Paper deep dive
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, Zhengmao Zhu, Jiayuan Song, Ke Ji, Junxian He, Pengyu Zhao, Yanghua Xiao
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 3/13/2026, 1:08:33 AM
Summary
DIVE is an evidence-driven framework for synthesizing diverse, verifiable, and executable agentic tasks to improve LLM tool-use generalization. By inverting the synthesis process—executing real-world tools first and reverse-deriving tasks from the resulting traces—DIVE ensures grounding by construction. The method scales structural diversity through tool-pool coverage and per-task toolset variety, significantly outperforming quantity-based scaling in OOD generalization benchmarks.
Entities (4)
Relation Signals (3)
DIVE → trains → Qwen3-8B
confidence 100% · Training Qwen3-8B on DIVE data (48k SFT + 3.2k RL) improves by +22 average points
DIVE → improves → Tool-Use Generalization
confidence 95% · Extensive experiments demonstrate that DIVE significantly improves tool-use generalization.
Evidence Collection–Task Derivation Loop → scales → Structural Diversity
confidence 90% · DIVE scales structural diversity along two controllable axes... and an Evidence Collection–Task Derivation loop further induces rich multi-step tool-use patterns
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2603.11076v1
- Canonical: https://arxiv.org/abs/2603.11076v1
Trouble viewing inline? Open PDF directly →
Full Text
91,944 characters extracted from source content.
Expand or collapse full text
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 ♣ Zhengmao Zhu ♣ Jiayuan Song ♣ Ke Ji ♣ Junxian He ♣ Pengyu Zhao ♣ Yanghua Xiao ♠† ♠ Fudan University ♣ MiniMax ♢ Independent alchen20@fudan.edu.cn shawyh@fudan.edu.cn https://sheep333c.github.io/DIVE/ Abstract Recent work synthesizes agentic tasks for post- training tool-using LLMs, yet robust generaliza- tion under shifts in tasks and toolsets remains an open challenge. We trace this brittleness to insufficient diversity in synthesized tasks. Scal- ing 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, exe- cuting diverse, real-world tools first and reverse- deriving tasks strictly entailed by the resulting traces, thereby providing grounding by construc- tion. 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 con- sistently outperforms quantity scaling for OOD generalization, even with 4× less data. 1. Introduction Recent work on agentic post-training increasingly relies on synthesized agentic tasks, improving LLMs’ ability to use general-purpose tools such as web search and code execution (Yao et al., 2024; Xu et al., 2024; Liu et al., 2025b; Froger et al., 2025). However, in practical deployments, these models often struggle with the open-ended diversity of tool use (Zhang et al., 2025): tasks range from open- domain queries (e.g., “what is the capital of Australia?”) to †Corresponding author. DIVE: Evidence-Driven Synthesis with Diverse, Real-world Tools ... ... Synthesize Diverse Agent Task for Generalization DeepResearch Search / Browse Pipeline APipeline B ... ... LLM / Search / Code Toolset Descrip;ons "tool_name": "rxnorm_get_drugs" "parameters": "name": "itraconazole" "response": "rxcui": “152854", “name": "Sporanox 10 MG/ML Oral Solution”...... Execute Real-world Tools ... ... Collect EvidenceDerive Tasks (Q, A) from Evidence Question: A patient with blastomycosis has severe dysphagia and can only swallow 10mL of liquid at a time. They need 400mg of an azole antifungal daily. Among itraconazole, fluconazole, and voriconazole oral suspensions, which one allows the patient to receive 400mg in a single ≤10mL dose? Answer: Fluconazole or Voriconazole Low Diversity Costly Scaling Insufficient Diversity in Training Tasks Weak Generalization Unverifiable Risk Unsolvable Risk 2. Extracting Data from Specified Pipelines 1. Synthesize Specific Tasks with Fixed Toolset Diverse Tasks but Unverifiable and Unsolvable Limits Training 3. Simulating Diverse Tools 4. Query-First Synthesis Task A Task B Simulated Environments Hypothe;cal Query Scaling diversity is close to (manually) scaling pipelines. Suffer from the unreliability of tools. Suffer from heavy quality checking. Synthesize Diverse Tasks with Varied Toolsets while Keeping Verifiable and Executable. Figure 1. Motivation and overview of DIVE. Top: Fixed-toolset synthesis and pipeline pooling limit diversity and weaken gener- alization. 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). domain-specific tasks such as clinical diagnosis, financial analysis, and software engineering (Jiang et al., 2025; Hu et al., 2025; Jimenez et al., 2023), while tools vary from general-purpose ones (e.g., web search) to specialized ones such as protein retrieval and email management (Mitchener et al., 2025; Xu et al., 2024). This motivates our primary research question: how can we improve the generalization of tool-using LLMs across real-world tasks and toolsets? 1 arXiv:2603.11076v1 [cs.AI] 10 Mar 2026 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use We argue that a key bottleneck is insufficient diversity in synthesized training tasks, which limits generalization un- der task and toolset shifts. Most existing synthesis recipes scale data primarily by quantity or difficulty, but remain confined to narrow task families and fixed toolsets (Fig- ure 1(1); e.g., deep-research tasks equipped with web search tools) (Liu et al., 2025b; Li et al., 2025d). As a result, while these agents perform well on in-distribution tasks, they of- ten over-rely on rigid routines (e.g.,search→browse loops) (He et al., 2025; Fu et al., 2025), leading to poor gen- eralization or even negative transfer on new task families and toolsets (Figure 1 radar), e.g., when asked to perform clinical diagnosis using tools like PatientLookup. However, scaling diversity while maintaining data quality is challenging because effective agentic training requires synthesized tasks to be both verifiable and executable for trajectory filtering and reward computation. This creates a fundamental tension: (1) Structural Diversity: beyond diverse tool types and per-task toolset combinations, tasks should involve heterogeneous multi-step tool-use patterns (e.g., retrieval-only→retrieval-then-analyze), rather than template substitutions (e.g., changing query entities); but (2) Grounded Validity: as diversity grows, ensuring every synthesized task remains solvable and verifiable under its specific toolset becomes increasingly difficult. Current ap- proaches fail to reconcile this tension (Figure 1): extracting data from specialized pipelines is costly, as scaling diver- sity requires manually scaling pipelines (Liu et al., 2025a); simulating tool environments with LLMs or generic tools suffers from the unreliability of simulated tools, leading to unverifiable risks (Castellani et al., 2025; Mitra et al., 2024); and query-first synthesis on real tools suffers from heavy quality checking to mitigate the unsolvable risk of hypothetical queries (Qin et al., 2023; Shen et al., 2024). To bridge this gap, we invert the synthesis order on di- verse, real-world tools. Rather than generating task queries first and then checking validity post hoc, we execute tools first and derive tasks from the resulting traces. This yields grounding by construction: executability follows from real tool traces, and verifiability follows from observable tool outputs. Simultaneously, we scale structural diversity by expanding tool-pool coverage and per-task toolset variety; real executions then yield tasks that are both grounded and structurally diverse, with heterogeneous tool-use patterns. Specifically, we propose DIVE (Figure 1 bottom), an evidence-first recipe that automatically synthesizes Diverse, Verifiable, and Executable agentic tasks. Starting from Retrieval and Processing tool-use primitives, we construct three resource pools: 373 validated tools spanning general- purpose and four expert domains, domain-specific seed con- cepts, and diverse query-only exemplars. Each synthesis cy- cle randomly samples a toolset, a seed, and exemplars, then runs a two-stage loop: (i) Evidence Collection interleaves multi-step reasoning with real tool use to gather logically related evidence and dynamically induce diverse tool-use patterns, and (i) Task Derivation observes and reorganizes the accumulated evidence to reverse-derive grounded query– answer pairs strictly entailed by the traces; as evidence grows across iterations, it further refines tasks to remain grounded while increasing diversity. Finally, we apply these synthesized tasks to train agents via SFT and RL, validating their effectiveness for robust generalization. In summary, our contributions are: •We investigate diversity scaling in agentic task synthesis for generalizable tool use, identifying two coupled data requirements: grounded validity (verifiable/executable under assigned toolsets) and structural diversity (hetero- geneous patterns beyond template variation). •We propose DIVE, an evidence-driven recipe that synthe- sizes diverse, verifiable, and executable agentic tasks at scale by inverting the synthesis order: executing diverse, real-world tools first and reverse-deriving tasks. •Extensive experiments demonstrate that DIVE signifi- cantly improves tool-use generalization. Our analysis reveals that diversity scaling outperforms quantity scaling, and that RL benefits are amplified by diverse training data. 2. Related Work Tool-Use Agents and Benchmarks.Tool-use agents must operate under diverse constraints: varying toolsets, invoca- tion protocols, and interaction environments (Qin et al., 2023; Tang et al., 2023; Yao et al., 2024). Benchmarks now reflect this diversity, spanning web research (Mialon et al., 2023; Wei et al., 2025; Chen et al., 2025), software engineer- ing (Jimenez et al., 2023), domain applications (Hu et al., 2025; Choi et al., 2025; Jiang et al., 2025), and universal tool suites (Li et al., 