DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use

Aili Chen; Chi Zhang; Junteng Liu; Jiangjie Chen; Chengyu Du; Yunji Li; Ming Zhong; Qin Wang

cs.AIMar 10, 2026

DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use

Authors: Aili Chen, Chi Zhang, Junteng Liu, Jiangjie Chen, Chengyu Du, Yunji Li, Ming Zhong, Qin Wang,

Organizations: Fudan University · MiniMax · Independent

Abstract

Recent work synthesizes agentic tasks for post-training tool-using LLMs, yet robust generalization under shifts in tasks and toolsets remains an open challenge. We trace this brittleness to insufficient diversity in synthesized tasks. Scaling diversity is difficult because training requires tasks to remain executable and verifiable, while generalization demands coverage of diverse tool types, toolset combinations, and heterogeneous tool-use patterns. We propose DIVE, an evidence-driven recipe that inverts synthesis order, executing diverse, real-world tools first and reverse-deriving tasks strictly entailed by the resulting traces, thereby providing grounding by construction. DIVE scales structural diversity along two controllable axes, tool-pool coverage and per-task toolset variety, and an Evidence Collection--Task Derivation loop further induces rich multi-step tool-use patterns across 373 tools in five domains. Training Qwen3-8B on DIVE data (48k SFT + 3.2k RL) improves by +22 average points across 9 OOD benchmarks and outperforms the strongest 8B baseline by +68. Remarkably, controlled scaling analysis reveals that diversity scaling consistently outperforms quantity scaling for OOD generalization, even with 4x less data.

Figures & tables

Figure 1: Motivation and overview of Dive . Top: Fixed-toolset synthesis and pipeline pooling limit diversity and weaken generalization. Middle: Simulated tools and query-first synthesis for diverse tasks increase unverifiability/unsolvability risk, limiting agentic training. Bottom: Dive performs evidence-first synthesis on diverse, real-world tools, producing verifiable and executable tasks. Radar: Gray: base model; Blue: trained on deep-research data synthesized with a fixed search/browse toolset (strong in-distribution but weak/negative transfer); Purple: trained on Dive with matched data and training budget (robust generalization).

Figure 2: Overview of the Dive framework. (1) Diverse Synthesis Resource Preparation (Left): We construct decoupled pools of tools (spanning general and expert domains), seed concepts, and query-only exemplars with implicit tool-use patterns. (2) Evidence-Driven Task Synthesis (Right): We randomly sample configurations and run an inverted loop where the model executes real tools to collect grounded evidence (a, b) and reverse-derives tasks (query-answer pairs) strictly entailed by traces (c, d), ensuring validity by construction. (3) Agentic Training (Bottom): The synthesized corpus supports effective SFT cold starts and RL using verifiable reference answers.

Tier	Task Family	Benchmark	OOD Factors	Tool Pool	Toolset	Protocol	Env
L1	In-Distribution	Dive-Eval	–	384 Tools (General + Expert)	Per-task	OpenAI	Stateless
L2	General DeepResearch	GAIA , HLE , BrowseComp , Xbench-DS	Task , Set	Search / Browse	Uniform	OpenAI	Stateless
L2	Domain DeepResearch	FinSearchComp (Global)	Task , Set	Search / Browse / Code Execution	Uniform	OpenAI	Stateless
L3	Financial Specialist	Finance Agent Benchmark	Task , Pool , Set	EDGAR / Web / Parse / Retrieve	Uniform	OpenAI	Stateless
	Medical Specialist	MedAgentBench	Task , Pool , Set , Proto , Env	FHIR GET / POST / Finish	Uniform	HTTP	Stateful
	Software Engineering	SWE-bench Verified	Task , Pool , Set , Env	Bash / Search / Editor / Finish	Uniform	OpenAI	Stateful

Table 1: Benchmark taxonomy and OOD factors w.r.t. Dive training data. L2 benchmarks use general-purpose tools; L3 benchmarks require specialized toolsets. OOD Factors : Task =shifted task distribution, Pool =unseen tool pool, Set =unseen toolset, Proto =non-OpenAI protocol, Env =stateful environment.

