SAP: State-Guided Data Synthesis with Argument Provenance for Multi-Turn Tool Use

Zichen Tian; Jinpeng Chen; Cheng Gong; Suiyun Zhang; Rui Liu

cs.AISep 5, 2026

SAP: State-Guided Data Synthesis with Argument Provenance for Multi-Turn Tool Use

Authors: Zichen Tian, Jinpeng Chen, Cheng Gong, Suiyun Zhang, Rui Liu

Organizations: Independent Researcher · Huawei Research

Abstract

High-quality multi-turn tool-use data is essential for training agentic models, yet existing data synthesis methods often underrepresent the argument-level dependencies that are critical to long-horizon tool use. As a result, even when a model selects the correct tool, task execution may still fail because the model fills tool arguments with fabricated, stale, or weakly grounded values. To address this problem, we propose State-Guided Data Synthesis with Argument Provenance (SAP). SAP combines state guidance, tool-argument provenance constraints, and turn-level validation to efficiently construct tool-use trajectories with long-range dependencies and high accuracy. Using data generated by SAP, we build SAP-4B, which is highly competitive even when compared with much larger models across multiple benchmarks. Source code, synthesized data, and trained weights are available at https://github.com/Zichen1024/SAP.

Figures & tables

Figure 1: Argument-dependency statistics on representative open-source multi-turn tool-use training sets and ours. Mean Chain Len. and Max Chain Len. denote the mean and maximum lengths of argument-value propagation chains across turns; Dep. Args (%) denotes the proportion of arguments involved in cross-turn dependencies.

Figure 2: Overview of the SAP pipeline. From left to right, $\mathcal{A}_{\text{FSM}}$ constructs an FSM from tool documentation and sampled initial states, and the pipeline samples a state trajectory from it. $\mathcal{A}_{\text{plan}}$ fills the tagged calls $\mathrm{Tool}(\mathrm{tag})$ with concrete parameters, yielding $\mathrm{Tool}(p)$ , and executes them with $\varepsilon$ , using environment rollback and local retry on failure. Finally, $\mathcal{A}_{\text{msg}}$ synthesizes the messages for the validated trajectory, which is filtered into the final data. Each argument is assigned a provenance source from $\{P_{i},P_{o},P_{c},P_{u},P_{f}\}$ .

Sym.	Type	Semantics
$P_{i}$	initial_state	A field of $c_{0}$ ; since the assistant cannot read $c_{0}$ directly, the value enters turn $k$ either from the user message or from a same-turn read-only call.
$P_{o}$	prev_output	Field $f^{\prime}$ of the $j^{\prime}$ -th call at a prior executed turn $k^{\prime}\!<\!k$ (all of $k^{\prime},j^{\prime},f^{\prime}$ are bound at runtime by $\mathcal{A}_{\text{plan}}$ ; $\mathcal{A}_{\text{FSM}}$ only declares the tag).
$P_{c}$	self_create	New user-side value introduced at this turn; later surfaced in $u_{k}$ .
$P_{u}$	prev_user_msg	Value semantically introduced in $u_{k^{\prime\prime}}$ for some $k^{\prime\prime}\!<\!k$ ; dialogue generated retroactively.
$P_{f}$	fallback	Runtime fallback when a declared source $s_{0}\in\{P_{i},P_{o},P_{u}\}$ cannot be used: no matching field of $c_{0}$ , no matching prior return, or no usable anchor value. The planner recovers the argument from another available $P_{i}$ value or from a new $P_{c}$ value when permitted; the resulting call is executor-validated after binding, and the original source $s_{0}$ is kept in fallback_from .

Table 1: Provenance Tag categories.

P_{i}

through

P_{u}

are declarable by

\mathcal{A}_{\text{FSM}}

at design time;

P_{f}

is produced only by

\mathcal{A}_{\text{plan}}

at runtime. Parametrized forms (

P_{o}(k^{\prime},j^{\prime},f^{\prime})

P_{u}(k^{\prime\prime})

P_{f}(s_{0})

) are used only in the definition of § 3.3 .

		BFCL v4 Multi-Turn					$\tau^{2}$ -bench (pass 1 )
Model	Size	Avg	Base	MFn	MPm	LCtx	Retail	Airline	Avg
Closed-source / large-scale reference (BFCL v4)
GPT-5.2-High	–	48.5	–	–	–	–	81.6	62.5	72.1
Claude Sonnet 4.5	–	61.4	69.0	65.0	52.5	59.0	86.2	70.1	78.2
Gemini-3 Pro	–	60.8	64.5	60.0	54.5	64.0	85.3	72.7	79.0
DeepSeek-V3.2-Exp	685B	37.4	41.5	39.5	33.5	35.0	–	–	–

Table 2: Main results on BFCL v4 Multi-Turn and

\tau^{2}

-bench. MFn, MPm, and LCtx denote MissFunc, MissParam, and LongCtx, respectively; Retail and Airline are the two

\tau^{2}

-bench domains. Row groups separate BFCL v4 (directly comparable) from BFCL v3 (reference). Sources: GPT-5.2 ( OpenAI, 2025 ) , Claude Sonnet 4.5 ( Anthropic, 2025 ) , Gemini-3 Pro ( Google DeepMind, 2025 ) , DeepSeek-V3.2-Exp ( DeepSeek-AI et al., 2025 ) , Qwen3-235B-A22B and backbone ( Yang et al., 2025a ) ; ToolWeave ( Khandelwal et al., 2026 ) , MAGNET ( Yin et al., 2025 ) , MUA-RL ( Zhao et al., 2025 ) ; ToolACE-2 / BitAgent results from ( Zhang et al., 2026b ) , CM2 ( Zhang et al., 2026a ) , TOUCAN ( Xu et al., 2025b ) , AWM ( Wang et al., 2026 ) .

