A State-Aware Dual-Tool Architecture for Improved Tool Use in Multi-Turn LLM Dialogue
Open accessThe DOI matches its Crossref record; 29 of the 29 references it cites were found in a registry or an index; 14 of the 19 sources its claims cite were checked for support.
Abstract
This study investigates whether the dominant single-tool paradigm is sufficient for reliable tool use in multi-turn dialogues in which data are provided incrementally and constrained by conditional validation rules. To address the limitations of this approach, we propose a state-aware dual-tool architecture that separates data preparation and validation from final action execution, and we complement it with Token-Optimized Notation Language (TONL) as a compact representation of tool-facing state. The approach was evaluated in a controlled Django-based test platform for an insurance quotation task. Two configurations were compared across four language models (GPT-4o, GPT-5.4, Qwen3.5-122B, Nemotron-3-Super-120B) under four scripted scenarios, with 80 runs per scenario, model, and configuration (2560 total runs). Scenario-level success was defined as correct tool-use behavior over a full scripted dialogue replay, including no premature execution, no fabrication of missing values, and no payload inconsistency with the final validated state. Under this definition, the single-tool baseline achieved 74.8% success, whereas the proposed dual-tool architecture achieved 99.4% success across all evaluated runs. The strongest gain was observed for Qwen3.5-122B in the pressure scenario, where success increased from 5% (4/80) to 100% (80/80). In addition, TONL reduced the size of a representative tool-state payload by 34.0%. The results indicate that explicit completeness feedback and deterministic readiness-based execution gating can substantially improve the reliability of tool-augmented LLM systems while reducing the recurring token cost of repeated tool interactions.
Overview
CoChat Summary
TL;DR
The paper proposes a state-aware dual-tool architecture that separates incremental data preparation/validation from final tool execution, with deterministic gating via a `ready_to_execute` signal and TONL-encoded tool outputs. Across four LLMs and 2,560 scripted multi-turn runs, it reports raising tool-use success from 74.8% to 99.4%.
Key findings
Dual-tool improved aggregate success from 74.8% (958/1280) to 99.4% (1272/1280), an absolute gain of 24.5 percentage points; a two-proportion z-test gave p < 0.001 and Cohen’s h = 0.85.
- The largest gain was Qwen3.5-122B in the pressure scenario: 5% (4/80) single-tool to 100% (80/80) dual-tool. Single-tool pressure failures were GPT-4o 36/80, Qwen3.5-122B 76/80, and Nemotron-3-Super-120B 68/80; GPT-5.4 had 0/80 pressure failures but over-confirmed elsewhere.
- Dual-tool produced only 8 recorded errors across 1280 runs: GPT-5.4 100%, GPT-4o and Qwen3.5-122B 99.7%, Nemotron-3-Super-120B 98.1%. Single-tool overall success was GPT-4o 79.4%, GPT-5.4 79.4%, Qwen3.5-122B 67.8%, Nemotron-3-Super-120B 72.8%.
- TONL reduced a representative tool-state payload from 144 tokens to 95 tokens, a 49-token (34.0%) reduction.
Methods
Controlled Django-based web test platform replaying scripted multi-turn dialogues for an insurance quotation task, with four models in tool-calling mode: GPT-4o, GPT-5.4, Qwen3.5-122B, and Nemotron-3-Super-120B.
- Four fixed scenarios (`scn_memory`, `scn_pressure`, `scn_corrections`, `scn_rules`) were each run 80 times per model and configuration, yielding 320 runs per model and 1280 runs per architecture (2560 total).
- Primary metric was scenario-level success over a full scripted replay, counting errors for failure to invoke the required tool, premature execution, fabricated missing values, or payload inconsistent with the final validated state; aggregate comparison used a two-proportion z-test.
Limitations
Stated in the paper: evaluation was in a controlled, scripted environment; the task was limited to a structured insurance quotation use case; architecture effects were not isolated from prompt effects because the two conditions used different system prompts; and no direct end-to-end latency measurements were provided.
- Also stated: residual dual-tool errors were associated with rare cases where the model did not invoke the prepare-tool early enough to preserve first relevant facts in session state.
- Also stated: no formal significance tests were conducted separately per model or scenario, and no multiple-comparison correction was applied.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 6 Not valid · 3 Not checked · 133 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Moreover, it complicates deterministic management of the tool execution process [9].”