		L1 In-distribution	L2 OOD w/ General Tools						L3 OOD w/ Specialized Tools
Category	Model	Dive -Eval	GAIA	HLE	BC	XB-DS	FSC 2	FSC 3	FAB	MAB	SWE	Toolathlon
Frontier ( $\gg$ 8B)	Gemini-3-Pro	45.3	80.3	42.9	49.0	76.0	70.6	52.4	39.0	74.8	76.2	36.4
	Claude-4-Sonnet	44.8	63.7	20.8	12.8	62.2	60.2	33.3	39.0	79.3	72.7	29.9
	Gemini-2.5-Pro	29.1	60.2	28.4	9.9	56.0	44.5	27.4	24.0	65.1	59.6	10.5
	DeepSeek-V3.2-Exp	40.4	61.0	17.9	40.1	67.2	61.3	27.4	26.0	67.3	67.8	20.1
	Kimi-K2-0905	32.9	60.0	26.9	14.1	61.0	47.1	10.7	28.0	61.2	69.2	13.0

Table 2: Overall comparison across L1–L3 benchmarks. L1 : in-distribution; L2 : OOD w/ general tools; L3 : OOD w/ specialized tools. BC=BrowseComp; XB-DS=Xbench-DeepSearch; FSC 2 /FSC 3 =FinSearchComp Global-T2/T3; FAB=Finance Agent Benchmark; MAB=MedAgentBench; SWE=SWE-bench Verified. 8B Baselines include specialized agentic models (WebExplorer-8B; our SWE-Dev-8B trained on SWE-Dev ( Wang et al., 2025a ) ) and generalizable agentic models (EnvScaler-8B). Scores are success rates (%). Toolathlon is averaged over 3 runs; all other benchmarks are averaged over 4 runs. Underline : best overall; Bold : best among 8B backbone.

Panel of Figure 3 — Figure 3: Scaling analysis. Gray dashed line: Qwen3-8B base. Left: Diversity-only vs. Quantity-only. Diversity-only expands the tool pool from 1 $\to$ 4 domains (12k fixed; representative path fin $\to$ fin+med $\to$ fin+med+bio $\to$ all). Quantity-only scales data 12k $\to$ 48k with tasks/tools fixed (Gen-DR; Search/Browse-only); diversity yields stronger OOD gains. Middle: Toolset-variety-only vs. Pool-Expansion+Variety. Both scale SFT data 12k $\to$ 48k from Finance. Toolset-variety-only: pool fixed. Pool-Expansion+Variety: pool expands across domains (multiple paths); pool expansion sustains gains. Right: All-path scaling (SFT $\to$ RL). 24 domain-expansion permutations; SFT 12k $\to$ 48k and RL 0.8k $\to$ 3.2k. Thin: paths; thick: mean; shaded: interquartile range; RL amplifies scaling.

Figure 4: RL training dynamics over 100 steps. Accuracy reward, tool calls/task, unique tool-call graphs, and unique R/P topologies. Light: per-step; dark: smoothed. Percent changes use smoothed start/end values. Each step uses a 512-task RL batch, so per-step unique counts are upper-bounded by 512. Full 16-panel version in Appendix Figure 6 .

Diversity Metric	Gen-DR	Dive	$\Delta$
Tools covered	2	373	+186 $\times$
Unique toolsets	1	46,398	+46k $\times$
Unique tool-call sequences	1,231	25,084	+20 $\times$
Unique tool-call graphs	19,442	39,810	+105%
Unique R/P topologies	12,315	23,450	+90%
R/P topology classes covered	65	153	+135%

Table 3: Diversity comparison: Gen-DR vs. Dive (48k each; same teacher model, rejection sampling).

Figure 5: R/P topology density and tool-frequency distributions (48k SFT). Left: Density over R/P topology classes (153 observed; retrieval-only

\to

mixed

\to

processing-only , i.e., PureR

\to

R+P

\to

PureP; taxonomy in Appendix D ). Right: Tool-call frequency over 373 tools (5 domains).

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Domain	Retrieval	Processing	Total	P%
Financial	155	14	169	8.3%
Medical	80	9	89	10.1%
Academic	43	7	50	14.0%
Biological	18	43	61	70.5%
General	3	1	4	25.0%
Total	299	74	373	19.8%

Appendix

Table 4: Tool pool summary. Distribution of Retrieval (R: information fetching) and Processing (P: computation/transformation) tools across domains.

Domain	Source Name	URL / Documentation
Financial	Tushare	https://tushare.pro/
Medical	RxNorm API	https://lhncbc.nlm.nih.gov/RxNav/APIs/RxNormAPIs.html
	RxClass API	https://lhncbc.nlm.nih.gov/RxNav/APIs/RxClassAPIs.html
	Clinical Tables	https://clinicaltables.nlm.nih.gov/
Academic	Semantic Scholar	https://www.semanticscholar.org/product/api
	OpenAlex	https://docs.openalex.org/

Appendix

Table 5: Tool sources and documentation. Official documentation or API endpoints for the tools used in Dive .

Category	Domain	R/P	Count	Tools
Conv. Bonds	Financial	P	1	cb_factor_pro
ETF Data	Financial	P	1	etf_adj
Mutual Funds	Financial	P	1	fund_factor
Indices	Financial	P	1	index_factor
Stock Metrics	Financial	P	10	stock_adj_factor, stock_ah_ratio, stock_cyq_perf, stock_daily_basic, stock_fina_indicator, stock_fina_indicator_vip, stock_stk_ah_comparison, stock_stk_factor, stock_stk_factor_pro, stock_stk_nine_turn
Bonds/Rates	Financial	R	7	bc_bestotcqt, bc_otcqt, libor, repo_daily, us_tbr, us_trycr, us_tycr

Appendix

Table 6: Complete tool pool (373 tools). Aggregated by category. R/P denotes Retrieval (R) or Processing (P) tool type.

