keyword="sikhae")
0.93
0.95
✓
(b) Argument value. gpt-5.4-mini , BFCL: “Find me the best Italian restaurants in New York City with average customer ratings of more than 4 and accepts credit cards.”
0
restaurant_search(accepts_credit_cards=true, cuisine="Italian", location="New York City, NY", rating=5)
0.94
0.19
✗
1
restaurant_search(…, rating=4)
0.96
0.57
✓
Table 4. Case studies of DM-Call with Jev at r⋆=3 , illustrating function correction, argument refinement, no-call decisions, and multi-step repair. Each step reports prob. ( pf , pa ) and correctness, with – indicating skipped argument verification.
Function calling represents the core capability of agentic large language models (LLMs). Existing research has focused on enhancing LLMs function-calling accuracy through fine-tuning, reinforcement learning (RL), and multi-agent frameworks, particularly for native function-calling LLMs. This work demonstrates that LLMs achieve superior accuracy in function calling in instruction-following contexts (i.e., standard user-assistant interactions) rather than a tool calling context. We introduce Instruction-Followed Function Calling (IFFC), a novel framework that decouples function-calling logic from the primary LLM and delegates it to a dedicated smaller model operating within the instruction-following paradigm. Our method consistently outperforms both native function calling (NFC) and prompt-based function calling (PFC) baselines, with particularly strong gains on reasoning-oriented LLMs. Furthermore, we demonstrate that IFFC maintains robust performance under aggressive quantization, enabling efficient on-device deployment without significant accuracy degradation. This work establishes a new paradigm for reliable, resource-efficient function calling in edge-computing scenarios.
Yalda Taheri, Mohammad Hassan Heydari, Erfan Naaman +1
Department of Engineering Islamic Azad University · Department of Computer Engineering University of Isfahan
We describe the QMUL_AK submission to AISA-ArabicFC 2026. The system adapts the Arabic-English ALLaM-7B-Instruct model to function calling with low-rank adaptation (LoRA). It combines an Arabic schema-aware prompt, dialect and no-call resampling, a token-weighted training objective, and deterministic validation of generated calls against the candidate tool schema. The same output supports Track A function calling and Track B reasoning-augmented calling. Training uses only the official task data. On the development set, the system achieves 0.9982 function accuracy and 0.8920 argument exact match, corresponding to overall A and B scores of 0.9345 and 0.9454. On the blind test set, the system reaches overall A and B scores of 0.9253 and 0.9380, ranking 3rd on both submission tracks and 1st of 21 teams on the Track C dialect-robustness diagnostic with a macro-averaged score of 0.9075 and the smallest dialect gap in the shared task (0.077). The results show that function selection is nearly saturated, while argument extraction remains the main source of error.
Function calling empowers large language models (LLMs) to interface with external tools, yet existing RL-based approaches suffer from misalignment between reasoning processes and tool-call decisions. We propose R2IF, a reasoning-aware RL framework for interpretable function calling, adopting a composite reward integrating format/correctness constraints, Chain-of-Thought Effectiveness Reward (CER), and Specification-Modification-Value (SMV) reward, optimized via GRPO. Experiments on BFCL/ACEBench show R2IF outperforms baselines by up to 34.62% (Llama3.2-3B on BFCL) with positive Average CoT Effectiveness (0.05 for Llama3.2-3B), enhancing both function-calling accuracy and interpretability for reliable tool-augmented LLM deployment.
Aijia Cheng, Kailong Wang, Ling Shi +1
Shanghai Key Laboratory of Trustworthy Computing, East China Normal University, Shanghai, China · Huazhong University of Science and Technology · Nanyang Technological University