2025b; Wang et al., 2025b; Guo et al., 2025). These benchmarks highlight a central challenge: generalizable tool use across shifting task distributions and toolsets, which is the focus of this work. Synthetic Data for Tool-Use Agent Training. Scaling synthetic tasks and trajectories for SFT and RL is a prevail- ing paradigm for training tool-use agents (Qin et al., 2023; Tang et al., 2023; Mitra et al., 2024; Liu et al., 2025b; Li et al., 2025d;c). Most prior work designs synthesis pipelines tailored to fixed task types and toolsets to optimize agent performance, e.g., deep-research tasks equipped with gen- eral web search tools (Liu et al., 2025b; Li et al., 2025d; Wu et al., 2025a;b; Qiao et al., 2025). However, this results in training data with limited diversity, hindering tool-use generalization to diverse unseen scenarios (He et al., 2025; Fu et al., 2025). A common engineering practice involves extracting data from specialized pipelines (Liu et al., 2025a; Team et al., 2025), but this heuristic is costly and scales 2 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use poorly as each task type or environment demands a cus- tomized synthesis pipeline. To inherently scale diversity, other works attempt to simulate diverse toolsets via LLMs or generic tools (e.g., search, code execution) (Mitra et al., 2024; Fang et al., 2025; Castellani et al., 2025; Li et al., 2025e). While achieving scalability, they risk unstable mock execution where tasks solvable during synthesis may fail verification during training. Conversely, methods targeting real toolsets typically follow a query-first paradigm (Qin et al., 2023; Shen et al., 2024; Li et al., 2025b; Guo et al., 2025), creating a verification bottleneck: tasks derived from documentation are often non-executable, and manual veri- fication is costly, hindering scalable RL. In contrast, DIVE guarantees executability and verifiability by construction via an inverted synthesis process on diverse, real-world tools. 3. DIVE In this work, we aim to improve tool-use generalization by scaling diversity in agentic task synthesis. We propose DIVE, an automated recipe designed to achieve this goal while ensuring training stability. After introducing prelimi- naries (§3.1), we describe DIVE in three phases: (1) Diverse Synthesis Resource Preparation (§3.2), which builds decou- pled pools of tools, seeds, and exemplars to support scalable synthesis; (2) Evidence-Driven Task Synthesis (§3.3), which reverse-derives tasks from grounded execution traces; and (3) Agentic Training with DIVE Tasks (§3.4), which opti- mizes the agentic LLM via supervised finetuning (SFT) and reinforcement learning (RL). 3.1. Problem Formulation Tool-Using Agent. We formulate tool use as a sequen- tial decision process. Given a task query푄and a toolset 풯, an agent policy휋 휃 performs interleaved reasoning and tool use (Yao et al., 2023) to solve the problem. At step푡, the agent generates a thought푟 푡 and an action 푎 푡 ∈ 풯based on the history; the environment executes 푎 푡 and returns an observation표 푡 . This yields a trajectory 휏 = (푟 1 ,푎 1 ,표 1 ,...,푟 푇 ,푎 푇 ,표 푇 ). Task Synthesis Objectives. We aim to synthesize an agentic task dataset풟 =(푄 (푖) ,퐴 (푖) ,풯 (푖) ), where each instance comprises a task query푄 (푖) , a reference answer퐴 (푖) (for verification), and a unique toolset풯 (푖) . To support general- ization and effective training,풟must satisfy four rigorous properties: (1) Structurally Diverse: Tasks should cover di- verse toolsets and exhibit heterogeneous tool-use patterns to support generalization. (2) Verifiable: Each task must have a deterministic verifier (e.g., by comparing the output to a reference answer) to ensure trajectory filtering and reward computation. (3) Executable: Each task must be solvable under its specific toolset풯 (푖) , guaranteeing at least one fea- sible solution path to avoid optimization noise. (4) Scalable: The synthesis pipeline must be autonomous, enabling data volume to scale with compute resources. 3.2. Diverse Synthesis Resource Preparation The diversity of synthesized data is inherently constrained by the richness of its underlying resources. Accordingly, prior to synthesis, we pre-construct a large, heterogeneous resource bank to support scalable and diverse task synthesis. We decompose this bank into three diverse and decoupled pools: (1) a tool pool defining a broad action space; (2) a seed pool providing long-tail semantic coverage; and (3) an exemplar pool offering heterogeneous structural priors. By decoupling these resources, we can exponentially expand task diversity through their independent sampling and re- combination, covering a vast space of domain knowledge, tool capabilities, and reasoning structures. Tool Pool: Broad Action Space. We start from two common generic tools: web search and code execution, which reveal a functional duality: search exemplifies the Re- trieval primitive for acquiring external information, while code execution represents the Processing primitive for performing deterministic transformation. To systemati- cally diversify tool use beyond generic utilities, we in- stantiate these two primitives with domain-specific tools across four expert domains: Finance, Biology, Medicine, and Academia. We construct the tool pool with a Crawl– Validate pipeline. (1) Crawl. Firstly, we crawl public APIs and wrap them for tool-calling, labeling each tool as Retrieval (e.g.,ncbi_search) or Processing (e.g., seq_translate). (2) Validate. To ensure training sta- bility, we filter candidates via unit tests for correctness, concurrency safety, and response consistency, yielding a final set of 373 robust tools (more details in Appendix B.1). Seed Pool: Diverse Semantic Anchors. Synthetic genera- tion can suffer from topic collapse, over-sampling generic, high-frequency concepts (Wang et al., 2023; Gudibande et al., 2024). To promote distributional diversity, we build a registry of seed concepts as anchors. We mine four do- mains: Wikipedia (Bridge, 2001), PubMed (National Li- brary of Medicine (US), 2026), NCBI (Sayers et al., 2025), and global stock exchanges (Yahoo Finance, 2026), yield- ing∼5,000 entity seeds per domain via LLM extraction. Anchoring synthesis on specific entities (e.g., “Erlotinib”) rather than generic terms (e.g., “medicine”) encourages exploration of sparse tool-space regions. Exemplar Pool: Heterogeneous Task Priors. In con- trast to synthesis methods constrained by fixed tasks and toolsets (Wang et al., 2025c; Li et al., 2025d; Wu et al., 2025a), open-ended generalization necessitates a broader spectrum of task forms. Accordingly, we construct a reposi- tory of query-only exemplars sourced from heterogeneous task families. Although exemplars contain no execution 3 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Config Sampling a. Evidence Collection Scaling Diverse Resources General Biological Academic Medical Financial ncbi_search(...) search(...) get_papers(...) find_rxcui(...) daily_stock(...) seq_translate(...) analyze_text(...) get_similarity(...) stock_forecast(...) code_execution(...) Biological Retrieval Tools Domain Processing Tools Scaling Diverse Tasks: Evidence-Driven Synthesis Implicit Tool-call Patterns Queries b. Grounded Evidence Retrieval: Processing: (a) (b) (c) (d)(e)......(f) Exemplar 2 Exemplar N Exemplar 1 Diverse To ol Pool Construction Diverse Seed Pool Construction Diverse Exemplar Pool Construction Agent Training with DIVE Dataset c. Task Derivation d. DIVE Tasks On December 20, 2024, which securities firm branches executed the "Quantitative Board Hitting" strategy in hot money trading? On December 20, 2024, what percentage of Huaxin Securities' total "Quantitative Board Hitting" net trading amount was conducted through its Shanghai Branch versus Chengdu Branch? Round to the nearest whole number. Reference answer is 57 ; 43 Exemplar List Query database: “What percentage of orders in 2023 were shipped to California?” “What integer-rounded percentage of the total length of the harlequin shrimp recorded...” ... Seed Concept Toolset Stock_Related Future_Related General_Tools Index_Related ...... 000 001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050 051 052 053 054 DIVE: Mining Diverse, Verifiable, Executable Tasks in Real Tool Worlds for Training Generalist Agents Anonymous Authors 1 Abstract 1. Introduction (Q k ,A k )=F evol (Q k→1 ,E k |X) E k =F exp (Q k→1 ,E k→1 |T) C=T,S,X Q 0 =S E 0 =→ ω=(r 1 ,a 1 ,o 1 ),...,(r T ,a T ,o T ) 2. Related Work 3. Method 4. Training on DiVE Tasks 5. Experimemts 6. Analysis Conclusion 1 Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>. Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute. 1 opt_daily fut_daily stock_express stock_hm_detail code_execution stock_margin X X Interleaved reasoning with real tools. <think><tool_call><tool_response> 000 001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050 051 052 053 054 DIVE: Mining Diverse, Verifiable, Executable Tasks in Real Tool Worlds for Training Generalist Agents Anonymous Authors 1 1. Introduction (Q k ,A k )=F der (Q k→1 ,E k |X) E k =F col (Q k→1 ,E k→1 |T) ω k =(r 1 ,a 1 ,o 1 ,...,r T ,a T ,o T ) C=T,S,X Q 0 =S E 0 =→ D task =(Q (i) ,A (i) ,T (i) ) N i=1 D task =(query,gold answer,toolset) N i=1 1 Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>. Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute. 1 000 001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050 051 052 053 054 DIVE: Mining Diverse, Verifiable, Executable Tasks in Real Tool Worlds for Training Generalist Agents Anonymous Authors 1 1. Introduction (Q k ,A k )=F der (Q k→1 ,E k |X) E k =F col (Q k→1 ,E k→1 |T) ω k =(r 1 ,a 1 ,o 1 ,...,r T ,a T ,o T ) C=T,S,X Q 0 =S E 0 =→ D task =(Q (i) ,A (i) ,T (i) ) N i=1 D task =(query,gold answer,toolset) N i=1 1 Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>. Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute. 1 000 001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050 051 052 053 054 DIVE: Mining Diverse, Verifiable, Executable Tasks in Real Tool Worlds for Training Generalist Agents Anonymous Authors 1 1. Introduction (Q k ,A k )=F der (Q k→1 ,E k |X) E k =F col (Q k→1 ,E k→1 |T) ω k =(r 1 ,a 1 ,o 1 ,...,r T ,a T ,o T ) C=T,S,X Q 0 =S E 0 =→ D task =(Q (i) ,A (i) ,T (i) ) N i=1 D task =(query,gold answer,toolset) N i=1 1 Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>. Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute. 1 stock_hm_detail Exemplars Old Task Evidence New Task Old Evidence New Evidence DIVE Task Dataset Reference Answer Toolset Agentic SFT for Cold StartAgentic RL Teacher DIVE-SFT Trajectories DIVE-RL RejecRon SamplingRL SFT Pass@k Filtered Tasks “Speculative Capital” Generate tasks with growing evidence Query 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. traces, each provides structural priors: (1) a query phrasing; and (2) an implicit tool-use pattern, e.g., “Query database: what percentage of orders in 2023 were shipped to Califor- nia?” (Figure 2) implies a retrieve-then-compute structure. Drawing from diverse exemplars broadens the space of de- rived task forms (see Appendix B.2 for sources). 3.3. Evidence-Driven Task Synthesis Given the diverse resource (§3.2), DIVE synthesizes train- ing tasks via an evidence collection and task derivation loop (Figure 2 Right). Each synthesis samples a synthesis configuration, then executes tools first to collect grounded evidence (Figure 2a,b), and finally derives tasks that are strictly supported by the evidence (Figure 2c,d). Configuration Random Sampling. To achieve diverse yet grounded synthesis, we begin each synthesis cycle by sam- pling a synthesis configuration퐶 =풯 ,푆,풳(Figure 2 Right-Config Sampling). (1) Seed Sampling: We draw a seed concept푆from the seed pool to anchor the semantic context. (2) Toolset Sampling: Conditioned on the seed’s do- main, we sample a compatible toolset풯(|풯|∈ [15, 50]) to define the execution environment. (3) Exemplar Sampling: We sample a small set of query-only exemplars풳(typically 3–5) as lightweight form-level cues. Randomly composing these components creates a vast space of configurations, providing diverse starting points for synthesis. Collect Evidence During Interleaved Reasoning with Real Tools. At each iteration푘, we invoke an evidence col- lector agent to expand the evidence frontier (i.e., grounded tool execution traces with outputs) by executing real tools under the configuration풯(Figure 2a). The collector is conditioned on the evolving synthesis context: (i) the cur- rent inquiry푄 푘−1 (initialized as푄 0 = 푆and updated via the derivation step); (i) the accumulated evidence퐸 푘−1 (where퐸 0 = ∅); and (i) the available toolset풯. Op- erating within these bounds, the agent performs a multi- step rollout (up to푇 max steps) to produce a trajectory 휏 푘 = (푟 1 ,푎 1 ,표 1 ,...,푟 푇 ,푎 푇 ,표 푇 ), where푟 푡 denotes the rea- soning thought,푎 푡 the tool invocation (function and argu- ments), and표 푡 the real execution return. We define this update as accumulating validated pairs(푎 푡 ,표 푡 )from휏 푘 into the evidence set (Figure 2b): 퐸 푘 =ℱ col (푄 푘−1 ,퐸 푘−1 |풯 ).(1) By enforcing execution-first collection, we ensure that every evidence item in퐸 푘 is grounded and replayable, imposing strict executability constraints on task derivation. Derive and Refine Tasks with Growing Evidence. Fol- lowing the collection step, we invoke a task generator LLM to synthesize a task grounded in the execution traces (Fig- ure 2c,d). The generator maps the prior query state and current evidence to a new query–answer pair: (푄 푘 ,퐴 푘 ) =ℱ der (푄 푘−1 ,퐸 푘 |풳),(2) 4 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use where푄 0 is initialized as the seed푆(and subsequent 푄 푘−1 are inherited). Conditioned on exemplars풳, the generator instantiates diverse query forms and implicit tool- use patterns (e.g., multi-hop retrieval or retrieve–compute pipelines) by composing evidence from퐸 푘 . Crucially, while 푄 푘 may vary in form, its content remains strictly grounded in퐸 푘 , and퐴 푘 is derived directly from this evidence (see Appendix A for synthesized examples). Iterative Synthesis Loop. We execute the collection– derivation loop for퐾iterations, progressively increasing the diversity of both the evidence set and synthesized tasks (Appendix C) while keeping them grounded in real tool execution. Each iteration is confined to a sampled toolset풯: collection expands퐸 푘 using only풯, and the derived query 푄 푘 becomes the basis for the next collection step, forming a closed-loop curriculum. This shared constraint guarantees executability and verifiability by construction: since푄 푘 is instantiated by composing elements from퐸 푘 , its implicit solution corresponds to a sub-trace of tool calls under풯; therefore, a valid trajectory over풯exists to recover the reference answer퐴; since퐴is derived from tool-returned outputs, it is deterministically verifiable. We store refined tasks and aggregate the dataset for agentic training (§3.4): 풟 task =(푄 (푖) 퐾 ,퐴 (푖) 퐾 ,풯 (푖) ) 푁 푖=1 .(3) 3.4. Agentic Training with DIVE Tasks We post-train an agent on the synthesized DIVE tasks with a two-stage scheme (Figure 2 Bottom): a supervised cold start to acquire reliable tool-calling, followed by reinforcement learning to improve robustness and generalization under diverse toolsets. Each training instance is a tuple(푄,퐴,풯 ), where풯is the same toolset under which the task was syn- thesized, and 퐴 is the reference answer. Agentic SFT for Cold Start. We generate SFT demonstra- tions by rolling out a strong teacher policy on(푄,퐴,풯 ) tuples and apply rejection sampling: a trajectory is accepted if ˆ 퐴 matches퐴; otherwise, the task is discarded. The re- sulting dataset풟 sft = (푄 (푖) ,퐴 (푖) ,풯 (푖) ,휏 (푖) )consists of tasks with verified trajectories. Agentic RL. Starting from the SFT checkpoint, we further optimize the agent’s policy in the same toolset풯to enhance robustness. Before RL updates, we estimate task learnability via self-sampling: for each task we generate푘 rl rollouts un- der풯and score their final answers against퐴. We then filter for frontier tasks where the success rate falls within a learn- able range, yielding a dataset풟 rl =(푄 (푖) ,퐴 (푖) ,풯 (푖) ) 푁 rl 푖=1 . During RL, the policy interacts with tools to produce a final answer ˆ 퐴 and receives a composite reward 푅 = 훼푅 format + 푅 correct ,(4) where푅 correct reflects the correctness of ˆ 퐴 w.r.t.퐴, and 푅 format penalizes invalid tool calls. 4. Experiments To investigate whether diversity scaling during synthesis translates to broad generalization, we evaluate DIVE across 9 benchmarks spanning diverse tasks and toolsets. We detail our experimental setup and multi-level benchmark taxon- omy (§4.1) and present main results in §4.2. 4.1. Experimental Setup DIVE Synthesis Details. We instantiate both the evidence collector and task generator with Claude-4-Sonnet (An- thropic, 2025). Each synthesis cycle samples a configura- tion: a seed concept, a 15–50 tool subset (randomly shuf- fled), and 3–5 query exemplars. The evidence collector performs up to 6 tool-calling steps per iteration; the task generator derives a grounded QA pair in a single reason- ing pass. We run퐾=3collection–derivation iterations per cycle. All tool executions are performed against live tools. Training Details. We use Qwen3-8B (Yang et al., 2025) as our backbone. (a) SFT: From a pool of 114k tasks, we use GPT-OSS-120B (Agarwal et al., 2025) as teacher to collect 48k trajectories (with rejection sampling) for fine-tuning (300 steps, batch size 64, learning rate 1e-5, max context 65,536 tokens, up to 50 tool-call turns), producing DIVE-8B (SFT). (b) RL: From a separate pool of 38k tasks, we select 3.2k frontier tasks (1–5 successes in pass@8 self-sampling) and train with GRPO (Shao et al., 2024) (100 steps, batch size 512, learning rate 5e-6, max context 131,072 tokens, up to 100 tool-call turns), producing DIVE-8B (RL). Benchmark suites. We evaluate DIVE across three tiers: in-domain (L1) and two OOD settings distinguished by tool pool: general-purpose tools (L2) vs. specialized toolsets (L3). See Table 1 for details. •L1 (In-distribution Tasks): 800 tasks with disjoint seed concepts but the same 373-tool pool. Each task samples a novel tool subset (15–50 tools). • L2 (OOD Tasks w/ General Tools): Benchmarks us- ing Search/Browse/Code Execution. This includes Gen- eral DeepResearch tasks (GAIA (Mialon et al., 2023), HLE (Phan et al., 2025), BROWSECOMP (Wei et al., 2025), XBENCH (Chen et al., 2025)) and Domain Deep- Research tasks (FINSEARCHCOMP (Hu et al., 2025)). We use the 103-sample text-only validation subset for GAIA and the DeepSearch subset for XBENCH. •L3 (OOD Tasks w/ Specialized Tools): Benchmarks re- quiring specialized toolsets, including FINANCE AGENT BENCHMARK (public validation set) (Choi et al., 2025) (financial APIs), MEDAGENTBENCH (Jiang et al., 2025) (EHR system), SWE-BENCH VERIFIED (Jimenez et al., 2023) (containerized codebase interaction), and TOOLATHLON (Li et al., 2025b) (diverse MCP toolsets). 