Variant	Avg	Base	MFn	MPm	LCtx
SAP -4B (Full)	30.4	38.0	23.0	24.5	36.0
A1: w/o Prov. Tag	24.0	29.5	19.5	21.5	25.5
A2: w/o $\Psi$	28.8	37.0	21.0	22.5	34.5

Table 3: Ablation on BFCL v4 Multi-Turn at matched data scale. A1 removes per-argument Provenance Tag declarations; A2 removes both rewrite operators

\Psi_{\text{param}},\Psi_{\text{func}}

Figure 3: Case study of SAP ’s error localization and hybrid validation. (a) Turn-level execution validation prevents an incorrect file binding from cascading to a later call. (b) Parameter fallback recovers a valid recipient and is checked by the executor, whereas disabling it causes message delivery to fail.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Marker	Trigger (when processing turn $k$ )	Expression constraint in $u_{k}$
[NEW]	$\mathrm{src}(p)=P_{c}$ , or $(p,v)\in\Delta_{v}[k]$ (anchor for downstream $P_{u}$ )	Explicit literal value: first occurrence in the dialogue.
[FROM_HISTORY]	$\mathrm{src}(p)\in\{P_{o},\,P_{u}\}$ (upstream at turn $<k$ )	Vague reference (“that file”, “the ID you found earlier”); literal must NOT be repeated.
[FROM_CONFIG]	$\mathrm{src}(p)=P_{i}$ , value must be passed verbatim to a tool argument	Explicit literal value (the assistant cannot read $c_{0}$ ): IDs, credentials, token-like fields.
[FALLBACK_from_ $s_{0}$ ]	$\mathrm{src}(p)=P_{f}$ (original source $s_{0}$ in metadata)	Treated as [NEW] : explicit literal value.

Appendix

Table 4: Must-mention markers dispatched by the orchestrator to

\mathcal{A}_{\text{msg}}

for user-message synthesis.

Backbone	Traj	Args	Fb.%	$P_{i}$ %	$P_{o}$ %	$P_{c}$ %
Gemini-3.1 Pro	500	3,962	2.02	42.7	25.5	29.8
Gemini-3 Pro	500	4,348	2.76	36.4	24.0	36.9
DeepSeek-V4-Flash	500	3,253	3.08	28.3	0 9.2	59.4
Qwen3.5-Plus	500	4,365	4.82	27.3	34.4	33.5

Appendix

Table 5: Cross-backbone

P_{f}

study on identical FSM skeletons. Traj is the number of trajectories and Args the total number of filled arguments across them; Fb.% is the runtime fallback rate; the remaining columns show the distribution of declared sources actually consumed by

\mathcal{A}_{\text{plan}}

Variant	Succ	$P_{f}$ %	$P_{o}$ %	$P_{c}$ %	$P_{i}$ %
Full SAP	89%	2.06	29.5	40.9	27.5
FSM+noTag	92%	0.00	0.0	53.8	46.2
noFSM+Tag	50%	15.75	16.2	43.9	24.2
noFSM+noTag	66%	0.00	0.0	54.2	45.8
noRetry	54%	3.45	24.8	42.3	29.4

Appendix

Table 6: Pipeline-level ablation. FSM = structure-first skeleton; Tag = Provenance Tag declarations; Succ = trajectories that pass executor validation;

P_{f}

= argument-level fallback rate; remaining columns show the fraction of arguments assigned to each source type (

P_{u}

counted under

P_{c}

Split	Backbone	SAP-4B	$\Delta$
Non-live	84.36	84.88	+0.52
Live	75.41	75.95	+0.54
Overall avg	79.89	80.42	+0.53

Appendix

Table 7: BFCL v4 single-turn results (accuracy, %). Backbone is Qwen3-4B-Instruct-2507; SAP -4B is the same backbone after SFT on SAP data. Multi-turn-only training neither degrades single-turn performance nor regresses on the noisier Live split, and yields small consistent gains on both subsets.

	Turns	$P_{i}$	$P_{o}$	$P_{c}$ / $P_{f}$	Span
Trajectory A (local)	2	0	0	3	0
Trajectory B (cross-turn)	3	0	2	1	2

Appendix

Table 8: Source distribution and max dependency span for the two case-study trajectories.