Benchmark	Task Description
WebArena ( Zhou et al., 2023 )	Long-horizon realistic web information seeking
AgentBench ( Liu et al., 2023 )	Comprehensive evaluation across multiple environments
ToolBench ( Qin et al., 2023 )	Diverse instruction following with real-world APIs
DSBench ( Jing et al., 2024 )	Data analysis and SQL/Python code generation
BrowseComp ( Wei et al., 2025 )	Web navigation and information extraction
Mind2Web ( Deng et al., 2023 )	Generalizable web interactions across domains

Appendix

Table 7: Exemplar sources. We sample 3,000 tasks from the following agentic benchmarks to serve as structural priors for task synthesis. These benchmarks are selected for their alignment with Dive ’s focus on domain-specific knowledge, multi-hop retrieval, and complex processing.

Metric	K=1	K=2	K=3	$\Delta$ (1 $\to$ 3)
GPT-OSS-120B Pass Rate (%)	66.8	51.1	41.2	$-$ 38%
Avg. Tool Calls per Task	4.6	7.5	10.1	+120%
Tool Coverage (%)	58.2	74.3	92.2	+58%
Unique Tool-call Sequences	892	1,321	2,348	+163%
Unique Tool-call Graphs	1,362	2,467	3,551	+161%
Unique R/P Topologies	542	1,004	2,393	+341%

Appendix

Table 8: Structural diversity across iteration rounds. We sample 4,000 tasks and measure diversity metrics in trajectories generated by GPT-OSS-120B. All diversity metrics increase substantially with more iterations, while pass rate decreases (indicating harder tasks).

Level	Dimension	Values	Definition
1. R/P Type	PureR	–	Only Retrieval tools
	R+P	–	Both Retrieval and Processing
	PureP	–	Only Processing tools
2. Structure (priority order)	Single	$n=1$	Single tool call
	Indep	$n>1\land\|E\|=0$	Multiple independent calls
	Phain	$e=n{-}1\land\max(\text{in/out})\leq 1$	Linear chain

Appendix

Table 9: Topology class dimensions (3-level hierarchy). Structure types are checked in priority order and are mutually exclusive.

Structure	Scale Params	Bins	Classes
Single	none	1	$3\times 1=3$
Indep	node count	5	$3\times 5=15$
Phain	depth only	4	$3\times 4=12$
Fork	depth $\times$ width	$4\times 4$	$3\times 16=48$
Join	depth $\times$ width	$4\times 4$	$3\times 16=48$
DAG	depth $\times$ width	$4\times 4$	$3\times 16=48$

Appendix

Table 10: Theoretical class count per structure type (

\times

3 R/P types).

Figure 6: RL training dynamics by domain over 100 steps. 4 rows (metrics: Accuracy, Calls/Task, Tool-call Graphs, R/P Topologies)

\times

4 columns (domains: Academic, Biological, Financial, Medical). All domains show consistent accuracy improvement, while structural diversity (tool-call graphs and R/P topologies) also increases during RL. Percentages are relative changes between smoothed start/end values.

Condition	Setting	FSC 2	FSC 3	FAB	MAB	GAIA	BC	HLE	XB-DS	SWE	Tool.
Scaling Diversity (12k fixed, tool pool expanding)
1 domain	fin (174 tools)	57.1	19.0	14.0	38.8	32.0	6.5	8.5	31.2	10.9	0.9
2 domain	fin_med (263 tools)	60.7	23.4	15.5	43.6	45.5	9.8	11.5	44.2	11.2	1.5
3 domain	fin_med_bio (324 tools)	61.2	27.4	17.0	45.2	47.1	10.8	12.3	48.6	11.8	2.5
4 domain	fin_med_bio_aca (373 tools)	62.0	29.8	18.0	47.6	50.5	12.2	13.4	50.1	12.5	2.9
Scaling Quantity – Gen-DR (search/browse only, 2 tools fixed)

Appendix

Table 11: Raw data for Figure 2(a): Diversity vs. Quantity Scaling. All results are SFT with 300 steps. Diversity Scaling : Fixed 12k trajectories, expanding tool pool (1

\to

4 domains, 174

\to

373 tools). Quantity Scaling : Fixed general-purpose search/browse tools (2 tools), expanding data (12k

\to

48k).