5 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use 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. Tier Task FamilyBenchmarkOOD FactorsTool PoolToolset ProtocolEnv L1In-DistributionDIVE-EVAL–384 Tools (General + Expert)Per-task OpenAI Stateless L2 General DeepResearch GAIA, HLE, BROWSECOMP, XBENCH-DS Task, SetSearch / BrowseUniform OpenAI Stateless Domain DeepResearch FINSEARCHCOMP (Global)Task, SetSearch / Browse / Code Execution Uniform OpenAI Stateless L3 Financial SpecialistFINANCE AGENT BENCHMARKTask, Pool, SetEDGAR / Web / Parse / RetrieveUniform OpenAI Stateless Medical SpecialistMEDAGENTBENCHTask, Pool, Set, Proto, EnvFHIR GET / POST / FinishUniformHTTPStateful Software EngineeringSWE-BENCH VERIFIEDTask, Pool, Set, EnvBash / Search / Editor / FinishUniform OpenAIStateful Zero-Shot GeneralistTOOLATHLONTask, Pool, Set, Env604 Tools (32 MCP Apps)Per-task OpenAIStateful 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. L1 In-distributionL2 OOD w/ General ToolsL3 OOD w/ Specialized Tools CategoryModelDIVE-EvalGAIAHLEBCXB-DSFSC 2 FSC 3 FABMABSWEToolathlon Frontier (≫8B) Gemini-3-Pro45.380.342.949.076.070.652.439.074.876.236.4 Claude-4-Sonnet44.863.720.812.862.260.233.339.079.372.729.9 Gemini-2.5-Pro 29.160.228.49.956.044.527.424.065.159.610.5 DeepSeek-V3.2-Exp40.461.017.940.167.261.327.426.067.367.820.1 Kimi-K2-0905 32.960.026.914.161.047.110.728.061.269.213.0 GPT-OSS-120B40.566.019.027.069.561.022.034.064.362.09.8 8B Baselines WebExplorer-8B19.150.017.315.753.735.918.14.017.87.00.3 SWE-Dev-8B13.823.26.91.631.630.53.63.014.219.50.0 EnvScaler-8B 15.425.82.81.745.740.710.814.056.611.52.2 Ours Qwen3-8B (base)13.022.46.41.324.028.67.12.038.410.80.9 DIVE-8B (SFT) 35.449.313.812.950.262.133.028.050.213.24.7 DIVE-8B (RL)42.561.217.816.458.167.337.334.057.318.38.3 OOD Factors. We categorize distribution shifts into five dimensions relative to DIVE’s training data (Table 1). Task distribution: the variety of user instructions; a shift means evaluation on tasks not synthesized by DIVE (e.g., human- curated benchmarks). Tool pool: the benchmark-level uni- verse of tool types; a shift involves unseen tools. Toolset: the task-level subset of tools; a shift tests unseen tool com- binations. Protocol: the invocation interface (e.g., function- calling vs. raw HTTP); a shift requires adapting to different schemas. Environment: the execution substrate; a shift introduces stateful dynamics (e.g., Docker containers). Baselines. We compare DIVE-8B against two categories of models: (i) 8B baselines (same backbone trained on other synthesized data), including specialized models for specific tasks/tools (WebExplorer-8B (for general Deep- Research) (Liu et al., 2025b), SWE-Dev-8B trained on the open-source SWE-Dev dataset (Wang et al., 2025a)) and generalizable models via query-first synthesis in simu- lated environments (EnvScaler-8B (Song et al., 2026)); (i) Frontier models (≫8B), including Gemini-3-Pro (Google DeepMind, 2025), Claude-4-Sonnet (Anthropic, 2025), Gemini-2.5-Pro (Comanici et al., 2025), DeepSeek-V3.2- Exp (DeepSeek-AI, 2025), Kimi-K2-0905 (Team et al., 2025), and GPT-OSS-120B (Agarwal et al., 2025). For all models, we evaluate with temperature = 1 and top-푝 = 1. 4.2. Main Results Table 2 presents the main evaluation results. In this sec- tion, we discuss the effectiveness of DIVE across diverse benchmarks and highlight the following observations: DIVE Delivers Robust and Substantial Generalization. It improves on in-distribution tasks (L1) and transfers con- sistently to all OOD benchmarks spanning both general and specialized toolsets (Table 1). Across the 9 OOD benchmarks, DIVE improves by +16.2 (SFT) and +22.2 (RL) points per benchmark on average, outperforming the strongest 8B baseline by +68% (Table 2). Despite its small backbone, DIVE is competitive with much larger models on deep-research and on challenging specialized benchmarks (e.g., FAB and MAB; Table 2). Notably, TOOLATHLON is a stringent zero-shot benchmark with per-task MCP app toolsets and stateful container environments, where DIVE improves from near-zero to 8.3 points, approaching GPT- OSS-120B and Gemini-2.5-Pro. DIVE Wins by Generalization, Even Against Specialists. Without task-specific training, DIVE matches or surpasses specialist agents on their home benchmarks (e.g., GAIA: 61.2 vs. 50.0 for WebExplorer-8B). In contrast, these spe- cialists transfer poorly under unseen shifts, often exhibiting negative transfer (e.g., WebExplorer-8B drops 5.8 points be- 6 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use 20 30 Success General DR 20 40 Expert DR 0 10 20 Success Finance (FAB) 20 40 Medical (MAB) S1S2S3S4 5 10 Success SWE-bench S1S2S3S4 0 2 Toolathlon BaseDiv-onlyQty-only (a) Diversity-only vs. Quantity-only 20 30 Success General DR 20 40 Expert DR 0 20 Success Finance (FAB) 40 50 Medical (MAB) 12k24k36k48k 10 14 Success SWE-bench 12k24k36k48k 2 4 Toolathlon BaseVar-onlyPoolExp+Var (b) Variety-only vs. Pool-Exp+Variety 20 40 Success General DR 20 40 Expert DR 0 20 Success Finance (FAB) 40 50 Medical (MAB) S1S2S3S4 10 15 Success SWE-bench S1S2S3S4 2.5 5.0 7.5 Toolathlon BaseSFTRL (c) All-Path Scaling: SFT→ RL Figure 3. Scaling analysis. Gray dashed line: Qwen3-8B base. Left: Diversity-only vs. Quantity-only. Diversity-only expands the tool pool from 1→4 domains (12k fixed; representative path fin→fin+med→fin+med+bio→all). Quantity-only scales data 12k→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→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→RL). 24 domain-expansion permutations; SFT 12k→48k and RL 0.8k→3.2k. Thin: paths; thick: mean; shaded: interquartile range; RL amplifies scaling. low the base model on L3 benchmarks). Compared to other generalization-oriented synthesis baselines (e.g., EnvScaler- 8B), DIVE achieves a 3.2×larger OOD lift, validating evidence-first synthesis on diverse, real tools. 5. Analysis Our main results show that DIVE generalizes broadly across shifts in tasks and toolsets. We now ask what drives these gains and how they scale. In our method, we controllably scale tool-pool coverage and toolset variety, and further induce richer tool-use patterns through our synthesis loop (cf. §3.3). In §5.1, we systematically study how these con- trollable diversity axes shape the scaling trend of OOD generalization; in §5.2, we further analyze the structural diversity induced by DIVE to explain this trend. 5.1. Scaling Analysis In this section, we systematically study the scaling trend of OOD generalization under tool shifts by controllably vary- ing tool-pool coverage and toolset variety (Fig. 3). Specifi- cally, we study the following three questions: How Necessary and Efficient Is Diversity Scaling for Generalizable Tool Use? We run an extreme scaling com- parison under matched synthesis and SFT settings (Fig. 3a). Diversity-only holds the total data budget fixed (12k) and scales tool-pool coverage from 1→4 domains, inducing richer tool-use patterns. Quantity-only holds the task and tool distribution fixed (Gen-DR: deep-research tasks syn- thesized with fixed Search/Browse tools) and scales data quantity from 12k→48k. The contrast is stark: diversity scaling yields consistent and stronger OOD gains, whereas Quantity-only scaling mainly improves in-distribution rou- tines. Even with more data, quantity-only cannot close the generalization gap from missing diversity, and at larger scale can further widen it on most OOD benchmarks. No- tably, with 4×less data (12k vs. 48k), Diversity-only still consistently outperforms Quantity-only across benchmarks, helping explain the gap to narrow specialist baselines. How Should We Scale Diversity for Faster Gains and a Higher Generalization Ceiling? We next ablate how to scale diversity in synthesis (Fig. 3b) with matched SFT scal- ing (12k→48k) starting from Finance. Toolset-variety-only keeps the Finance tool pool fixed; scaling data increases distinct toolset variants, but the pool’s tool types stay con- stant. Pool-Expansion+Variety similarly increases toolset variants, while also expanding the pool across domains to introduce new tool types. Our statistics show toolset- variant growth is similar under both routes; thus the key difference is whether new tool types/capabilities enter the pool. Empirically (Fig. 3b), both improve generalization, but toolset-variety-only yields limited gains and exhibits faster saturation, while Pool-Expansion+Variety yields faster gains and a higher, slower-saturating ceiling as the pool grows—making it a stronger scaling strategy. 7 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use 0255075100 Steps 0.3 0.4 0.5 0.6 Accuracy +32% 0255075100 Steps 7.5 10.0 12.5 Calls / Task +35% 0255075100 Steps 300 400 Tool-call Graphs +20% 0255075100 Steps 100 150 R/P Topologies +55% 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. Does Exploration (RL) Further Amplify the Diversity- Scaling Trend Beyond Imitation (SFT)? To robustly evaluate this trend, we evaluate SFT and RL across all 24 scaling paths (Fig. 3c; Table 13). The trend is already visible under SFT, indicating the model can imitate diverse tool-use patterns in expert trajectories. RL amplifies this diversity- scaling trend, suggesting exploration beyond imitation: the RL–SFT gap grows with diversity (Avg RL –Avg SFT : +4.6 at 1 domain vs. +5.6 at 4 domains), and at 4 domains the mean rises from 29.1→34.8 with narrow interquartile bands. In §5.2, we quantify structural diversity induced by DIVE to validate and help explain this mechanism. 