What the sources say
- Partly supportedcites [9] Sun, 2025, Evaluating the reliability of large language models for clinical data extraction in bladder cancer prognosis, Sci. Rep. sourceExperimental — accuracy not yet measured
The Introduction says that LLMs’ probabilistic nature can produce inconsistent outcomes, even with identical prompts and contexts. That supports a concern about reproducibility, but the paper does not address shifting normalization and validation onto an LLM or the deterministic management of tool execution. (LLM, conf=0.77)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“Despite substantial progress, the tool use paradigm exhibits significant limitations in multi-turn dialogue scenarios in which required information is provided incrementally and in diverse linguistic forms [15].”
What the sources say
- Could not verifycites [15] Kennington, C., Lison, P., and Schlangen, D. (2025). Prior Lessons of Incremental Dialogue and Robot Action Management for the Age of Language Models. arXiv. sourceExperimental — accuracy not yet measured
Our copy of Kennington et al. (2025) never takes up a 'tool use paradigm' (function/tool calling) or its limitations; searches for 'tool use', 'function calling', 'API calls' and 'multi-turn' returned only passages on incremental dialogue management in spoken/HRI settings. The nearest material is the Introduction (p. 1) statement that LLMs 'still suffer from a number of limitations when it comes to incrementality' because they are 'trained to act upon complete, sentence-level input' and are strictly monotonic, plus the §3.1.1 note that slot-filling dialogue state is 'restricted to a fixed list of predefined slots' — neither is about tool use across turns, and nothing addresses information arriving 'in diverse linguistic forms'. So the claim's specific point is not covered by our copy (neither supported nor contradicted). (LLM, conf=0.62)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“shifting responsibility for data extraction, normalization, and validation directly onto the language model reduces control over the consistency and completeness of the collected information.”
What the sources say
- Partly supportedcites [9] Sun, 2025, Evaluating the reliability of large language models for clinical data extraction in bladder cancer prognosis, Sci. Rep. sourceExperimental — accuracy not yet measured
The Introduction says LLM variability can lead to inconsistent outcomes, and the Discussion’s Future directions section calls for external validation and quality assurance. The paper does not establish that assigning extraction, normalization, and validation directly to an LLM reduces control over the completeness of collected information; it does not evaluate that delegation as a workflow. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Within structured tool invocation approaches, the language model implicitly performs additional tasks such as information extraction and named entity recognition (NER). This process involves mapping fragments of unstructured user input onto the fields of a structured tool input schema [8].”
What the sources say
- Partly supportedcites [8] Xu, D., Chen, W., Peng, W., Zhang, C., Xu, T., Zhao, X., Wu, X., Zheng, Y., Wang, Y., and Chen, E. (2024). Large Language Models for Generative Information Extraction: A Survey. Front. Comput. Sci., 18. sourceExperimental — accuracy not yet measured
Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Furthermore, information extraction performed implicitly by LLMs may exhibit instability across dialogue turns, leading to inconsistent or incomplete tool input data [9,18].”
What the sources say
- Partly supportedcites [9] Sun, 2025, Evaluating the reliability of large language models for clinical data extraction in bladder cancer prognosis, Sci. Rep. sourceExperimental — accuracy not yet measured
The Information extraction consistency section reports variation in clinical-variable extraction across repeated runs, and the Discussion’s “LLM extraction error propagation” section reports that extraction errors affected downstream survival predictions. The paper does not examine instability across dialogue turns or inputs to tools; its downstream inputs were clinical variables supplied to prediction models. Page numbers were not available in our copy. (LLM, conf=0.84)
- Partly supportedcites [18] Ji, 2023, Survey of Hallucination in Natural Language Generation, ACM Comput. Surv. sourceExperimental — accuracy not yet measured
Sections 8.2 and 8.4 describe inconsistent dialogue responses and difficulty retaining dialogue history across turns. Section 8.3 discusses missing slots in generated task-oriented dialogue. These passages support the general concern about inconsistency or incomplete information, but do not establish that implicit LLM information extraction produces inconsistent or incomplete tool input data. (LLM, conf=0.78)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
56 more claims
Create a free account to see all 150 claims and to verify the 4 references still open.