Panel of Figure 4 — Figure 4: Trajectory A with only local sources. The three argument slots of compute_exchange_rate are user-supplied literals ( $P_{c}$ ); the preceding zero-argument call contributes no argument slot, and no argument depends on a prior turn ( $L{=}0$ ).

Panel of Figure 5 — Figure 5: Trajectory B with cross-turn $P_{o}$ dependencies: Turn 3’s mention consumes tweet_id from Turn 2’s post_tweet return and mentioned_usernames from Turn 1’s list_all_following return (maximum span $L^{\star}{=}2$ ).

Panel of Figure 6 — Figure 6: Prompt template for $\mathcal{A}_{\text{FSM}}$ (FSM skeleton synthesis).

Panel of Figure 7 — Figure 7: Prompt template for $\mathcal{A}_{\text{plan}}$ (per-call argument filling). The two-track output $(\boldsymbol{\theta}^{\text{exec}},\boldsymbol{\theta}^{\text{prov}})$ corresponds to args_for_exec and args_provenance .

Panel of Figure 8 — Figure 8: Prompt template for $\mathcal{A}_{\text{msg}}$ on user-message synthesis. The must-mention list is pre-rendered by the dispatcher of Appendix A.6 from $\boldsymbol{\theta}^{\text{prov}}$ and $\Delta_{v}$ , with each entry already tagged [NEW] / [FROM_HISTORY] / [FROM_CONFIG] / [FALLBACK from $\ldots$ ] .

Panel of Figure 9 — Figure 9: Prompt template for $\mathcal{A}_{\text{msg}}$ on assistant text responses (turns with $\mathcal{C}_{k}=\varnothing$ ).

Panel of Figure 10 — Figure 10: Hint templates injected by the rewrite operators $\Psi_{\text{param}}$ and $\Psi_{\text{func}}$ . Placeholders are filled at injection time.

Synthetic Data Generation for Language Models

Synthetic data generation for language models creates artificial text, tasks, and interaction traces to train or evaluate models.

See all 293 papers

Data Turnstile: A Scalable Open Framework for Function-Calling Data Generation
Jul 31, 2026Goutham Ramakrishnan, Megha SharmaTool-Augmented Language Model Agents Small Language Models
Amazon AGI Bengaluru, India
State-Grounded Multi-Agent Synthetic Data Generation for Tool-Augmented LLMs
Jun 15, 2026Rahul Khedar, Eshita, Sneha Teja Sree Reddy Thondapu +10Multi-Agent LLM Systems Tool-Augmented Language Model Agents
PayPal AI
MidTool: Mid-training Data Synthesis for Agentic Tool Use
Aug 20, 2026Fengqing Jiang, Yite Wang, Boyi Liu +5LLM Tool Use LLM Agent Training
University of Washington · Snowflake · University of North Carolina at Chapel Hill

Jul 31, 2026J/K move · Enter open · S save

Data Turnstile: A Scalable Open Framework for Function-Calling Data Generation

Goutham Ramakrishnan, Megha Sharma

Amazon AGI Bengaluru, India

Tool-Augmented Language Model Agents Small Language Models LLM Tool Use Synthetic Data Generation for Language Models Tool Use Synthetic Data Generation for LLM Agents

Open paper AskPDF arXiv

Small language models (SLMs) are attractive for agentic deployment due to low latency, reduced cost, and on-device privacy, yet they struggle with tool-use tasks where training data is scarce and noisy. Unlike larger models, SLMs cannot compensate for low-quality supervision through sheer capacity, making data quality the critical bottleneck. We present Data Turnstile, an open-source framework that takes user-defined API specifications and generates high-quality synthetic training data for function calling. Turnstile decomposes multi-turn tool-use interactions into constrained, stepwise generation with validation and error-feedback loops, providing fine-grained control over API diversity, conversation complexity, and output correctness. We demonstrate effectiveness of domain adaptation with Turnstile data on two challenging function calling benchmarks. On the BFCL single-turn benchmark, a Qwen3-0.6B fine-tuned on Turnstile data without chain-of-thought achieves 75.9% overall accuracy (versus 67.4% for the base model with thinking enabled), closing the gap with thinking-enabled Qwen3-1.7B (78.4%) and Qwen3-4B (79.9%) despite being 3

\times

and 7

\times

smaller respectively. On

τ^2

-bench, a multi-turn agentic benchmark, Turnstile-trained Qwen3-1.7B achieves 31.1% pass^1 on the Telecom domain, improving 4.7

\times

over its 6.6% base and surpassing Qwen2.5-32B-Instruct (27.4%), a model 19

\times

larger. Turnstile-trained Qwen3-0.6B achieves 24.6%, improving 7

\times

over its 3.5% base and approaching the 32B model (53

\times

larger). We release Data Turnstile along with a dataset spanning 1,000+ APIs and 100K+ multi-turn interactions.

Abstract

Synthetic Data Generation for Language Models

Explore similar work

Data Turnstile: A Scalable Open Framework for Function-Calling Data Generation

State-Grounded Multi-Agent Synthetic Data Generation for Tool-Augmented LLMs

MidTool: Mid-training Data Synthesis for Agentic Tool Use