5.2. Structural Diversity Analysis At this stage, we analyze the structural diversity of tool use in DIVE’s SFT trajectories and RL roll- outs to explain the diversity-scaling trend in OOD generalization.We measure diversity at three lev- els: (i) tool-pool coverage (tool types exercised, e.g., tools coveredin the dataset anddistinct tool types per task); (i) toolset variety (unique toolsets across tasks); and (i) tool-use patterns, including unique tool-call se- quences, unique tool-call graphs capturing tool–reasoning dependencies (inferred by Claude-4-Sonnet), and ab- stractRetrieval/Processing (R/P) topologies(222-class tax- onomy; Appendix D). We compare DIVE’s 48k SFT trajec- tories against Gen-DR (Table 3, Figure 5), then track how these patterns evolve during RL (Figure 4). SFT Data: Structural Diversity. To interpret §5.1, we inspect SFT trajectories: DIVE exhibits higher diversity than Gen-DR in tool coverage, toolset variety, and tool-use patterns (Table 3). Figure 5 highlights two key differences: in topology space (left), Gen-DR concentrates in retrieval- only (PureR), while DIVE shifts mass to mixed (R+P) and processing-only (PureP) and covers more topology classes; in tool usage (right), DIVE exhibits a long-tail frequency over a broad tool pool. These shifts match the OOD setting, where tasks often require retrieval–processing composition and specialized tools. Further, we analyze how synthesis- loop iterations increase diversity in Appendix C. RL Amplifies Diversity. During RL, accuracy improves while structural diversity continues to increase (Figure 4). Table 3. Diversity comparison: Gen-DR vs. DIVE (48k each; same teacher model, rejection sampling). Diversity MetricGen-DRDIVEΔ Tools covered2373+186× Unique toolsets146,398+46k× Unique tool-call sequences1,23125,084+20× Unique tool-call graphs19,44239,810+105% Unique R/P topologies12,31523,450+90% R/P topology classes covered65153+135% Avg. tool calls per task15.2111.11-27% Distinct tool types per task1.713.26+91% Avg. score after SFT22.5132.15+43% 165123153 R/P topology class 0.00 0.03 0.05 0.08 Density PureRR+PPureP Gen-DR DIVE 1100200300373 Tool rank 10 1 10 3 10 5 Count General Biology Academic Medical Financial Figure 5. R/P topology density and tool-frequency distri- butions (48k SFT). Left: Density over R/P topology classes (153 observed;retrieval-only→mixed→processing-only, i.e., PureR→R+P→PureP; taxonomy in Appendix D). Right: Tool- call frequency over 373 tools (5 domains). Over 100 RL iterations, reward improves while diverse tool- use patterns persist and expand (e.g., unique tool-call graphs and R/P topologies). See Appendix Figure 6 for a per- domain breakdown of these dynamics. Together with §5.1, this supports our hypothesis that RL amplifies generalization by exploring and reinforcing a broader set of effective tool- use structures rather than collapsing to a single routine. Conclusion We presented DIVE, an execution-first framework for train- ing tool-using agents on real, diverse toolsets with built-in executability and verifiability. By inverting synthesis (evi- dence first, tasks derived from traces), DIVE scales diversity while keeping supervision grounded. Across three bench- mark tiers, DIVE improves OOD generalization, and scaling studies show tool-pool diversity matters more than data quantity. Structural analyses reveal richer tool-use patterns (sequences, graphs, R/P topologies), a trend visible under SFT and amplified by RL. 8 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Impact Statement This work studies data synthesis and post-training for tool- using language agents. By generating grounded tasks and training on diverse, real-world tools, we aim to expand the supply of diverse agentic training data, improve general- ization across tasks and toolsets, and enable more reliable agent evaluation. References Agarwal, S., Ahmad, L., Ai, J., Altman, S., Applebaum, A., Arbus, E., Arora, R. K., Bai, Y., Baker, B., Bao, H., et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925, 2025. Anthropic.Introducing claude 4.https://w. anthropic.com/news/claude-4, 2025. Arora, R. K., Wei, J., Hicks, R. S., Bowman, P., Quiñonero- Candela, J., Tsimpourlas, F., Sharman, M., Shah, M., Vallone, A., Beutel, A., et al. Healthbench: Evaluating large language models towards improved human health. arXiv preprint arXiv:2505.08775, 2025. Bridge, A.-M. Wikipedia, the free encyclopedia. San Fran- cisco (CA): Wikimedia Foundation, 2001. Castellani, T., Ye, N., Mittal, D., Yen, T., and Namkoong, H. Synthtools: A framework for scaling synthetic tools for agent development. arXiv preprint arXiv:2511.09572, 2025. Chen, K., Ren, Y., Liu, Y., Hu, X., Tian, H., Xie, T., Liu, F., Zhang, H., Liu, H., Gong, Y., et al. xbench: Track- ing agents productivity scaling with profession-aligned real-world evaluations. arXiv preprint arXiv:2506.13651, 2025. Chen, Z., Chen, W., Smiley, C., Shah, S., Borova, I., Lang- don, D., Moussa, R., Beane, M., Huang, T.-H., Chiang, B., et al. Finqa: A dataset of numerical reasoning over financial data. arXiv preprint arXiv:2109.00122, 2021. Choi, C., Kwon, J., Lopez-Lira, A., Kim, C., Kim, M., Hwang, J., Ha, J., Choi, H., Yun, S., Kim, Y., et al. Fi- nagentbench: A benchmark dataset for agentic retrieval in financial question answering. In Proceedings of the 6th ACM International Conference on AI in Finance, p. 632–637, 2025. Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. DeepSeek-AI. Introducing deepseek-v3.2-exp.https:// api-docs.deepseek.com/news/news250929, 2025. Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H., and Su, Y. Mind2web: Towards a general- ist agent for the web. Advances in Neural Information Processing Systems, 36:22485–22517, 2023. Fang, R., Cai, S., Li, B., Wu, J., Li, G., Yin, W., Wang, X., Wang, X., Su, L., Zhang, Z., et al. Towards general agen- tic intelligence via environment scaling. arXiv preprint arXiv:2509.13311, 2025. Froger, R., Andrews, P., Bettini, M., Budhiraja, A., Cabral, R. S., Do, V., Garreau, E., Gaya, J.-B., Laurençon, H., Lecanu, M., et al. Are: Scaling up agent environments and evaluations. arXiv preprint arXiv:2509.17158, 2025. Fu, D., He, K., Wang, Y., Hong, W., Gongque, Z., Zeng, W., Wang, W., Wang, J., Cai, X., and Xu, W. Agentrefine: En- hancing agent generalization through refinement tuning. arXiv preprint arXiv:2501.01702, 2025. Google DeepMind. Introducing gemini 3 pro.https:// deepmind.google/technologies/gemini/, 2025. Gudibande, A., Wallace, E., Snell, C. V., Geng, X., Liu, H., Abbeel, P., Levine, S., and Song, D. The false promise of imitating proprietary language models. In The Twelfth International Conference on Learning Representations, 2024. Guo, Z., Xu, B., Zhu, C., Hong, W., Wang, X., and Mao, Z. Mcp-agentbench: Evaluating real-world language agent performance with mcp-mediated tools. arXiv preprint arXiv:2509.09734, 2025. He, J., Neville, J., Wan, M., Yang, L., Liu, H., Xu, X., Song, X., Pan, J. Z., and Zhou, P. Gentool: Enhanc- ing tool generalization in language models through zero- to-one and weak-to-strong simulation. arXiv preprint arXiv:2502.18990, 2025. Hu, L., Jiao, J., Liu, J., Ren, Y., Wen, Z., Zhang, K., Zhang, X., Gao, X., He, T., Hu, F., et al. Finsearchcomp: Towards a realistic, expert-level evaluation of financial search and reasoning. arXiv preprint arXiv:2509.13160, 2025. Jiang, Y., Black, K. C., Geng, G., Park, D., Zou, J., Ng, A. Y., and Chen, J. H. Medagentbench: a virtual ehr environment to benchmark medical llm agents. Nejm Ai, 2(9):AIdbp2500144, 2025. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., and Narasimhan, K. Swe-bench: Can language mod- els resolve real-world github issues?arXiv preprint arXiv:2310.06770, 2023. 9 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Jin, Q., Dhingra, B., Liu, Z., Cohen, W. W., and Zhu, X. Pubmedqa: A dataset for biomedical research question answering. arXiv preprint arXiv:1909.06146, 2019. Jing, L., Huang, Z., Wang, X., Yao, W., Yu, W., Ma, K., Zhang, H., Du, X., and Yu, D. Dsbench: How far are data science agents from becoming data science experts? arXiv preprint arXiv:2409.07703, 2024. Krishna, S., Krishna, K., Mohananey, A., Schwarcz, S., Stambler, A., Upadhyay, S., and Faruqui, M. Fact, fetch, and reason: A unified evaluation of retrieval-augmented generation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), p. 4745–4759, 2025. Li, H., Chen, J., Yang, J., Ai, Q., Jia, W., Liu, Y., Lin, K., Wu, Y., Yuan, G., Hu, Y., et al. Legalagentbench: Evaluat- ing llm agents in legal domain. In Proceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), p. 2322–2344, 2025a. Li, J., Zhao, W., Zhao, J., Zeng, W., Wu, H., Wang, X., Ge, R., Cao, Y., Huang, Y., Liu, W., et al. The tool decathlon: Benchmarking language agents for diverse, realistic, and long-horizon task execution. arXiv preprint arXiv:2510.25726, 2025b. Li, K., Zhang, Z., Yin, H., Ye, R., Zhao, Y., Zhang, L., Ou, L., Zhang, D., Wu, X., Wu, J., et al. Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning. arXiv preprint arXiv:2509.13305, 2025c. Li, K., Zhang, Z., Yin, H., Zhang, L., Ou, L., Wu, J., Yin, W., Li, B., Tao, Z., Wang, X., et al. Websailor: Navigating super-human reasoning for web agent. arXiv preprint arXiv:2507.02592, 2025d. Li, M., Song, F., Yu, B., Yu, H., Li, Z., Huang, F., and Li, Y. Api-bank: A comprehensive benchmark for tool- augmented llms. arXiv preprint arXiv:2304.08244, 2023. Li, Y., Inan, H. A., Yue, X., Chen, W.-N., Wutschitz, L., Kulkarni, J., Poovendran, R., Sim, R., and Rajmohan, S. Simulating environments with reasoning models for agent training. arXiv preprint arXiv:2511.01824, 2025e. Liu, A., Mei, A., Lin, B., Xue, B., Wang, B., Xu, B., Wu, B., Zhang, B., Lin, C., Dong, C., et al. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556, 2025a. Liu, J., Li, Y., Zhang, C., Li, J., Chen, A., Ji, K., Cheng, W., Wu, Z., Du, C., Xu, Q., et al. Webexplorer: Explore and evolve for training long-horizon web agents. arXiv preprint arXiv:2509.06501, 2025b. Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., et al. Agentbench: Evalu- ating llms as agents. arXiv preprint arXiv:2308.03688, 2023. Mialon, G., Fourrier, C., Wolf, T., LeCun, Y., and Scialom, T. Gaia: a benchmark for general ai assistants. In The Twelfth International Conference on Learning Represen- tations, 2023. Mitchener, L., Laurent, J. M., Andonian, A., Tenmann, B., Narayanan, S., Wellawatte, G. P., White, A., Sani, L., and Rodriques, S. G. Bixbench: a comprehensive benchmark for llm-based agents in computational biology. arXiv preprint arXiv:2503.00096, 2025. Mitra, A., Del Corro, L., Zheng, G., Mahajan, S., Rouhana, D., Codas, A., Lu, Y., Chen, W.-g., Vrousgos, O., Rosset, C., et al. Agentinstruct: Toward generative teaching with agentic flows. arXiv preprint arXiv:2407.03502, 2024. National Library of Medicine (US). Pubmed, 2026. URL https://pubmed.ncbi.nlm.nih.gov/.Ac- cessed: 2026-01-29. Phan, L., Gatti, A., Han, Z., Li, N., Hu, J., Zhang, H., Zhang, C. B. C., Shaaban, M., Ling, J., Shi, S., et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249, 2025. Qiao, Z., Chen, G., Chen, X., Yu, D., Yin, W., Wang, X., Zhang, Z., Li, B., Yin, H., Li, K., et al. Webre- searcher: Unleashing unbounded reasoning capability in long-horizon agents. arXiv preprint arXiv:2509.13309, 2025. Qin, Y., Liang, S., Ye, Y., Zhu, K., Yan, L., Lu, Y., Lin, Y., Cong, X., Tang, X., Qian, B., et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. Sayers, E. W., Beck, J., Bolton, E. E., Brister, J. R., Chan, J., Connor, R., Feldgarden, M., Fine, A. M., Funk, K., Hoffman, J., et al. Database resources of the national center for biotechnology information in 2025. Nucleic acids research, 53(D1):D20–D29, 2025. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al. Deepseekmath: Push- ing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. Shen, Y., Song, K., Tan, X., Zhang, W., Ren, K., Yuan, S., Lu, W., Li, D., and Zhuang, Y. Taskbench: Benchmarking large language models for task automation. Advances in Neural Information Processing Systems, 37:4540–4574, 2024. 10 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Song, X., Chang, H., Dong, G., Zhu, Y., Dou, Z., and Wen, J.-R. Envscaler: Scaling tool-interactive environments for llm agent via programmatic synthesis. arXiv preprint arXiv:2601.05808, 2026. Tang, Q., Deng, Z., Lin, H., Han, X., Liang, Q., Cao, B., and Sun, L. Toolalpaca: Generalized tool learning for language models with 3000 simulated cases, 2023. URL https://arxiv.org/abs/2306.05301. Team, K., Bai, Y., Bao, Y., Chen, G., Chen, J., Chen, N., Chen, R., Chen, Y., Chen, Y., Chen, Y., et al. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534, 2025. Tian, M., Chen, L., Zhu, S., et al. Scicode: A research coding benchmark curated by scientists. arXiv preprint arXiv:2407.13168, 2024. Tsatsaronis, G., Balikas, G., Malakasiotis, P., Partalas, I., Zschunke, M., Alvers, M. R., Weissenborn, D., Krithara, A., Petridis, S., Polychronopoulos, D., et al. An overview of the bioasq large-scale biomedical semantic indexing and question answering competition. BMC bioinformat- ics, 16:1–28, 2015. Wang, H., Hou, Z., Wei, Y., Tang, J., and Dong, Y. Swe- dev: Building software engineering agents with training and inference scaling. arXiv preprint arXiv:2506.07636, 2025a. Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N. A., Khashabi, D., and Hajishirzi, H. Self-instruct: Aligning language models with self-generated instructions. In Pro- ceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 13484–13508, 2023. Wang, Z., Chang, Q., Patel, H., Biju, S., Wu, C.-E., Liu, Q., Ding, A., Rezazadeh, A., Shah, A., Bao, Y., et al. Mcp-bench: Benchmarking tool-using llm agents with complex real-world tasks via mcp servers. arXiv preprint arXiv:2508.20453, 2025b. Wang, Z., Liang, Y., Zhang, X., Wu, Q., Han, S., Bastos, A., Wang, R., Bansal, C., Peng, B., Gao, J., et al. Adapting web agents with synthetic supervision. arXiv preprint arXiv:2511.06101, 2025c. Wei, J., Sun, Z., Papay, S., McKinney, S., Han, J., Fulford, I., Chung, H. W., Passos, A. T., Fedus, W., and Glaese, A. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv preprint arXiv:2504.12516, 2025. Wu, J., Li, B., Fang, R., Yin, W., Zhang, L., Tao, Z., Zhang, D., Xi, Z., Fu, G., Jiang, Y., et al. Webdancer: Towards autonomous information seeking agency. arXiv preprint arXiv:2505.22648, 2025a. Wu, J., Yin, W., Jiang, Y., Wang, Z., Xi, Z., Fang, R., Zhang, L., He, Y., Zhou, D., Xie, P., et al. Webwalker: Benchmarking llms in web traversal. arXiv preprint arXiv:2501.07572, 2025b. Xie, J., Zhang, K., Chen, J., Zhu, T., Lou, R., Tian, Y., Xiao, Y., and Su, Y. Travelplanner: A benchmark for real-world planning with language agents. In Forty-first International Conference on Machine Learning, 2024. Xu, F. F., Song, Y., Li, B., Tang, Y., Jain, K., Bao, M., Wang, Z. Z., Zhou, X., Guo, Z., Cao, M., et al. Theagentcom- pany: benchmarking llm agents on consequential real world tasks. arXiv preprint arXiv:2412.14161, 2024. Yahoo Finance. Yahoo finance market data, 2026. URL https://finance.yahoo.com/. Accessed: 2026- 01-15. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Yang, Z., Qi, P., Zhang, S., Bengio, Y., Cohen, W. W., Salakhutdinov, R., and Manning, C. D. Hotpotqa: A dataset for diverse, explainable multi-hop question an- swering. arXiv preprint arXiv:1809.09600, 2018. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. ReAct: Synergizing reasoning and act- ing in language models. In International Conference on Learning Representations (ICLR), 2023. Yao, S., Shinn, N., Razavi, P., and Narasimhan, K.휏-bench: A benchmark for tool-agent-user interaction in real-world domains. arXiv preprint arXiv:2406.12045, 2024. Zhang, M., Yang, Y., Xie, R., Dhingra, B., Zhou, S., and Pei, J. Generalizability of large language model- based agents: A comprehensive survey. arXiv preprint arXiv:2509.16330, 2025. Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Bisk, Y., Fried, D., Alon, U., et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. Zhu, F., Lei, W., You, C., Gu, S., Kuang, Z., Feng, Z., Zhang, L., Wu, T., Deng, X., and Chen, Y. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624, 2021. 11 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use A. Synthesized Task Examples We present representative examples of synthesized tasks from each domain to illustrate the complexity and diversity achieved by DIVE. Each task requires multi-step reasoning across multiple tools selected from a domain-specific candidate set. In the examples below, the Tools field lists the full candidate toolset available for the task, wheregreen tagsindicate tools effectively used by the agent andgray tagsindicate available but unused tools. This visualization highlights the agent’s ability to precisely select relevant tools from a noisy candidate set to solve complex queries. The Stats line provides quantitative metrics: Calls (total tool calls), Available (size of toolset), and Unique (count of distinct tools used). A.1. Academic Domain Academic Case: Cross-Journal Researcher Tracking Tools:arxiv_advanced_searcharxiv_get_papers_by_idsarxiv_search_by_authorarxiv_search_by_category arxiv_search_by_date_rangearxiv_search_paperscrossref_funder_workscrossref_funders crossref_get_journalcrossref_get_workcrossref_journal_workscrossref_journalscrossref_licenses crossref_member_workscrossref_work_agencyopenalex_get_authoropenalex_get_institution openalex_get_sourceopenalex_get_workopenalex_search_authorsopenalex_search_funders openalex_search_institutionsopenalex_search_publishersopenalex_search_sourcesopenalex_search_topics openalex_search_works Stats: Calls: 50Available: 26Unique: 9 Query: A Stanford University computer science researcher with an ORCID identifier ending in 3426 has exactly 260 publications cited at least 10 times each, and their most highly cited work was published in 1981. This researcher’s total citation count exceeds 65,000 but is less than 66,000, and they have been affiliated with Stanford University continuously from 2014 through 2023. Among their publications from 2020-2023, what is the title of their work that appears in “ACM Transactions on Management Information Systems” and has been cited more than 40 times? Answer: PANDA: Partitioned Data Security on Outsourced Sensitive and Non-sensitive Data. A.2. Biological Domain Biological Case: Enzyme Characterization with Taxonomic Constraints Tools:bio_seq_transcribebio_sequtils_gc_skewbio_sequtils_nt_searchbio_sequtils_seq1 cluster_treeclustercodon_table_standardexpasy_prodocexpasy_prosite_rawncbi_entrez_fetch ncbi_entrez_searchncbi_entrez_summarypairwise2_globalxxpairwise2_localxxprotparam_aromaticity restriction_all_enzymesseqfeature_locationsvd_superimposetogows_converttogows_entrytogows_search Stats: Calls: 47Available: 20Unique: 6 Query: I need the UniProt accession number for a thermostable cutinase that meets these interconnected criteria: (1) It originates from a bacterial species containing “fusca” in its scientific name and belongs to the same taxonomic family as the organism that produces BTA-hydrolase 2, (2) It has an alternative name containing “BTA-hydrolase” with the same numerical designation as the number of distinct BTA-hydrolase variants found in the NCBI protein database for this species, (3) It can degrade PET with EC number 3.1.1.101 and shares >95% sequence identity with CAH17553, (4) Its UniProt entry was most recently modified in the same month as the Zeikus thermozyme review publication but in 2025, (5) The protein must have the exact same number of amino acid residues as BTA-hydrolase 2 from the same organism, and (6) Its gene has a synonym that matches the primary gene name of its homolog minus one character. Answer: Q6A0I4 12 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use A.3. Financial Domain Financial Case: Cross-Market Portfolio Analysis Tools:browsecb_callcb_shareetf_dailyetf_indexfilm_recordfund_navfund_portfoliofut_settle fut_weekly_detailindex_sw_dailyjupyter_execute_code_cellparallel_searchsearchstock_balancesheet stock_ccas_holdstock_ccas_hold_detailstock_disclosure_datestock_dividendstock_em_hot stock_hsgt_top10stock_kpl_topicstock_limit_boardstock_managersstock_stk_auctionstock_suspend stock_tdx_indextv_record Stats: Calls: 47Available: 28Unique: 8 Query: Among the fund holdings in portfolio 001753.OF as of December 31, 2023, identify the healthcare/medical diagnostics sector stock with the highest market value that also has a stock float ratio below 0.03 and was established before 2000. Then, cross-reference this with the Shanghai-Shenzhen-Hong Kong Stock Connect top 10 trading data for December 20, 2024, to find if any semiconductor industry stock with a trading amount exceeding 1.1 billion yuan shares the same first three digits in its stock code and has “innovation” in its Chinese company name. Additionally, verify that both companies are listed on the same exchange (Shanghai vs Shenzhen) and that the semiconductor company’s English name contains “Semiconductor”. What is the stock code of the fund holding that meets these industry-specific criteria, what is the English