Condition	Setting	FSC 2	FSC 3	FAB	MAB	GAIA	BC	HLE	XB-DS	SWE	Tool.
Config Scaling (fin domain fixed, 174 tools)
12k	fin	57.1	19.0	14.0	38.8	32.0	6.5	8.5	31.2	10.9	0.9
24k	fin	60.6	22.9	15.0	41.3	41.6	8.9	9.8	37.6	11.3	1.2
36k	fin	61.6	24.2	16.0	41.8	45.5	9.8	10.5	40.5	11.5	1.5
48k	fin	62.1	25.3	16.5	42.0	47.0	10.3	11.0	42.8	11.5	1.5
Pool+Config Scaling (all paths starting from fin, tool pool expanding)

Appendix

Table 12: Raw data for Figure 2(b): Config Scaling vs. Pool+Config Scaling. All results are SFT with 300 steps, starting from Financial domain (12k). Config Scaling : Fixed tool pool (fin, 174 tools), expanding data (12k

\to

48k) and configurations. Pool+Config Scaling : Jointly expanding tool pool diversity (174

\to

373 tools) and data quantity, all paths starting from fin.

Domains	Combo	L2-A	L2-B	L3-A	L3-B	L3-C	L3-D	Avg	$\boldsymbol{\Delta}$ Avg
1 domain 12k	fin	19.6 $\to$ 29.0	38.0 $\to$ 44.8	14.0 $\to$ 22.0	38.8 $\to$ 49.9	10.9 $\to$ 11.9	0.9 $\to$ 3.7	20.4 $\to$ 26.9	+6.5
	med	20.8 $\to$ 31.1	37.0 $\to$ 44.0	12.0 $\to$ 16.0	47.1 $\to$ 52.1	12.2 $\to$ 12.4	2.3 $\to$ 4.4	21.9 $\to$ 26.7	+4.8
	aca	30.1 $\to$ 31.0	37.6 $\to$ 46.5	10.0 $\to$ 16.0	42.3 $\to$ 44.8	14.2 $\to$ 14.7	1.5 $\to$ 2.5	22.6 $\to$ 25.9	+3.3
	bio	25.3 $\to$ 30.1	38.2 $\to$ 37.1	6.0 $\to$ 8.0	39.9 $\to$ 50.3	11.0 $\to$ 15.9	1.0 $\to$ 2.5	20.2 $\to$ 24.0	+3.7
	Avg	23.9 $\to$ 30.3	37.7 $\to$ 43.1	10.5 $\to$ 15.5	42.0 $\to$ 49.3	12.1 $\to$ 13.7	1.4 $\to$ 3.3	21.3 $\to$ 25.9	+4.6
2 domain 24k	fin-med	29.5 $\to$ 31.8	42.8 $\to$ 45.5	16.0 $\to$ 26.0	48.3 $\to$ 56.4	11.9 $\to$ 14.2	3.1 $\to$ 4.0	25.3 $\to$ 29.7	+4.4

Appendix

Table 13: Scaling diversity & quantity (SFT

\to

RL). Each cell shows category score after SFT and after RL (SFT

\to

RL).

\Delta

Avg denotes Avg RL -Avg SFT . L2-A averages GAIA/XB-DS/BC/HLE; L2-B averages FSC 2 /FSC 3 .

Tool-Augmented Language Model Agents

Tool-augmented language model agents use external tools, APIs, and environments to solve tasks beyond text generation alone.

See all 167 papers

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis
Oct 27, 2026Andrew Szot, Samira Abnar, Omar Attia +5Tool-Using Agents LLM Agent Training
Apple
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
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

Oct 27, 2026J/K move · Enter open · S save

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

Andrew Szot, Samira Abnar, Omar Attia, Abbas Kazerouni+4

Apple

Tool-Using Agents LLM Agent Training Synthetic Data Generation for LLM Agents Scaling Laws

Open paper AskPDF arXiv

We explore how to build performant large language model (LLM) tool-use agents that complete complex, multi-step tasks. Specifically, we study how data composition in synthetically generated tool-use environments affects the training of such agents. To do so, we propose the SynthAPI pipeline, which explicitly controls the diversity of generated environments along the axes of tools, tasks, and the databases on which the tools operate. By systematically varying SynthAPI's generation controls in 458,752 tasks across 849 trained models of four different sizes, we derive a scaling law that predicts downstream tool-use agent generalization based on the scale and composition of the synthesized environments. These predictable relationships reveal several lessons for scaling tool-use agent data. As data budgets grow, scaling tool and database diversity becomes more effective than increasing the number of tasks, and increasing data complexity yields larger gains. Scaling according to these findings yields agents that outperform models trained on data from prior work in large-scale tool-use agent and synthetic environment generation across four evaluation benchmarks.

Abstract

Tool-Augmented Language Model Agents

Explore similar work

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

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

MidTool: Mid-training Data Synthesis for Agentic Tool Use