- It may state two separate claims as one. 71
- Probably not a substantial claim. 35
- Borderline: one of the tests for a substantial claim is unclear. 2
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3.1. Problem Formulation and Conceptual Framework
“Once full completeness is reached, the language model is permitted to invoke the execute-tool exactly once, and execution is triggered at the first dialogue turn in which the completeness condition is satisfied.”
§ 3.4. Tool-Calling Flow and LLMService Orchestration
“The model is explicitly instructed to call prepare\_calculate\_offer whenever any new schema-relevant data appear in the dialogue, and to call execute\_calculate\_offer only after the prepare-tool returns ready\_to\_execute = true and the user confirms the intent to execute.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3.4. Tool-Calling Flow and LLMService Orchestration
“| 3 | The VIN is N7D490L2EHDDYV26A, and the fuel type is petrol. | prepare_calculate_offer({vin:'N7D490L2EHDDYV26A', fuel_type:'petrol'}) | missing = [], ready = true |”
§ 3.8. Conditional Field Requirements and the Completeness Rule Layer
“- tax\_id (taxpayer identification number) is required only if crn\_id (company registration number) is not provided, and vice versa; exactly one identifier must be supplied per quotation request - vin is required only for vehicles older than two years - engine\_power\_kw is required only for gasoline vehicles”
Citations
Cites
Survey of Hallucination in Natural Language Generation
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023) · ACM Computing Surveys · doi:10.1145/3571730 · Cited by 5,025
Self-Refine: Iterative Refinement with Self-Feedback
Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al. (2023) · arXiv · doi:10.48550/arxiv.2303.17651 · Cited by 2,314
Large language models are zero-shot reasoners
Kojima, T., Gu, S. (., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1613 · Cited by 2,013
A survey on large language model based autonomous agents
Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al. (2024) · Frontiers of Computer Science · doi:10.1007/s11704-024-40231-1 · Cited by 1,941
Lost in the Middle: How Language Models Use Long Contexts
Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00638 · Cited by 1,612
Emergent Abilities of Large Language Models
Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al. (2022) · arXiv · doi:10.48550/arxiv.2206.07682 · Cited by 1,264
Toolformer: Language Models Can Teach Themselves to Use Tools
Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023) · arXiv · doi:10.48550/arxiv.2302.04761 · Cited by 1,150
Reflexion: Language Agents with Verbal Reinforcement Learning
Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023) · arXiv · doi:10.48550/arxiv.2303.11366 · Cited by 1,087
MultiWOZ - A Large-Scale Multi-Domain Wizard-of-Oz Dataset for Task-Oriented Dialogue Modelling
Budzianowski, P., Wen, T.-H., Tseng, B.-H., Casanueva, I., Ultes, S., Ramadan, O., & Gašić, M. (2018) · Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d18-1547 · Cited by 1,028
Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?
Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., & Zettlemoyer, L. (2022) · Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2022.emnlp-main.759 · Cited by 755
Research works printed but not linked: 1
- Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., and Awadallah, A.H. (2024, January 7–9). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Framework. Proceedings of the Conference on Language Modeling (COLM 2024), Philadelphia, PA, USA.
Cited by
Shi, Y., Liu, Y., Du, G., Liu, H., & Feng, S. (2026) · IEEE Access · doi:10.1109/access.2026.3736562
Metadata
- Authors
- Pałgan, B., Wawryszczuk, M., Chwał, J., Tkacz, E., & Dzik, R.
- Affiliations
- Akademia Śląska (PL)
- Corresponding author
- Wawryszczuk, Marcin
- Venue
- Applied Sciences
- Publisher
- MDPI AG
- Published
- 2026-04-29
- Volume
- 16
- Issue
- 9
- Pages
- 4329
- ISSN
- 2076-3417 (ISSN-L)
- OpenAlex
- W7160130809
- Cited by
- 1
- References
- 11
- Citations by year
- 2026: 1
- FWCI
- 3.46
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Applied Sciences (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Ministry of Science and Higher Education
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- multi-turn dialogue, data validation, token cost reduction, large language models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: A State-Aware Dual-Tool Architecture for Improved Tool Use in Multi-Turn LLM Dialogue. https://app.cochat.ai/papers/10.3390/app16094329
- The paper
- Pałgan, B., Wawryszczuk, M., Chwał, J., Tkacz, E., & Dzik, R. (2026). A State-Aware Dual-Tool Architecture for Improved Tool Use in Multi-Turn LLM Dialogue. Applied Sciences. https://doi.org/10.3390/app16094329
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.