company name of its matching trading counterpart, and what year was the healthcare company established? Answer: Stock code: 603658.SH; Chinese company name: GigaDevice Semiconductor (Beijing) Inc.; Healthcare company established: 1998 13 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use A.4. Medical Domain Medical Case: Contraindication and Drug Classification Filtering Tools:dbvar_germline_datagenetic_diseaseshcpcs_procedure_codesicd10cm_diagnosis_codes icd9cm_diagnosis_codespharmvar_star_allelesprescribable_find_rxcui_by_id prescribable_find_rxcui_by_stringprescribable_get_all_concepts_by_ttyprescribable_get_drugs prescribable_get_prop_categoriesprescribable_get_related_by_typeprescribable_get_rxnorm_name prescribable_get_source_typesprescribable_get_spelling_suggestionsprescribable_get_term_types rxclass_find_similar_classes_by_drug_listrxclass_get_all_classesrxclass_get_class_by_rxnorm_drug_name rxclass_get_class_contextsrxclass_get_class_membersrxclass_get_class_tree rxclass_get_rela_source_versionrxclass_get_relasrxclass_get_sources_of_drug_class_relations rxclass_get_spelling_suggestionsrxnorm_find_related_ndcsrxnorm_find_rxcui_by_id rxnorm_find_rxcui_by_stringrxnorm_get_all_propertiesrxnorm_get_all_related_inforxnorm_get_id_types rxnorm_get_ndc_statusrxnorm_get_proprietary_informationrxnorm_get_related_by_relationship rxnorm_get_rx_concept_propertiesrxnorm_get_rxcui_history_statusrxnorm_get_source_types rxnorm_get_term_typesrxterms_prescription_drugs Stats: Calls: 44Available: 40Unique: 14 Query: Among all drugs in the ATC therapeutic subgroup A10BK that share the same specific mechanism of action classification (MOA class N0000187058) and have contraindications with chronic kidney failure (MEDRT disease class D007676), identify which drug has UNII code 6C282481IP. For this drug: (1) calculate the numerical difference between its two available tablet strengths, (2) determine how many other active ingredients in the same MOA class have anhydrous or propanediol salt forms available in RxNorm, and (3) among the drugs in this MOA class that also have contraindications with chronic kidney failure, identify which one has the most brand name variations and state the total count of those brand names. Answer: Ertugliflozin (UNII 6C282481IP). (1) 10 MG (15 MG− 5 MG); (2) 2 other ingredients: canagliflozin anhydrous, dapagliflozin propanediol; (3) Empagliflozin with 4 brand names: Jardiance, Glyxambi, Synjardy, Trijardy. B. Data Synthesis Details B.1. Tool Pool Details We curate a diverse tool pool spanning five domains with both Retrieval (R) and Processing (P) primitives. Retrieval tools fetch data from external sources (APIs, databases) without significant transformation. Processing tools perform calculations, analysis, or data transformations (e.g., technical indicators, sequence alignment, similarity matching). Table 4. Tool pool summary. Distribution of Retrieval (R: information fetching) and Processing (P: computation/transformation) tools across domains. DomainRetrievalProcessingTotalP% Financial155141698.3% Medical8098910.1% Academic4375014.0% Biological18436170.5% General31425.0% Total2997437319.8% 14 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Table 5. Tool sources and documentation. Official documentation or API endpoints for the tools used in DIVE. DomainSource NameURL / Documentation FinancialTushare 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://w.semanticscholar.org/product/api OpenAlex https://docs.openalex.org/ Crossref https://w.crossref.org/documentation/ retrieve-metadata/rest-api/ arXiv https://arxiv.org/help/api Biological NCBI Entrez https://w.ncbi.nlm.nih.gov/books/NBK25501/ BioPython https://biopython.org/ TogoWS http://togows.dbcls.jp/ General Google Search https://w.google.com/ Python SandboxCustom Implementation (based on standard library) The complete list of all 373 tools is provided in Table 6. Table 6. Complete tool pool (373 tools). Aggregated by category. R/P denotes Retrieval (R) or Processing (P) tool type. CategoryDomainR/PCountTools Conv. BondsFinancialP1cb_factor_pro ETF DataFinancialP1etf_adj Mutual FundsFinancialP1fund_factor IndicesFinancialP1index_factor Stock MetricsFinancialP10stock_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/RatesFinancialR7bc_bestotcqt, bc_otcqt, libor, repo_daily, us_tbr, us_trycr, us_tycr Conv. BondsFinancialR6cb_basic, cb_call, cb_daily, cb_issue, cb_rate, cb_share ETF DataFinancialR3etf_basic, etf_daily, etf_index FX & HKFinancialR4fx_daily, hibor, hk_basic, hk_tradecal Mutual FundsFinancialR6fund_basic, fund_div, fund_manager, fund_nav, fund_portfolio, fund_share FuturesFinancialR9fut_basic, fut_daily, fut_holding, fut_limit, fut_mapping, fut_settle, fut_weekly_detail, fut_weekly_monthly, fut_wsr IndicesFinancialR12index_basic, index_ci_daily, index_ci_member, index_daily, index_daily_info, index_global, index_monthly, index_sw_daily, index_sw_member, index_sz_market, index_weekly, in- dex_weight Macro/OtherFinancialR13anns_d, cn_gdp, cn_m, cn_pmi, cn_ppi, film_record, news_cctv, sse_qa, szse_qa, tmt_twincome, tmt_twincome_detail, trade_cal, tv_record OptionsFinancialR2opt_basic, opt_daily 15 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use (continued from previous page) CategoryDomainR/PCountTools Stock DataFinancialR90stock_auction_close,stock_auction_open,stock_bak_daily,stock_balancesheet, stock_balancesheet_vip,stock_block_trade,stock_broker_forecast, stock_broker_recommend, stock_broker_recommend, stock_cashflow, stock_cashflow_vip, stock_ccas_hold, stock_ccas_hold_detail, stock_ccass_hold, stock_ccass_hold_detail, stock_company, stock_concept_detail, stock_daily, stock_dc_daily, stock_dc_index, stock_dc_member, stock_disclosure_date, stock_dividend, stock_em_hot, stock_express, stock_fina_mainbz, stock_forecast, stock_ggt_daily, stock_ggt_monthly, stock_ggt_top10, stock_hk_hold, stock_hm_detail, stock_hm_list, stock_holder_number, stock_hs_const, stock_hsgt_top10,stock_income,stock_income_vip,stock_index_member, stock_kpl_list, stock_kpl_topic, stock_lhb_detail, stock_limit_board, stock_limit_list_d, stock_limit_list_ths,stock_limit_step,stock_managers,stock_margin, stock_margin_detail, stock_market_money_flow, stock_money_flow, stock_moneyflow, stock_moneyflow_hsgt,stock_moneyflow_ths,stock_monthly,stock_namechange, stock_pledge_detail, stock_pledge_stat, stock_pro_bar, stock_report_rc, stock_repurchase, stock_sector_money_flow,stock_share_float,stock_slb_len_m,stock_slb_sec, stock_slb_sec_detail,stock_stk_auction,stock_stk_holdertrade,stock_stk_holds, stock_stk_limit, stock_stk_limit_pool, stock_stk_rewards, stock_stk_splits, stock_stk_surv, stock_stk_surv, stock_suspend, stock_suspend_d, stock_tdx_daily, stock_tdx_index, stock_tdx_member, stock_ths_daily, stock_ths_hot, stock_ths_index, stock_ths_member, stock_top10_floatholders,stock_top10_holders,stock_top_inst,stock_top_list, stock_trade_cal, stock_weekly Stock MetricsFinancialR5stock_bak_basic, stock_basic, stock_cyq_chips, stock_cyq_perc, stock_index_dailybasic MatchingMedicalP8prescribable_get_approximate_match,prescribable_get_spelling_suggestions,rx- class_find_similar_classes_by_class,rxclass_find_similar_classes_by_drug_list, rxclass_get_similarity_information,rxclass_get_spelling_suggestions, rxnorm_get_approximate_match, rxnorm_get_spelling_suggestions PrescribableMedicalP1prescribable_find_rxcui_by_string PrescribableMedicalR21prescribable_filter_by_property,prescribable_find_rxcui_by_id,prescrib- able_get_all_concepts_by_tty,prescribable_get_all_properties,prescrib- able_get_all_related_info,prescribable_get_display_terms,prescribable_get_drugs, prescribable_get_id_types, prescribable_get_multi_ingred_brand, prescribable_get_ndcs, pre- scribable_get_prop_categories, prescribable_get_prop_names, prescribable_get_rela_paths, prescribable_get_rela_types,prescribable_get_related_by_relationship,prescrib- able_get_related_by_type,prescribable_get_rx_concept_properties,prescrib- able_get_rx_property, prescribable_get_rxnorm_name, prescribable_get_source_types, prescribable_get_term_types ReferenceMedicalR14cytogenetic_chromosome_locations,dbvar_germline_data,genetic_diseases, hcpcs_procedure_codes, hugo_genes, icd10cm_diagnosis_codes, icd9cm_diagnosis_codes, pharmvar_star_alleles,reference_sequences,rxterms_get_all_concepts, rxterms_get_all_rxterm_info,rxterms_get_rxterm_display_name,rx- terms_get_rxterms_version, rxterms_prescription_drugs RxClassMedicalR13rxclass_find_class_by_name,rxclass_find_classes_by_id,rxclass_get_all_classes, rxclass_get_class_by_rxnorm_drug_id,rxclass_get_class_by_rxnorm_drug_name, rxclass_get_class_contexts, rxclass_get_class_graph_by_source, rxclass_get_class_members, rxclass_get_class_tree,rxclass_get_class_types,rxclass_get_rela_source_version, rxclass_get_relas, rxclass_get_sources_of_drug_class_relations RxNormMedicalR32rxnorm_filter_by_property,rxnorm_find_related_ndcs,rxnorm_find_rxcui_by_id, rxnorm_find_rxcui_by_string,rxnorm_get_all_concepts_by_status, rxnorm_get_all_concepts_by_tty,rxnorm_get_all_historical_ndcs, rxnorm_get_all_ndcs_by_status, rxnorm_get_all_properties, rxnorm_get_all_related_info, rxnorm_get_display_terms,rxnorm_get_drugs,rxnorm_get_id_types, rxnorm_get_multi_ingred_brand, rxnorm_get_ndc_properties, rxnorm_get_ndc_status, rxnorm_get_ndcs,rxnorm_get_prop_categories,rxnorm_get_prop_names, rxnorm_get_proprietary_information,rxnorm_get_reformulation_concepts, rxnorm_get_rela_paths,rxnorm_get_rela_types,rxnorm_get_related_by_relationship, rxnorm_get_related_by_type, rxnorm_get_rx_concept_properties, rxnorm_get_rx_property, rxnorm_get_rxcui_history_status, rxnorm_get_rxnorm_name, rxnorm_get_rxnorm_version, rxnorm_get_source_types, rxnorm_get_term_types 16 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use (continued from previous page) CategoryDomainR/PCountTools AnalysisAcademicP4openalex_analyze_text,semantic_scholar_paper_autocomplete,seman- tic_scholar_paper_recommendations, semantic_scholar_recommend_papers S. ScholarAcademicP3semantic_scholar_paper_search,semantic_scholar_paper_title_search,seman- tic_scholar_snippet_search CrossrefAcademicR15crossref_funder_works, crossref_funders, crossref_get_funder, crossref_get_journal, cross- ref_get_member, crossref_get_prefix, crossref_get_type, crossref_get_work, cross- ref_journal_works, crossref_journals, crossref_licenses, crossref_member_works, cross- ref_members, crossref_types, crossref_work_agency OpenAlexAcademicR11openalex_get_author, openalex_get_institution, openalex_get_source, openalex_get_work, openalex_search_authors,openalex_search_funders,openalex_search_institutions, openalex_search_publishers,openalex_search_sources,openalex_search_topics, openalex_search_works S. ScholarAcademicR11semantic_scholar_author_batch,semantic_scholar_author_papers,seman- tic_scholar_author_search, semantic_scholar_get_author, semantic_scholar_get_paper, semantic_scholar_paper_authors,semantic_scholar_paper_batch,seman- tic_scholar_paper_bulk_search,semantic_scholar_paper_citations,seman- tic_scholar_paper_references, semantic_scholar_release_list arXivAcademicR6arxiv_advanced_search,arxiv_get_papers_by_ids,arxiv_search_by_author, arxiv_search_by_category, arxiv_search_by_date_range, arxiv_search_papers AlignmentBiologicalP4pairwise2_global_align, pairwise2_globalxx, pairwise2_local_align, pairwise2_localxx ClusteringBiologicalP3cluster_distancematrix, cluster_pca, cluster_treecluster MotifsBiologicalP3motifs_create, motifs_reverse_complement, motifs_reverse_complement_rna NCBIBiologicalP2ncbi_entrez_ecitmatch, ncbi_entrez_espell OtherBiologicalP2svd_superimpose, togows_convert PDBBiologicalP4pdb_calc_angle, pdb_calc_dihedral, pdb_is_a, pdb_is_nucleic ProteinBiologicalP4protparam_analysis,protparam_aromaticity,protparam_isoelectric_point,prot- param_molecular_weight RestrictionBiologicalP2restriction_catalyse, restriction_search FeaturesBiologicalP2seqfeature_compound, seqfeature_location SequenceBiologicalP17bio_seq_back_transcribe,bio_seq_complement,bio_seq_complement_rna, bio_seq_count,bio_seq_find,bio_seq_pattern,bio_seq_reverse_complement, bio_seq_reverse_complement_rna,bio_seq_transcribe,bio_seq_translate, bio_sequtils_gc123, bio_sequtils_gc_content, bio_sequtils_gc_skew, bio_sequtils_nt_search, bio_sequtils_seq1, bio_sequtils_seq3, bio_sequtils_six_frame NCBIBiologicalR5ncbi_entrez_einfo,ncbi_entrez_elink,ncbi_entrez_fetch,ncbi_entrez_search, ncbi_entrez_summary OtherBiologicalR11codon_table_by_id, codon_table_list, codon_table_standard, expasy_prodoc, expasy_prosite, expasy_prosite_raw, iupac_data_letters, iupac_data_weights, togows_entry, togows_search, togows_search_count RestrictionBiologicalR2restriction_all_enzymes, restriction_enzyme_info GeneralGeneralP1code_execution GeneralGeneralR3browse, parallel_search, search B.2. Exemplar Sources B.3. Synthesis Prompts We provide the core prompts used in the DIVE synthesis pipeline. Variable names in brackets (e.g.,domain) are placeholders filled during runtime. 17 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use 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. BenchmarkTask 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 HotpotQA (Yang et al., 2018)Multi-hop information retrieval and reasoning GAIA (Mialon et al., 2023)Complex multi-step reasoning and tool planning HLE (Phan et al., 2025)Deep multidisciplinary reasoning and knowledge FinQA (Chen et al., 2021)Numerical reasoning over financial reports TAT-QA (Zhu et al., 2021)Hybrid reasoning over tabular and textual data PubMedQA (Jin et al., 2019)Biomedical research question answering BioASQ (Tsatsaronis et al., 2015)Biomedical semantic indexing and question answering API-Bank (Li et al., 2023)Tool usage and dialogue management Frames (Krishna et al., 2025)Unified evaluation of retrieval-augmented generation SciCode (Tian et al., 2024)Research-level scientific coding problems 휏 -bench (Yao et al., 2024)Tool-agent-user interaction in real-world domains TheAgentCompany (Xu et al., 2024)Consequential real-world professional tasks TravelPlanner (Xie et al., 2024)Real-world travel planning with constraints LegalAgentBench (Li et al., 2025a)Legal reasoning and document analysis with tools HealthBench (Arora et al., 2025)Comprehensive health-related LLM evaluation MCP-Bench (Wang et al., 2025b)Complex real-world tasks via MCP servers Evidence Collection. The agent interacts with the sampled toolset to accumulate grounded evidence (tool execution traces with outputs). Evidence Collection (Round 1) Research “seed_concept” in domain domain. Use multiple tools to retrieve and process comprehensive and verifiable information from various sources. Step budget: max_steps Note: Investigate the topic from multiple angles and explore its connections to related entities or concepts. Strategy: If direct search has limited results, try related concepts, broader categories, or alternative terms. Consider how the different aspects of the topic relate to each other. Evidence Collection (Round 퐾 > 1) Continue research on “current_query” in domain domain. Step budget: max_steps Previous findings: accumulated_evidence Based on previous findings, expand the research to broader or deeper aspects. Use diverse tools to retrieve and process new information. Avoid repeating previous findings. Task Derivation.Based on the accumulated evidence, the model derives a query-answer pair strictly grounded in the execution traces. 18 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Task Derivation (Round 1) Exemplars: exemplars Seed: seed_concept Evidence collected: accumulated_evidence Derive a specific and realistic query using the collected data. Base the answer on actual tool results only. QUERY: [specific query grounded in evidence] ANSWER: [concise factual answer from tool results only - no explanations, no reasoning, just the key values/facts] REASONING: [how the evidence supports this query-answer pair] Task Derivation (Round 퐾 > 1) Exemplars: exemplars Current: current_query Evidence collected: accumulated_evidence Refine the question to be more challenging, specific and realistic using the diverse collected data. Base answer on actual tool results only. EVOLVED_QUERY: [more complex question using collected data] EVOLVED_ANSWER: [brief, factual answer from tool results - be concise, specific to the question] REASONING: [what complexity was added] Verification.We use Claude-4-Sonnet and DeepSeek-V3.2 as cross-verifiers; an answer is marked correct only if both models agree. On a 200-sample human audit, verifiers achieved 100% agreement, owing to concise, unambiguous reference answers. Answer Verification Evaluate the correctness of the model’s answer. QUERY: query REFERENCE ANSWER: ground_truth MODEL ANSWER: model_response Evaluation criteria: - Compare factual content, not surface format - Ignore differences in phrasing or presentation - Focus on whether the core factual claims are correct Output format: JUDGEMENT: [correct/partial/incorrect] EXPLANATION: [Brief justification] Use “correct” if all key facts match, “partial” if the core answer is right but some details are wrong or missing, “incorrect” if the main answer is wrong. C. Diversity Analysis To analyze how iterative synthesis increases task diversity, we sample 4,000 tasks and evaluate structural diversity metrics across iteration rounds. For each iteration count퐾 ∈ 1, 2, 3, we use GPT-OSS-120B to solve the synthesized tasks and measure diversity in the resulting trajectories. Key observations. (1) Diversity scales with iterations: All structural diversity metrics increase substantially from퐾=1to퐾=3. Tool coverage grows from 58.2% to 92.2%, unique R/P topologies increase by 341%, and topology class coverage expands from 42 to 19 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use 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). MetricK=1K=2K=3Δ (1→3) GPT-OSS-120B Pass Rate (%)66.851.141.2 −38% Avg. Tool Calls per Task4.67.510.1+120% Tool Coverage (%)58.274.392.2+58% Unique Tool-call Sequences8921,3212,348+163% Unique Tool-call Graphs1,3622,4673,551+161% Unique R/P Topologies5421,0042,393+341% R/P Topology Classes42121165+293% Distinct Tools per Task1.892.243.15+67% 165 classes. (2) Task complexity increases: The decreasing pass rate (66.8%→41.2%) and increasing tool calls per task (4.6→10.1) indicate that iterative synthesis produces more challenging tasks requiring longer solution trajectories. (3) Richer tool compositions: Distinct tools per task grows from 1.89 to 3.15, showing that later iterations induce tasks requiring more diverse tool combinations rather than repetitive single-tool patterns. These results validate the iterative synthesis design: each additional iteration explores new regions of the tool space, producing tasks that cover more tools, exhibit more structural variety, and require more sophisticated reasoning. D. Topology Class Definition We define topology classes using a 3-level hierarchy to systematically categorize tool-call graph patterns. Table 9 summarizes the classification scheme. Table 9. Topology class dimensions (3-level hierarchy). Structure types are checked in priority order and are mutually exclusive. LevelDimensionValuesDefinition 1. R/P Type PureR–Only Retrieval tools R+P–Both Retrieval and Processing PureP–Only Processing tools 2. Structure (priority order) Single푛 = 1Single tool call Indep푛 > 1∧|퐸| = 0Multiple independent calls Phain푒 = 푛−1∧ max(in/out)≤ 1Linear chain Forksources = 1∧ sinks > 1∧ max(in)≤ 1One-to-many Joinsinks = 1∧ sources > 1∧ max(out)≤ 1 Many-to-one DAGmax(in) > 1∧ max(out) > 1Complex DAG MixOtherMixed structure 3. Scale Depth (Phain/Fork/Join/DAG/Mix) d1-2, d3-4, d5-7, d8+Longest path length Width (Fork/Join/DAG/Mix)w1-2, w3-5, w6-10, w11+Max BFS layer width Node count (Indep only)n2-3, n4-6, n7-10, n11-20, n21+Number of independent calls Table 10. Theoretical class count per structure type (× 3 R/P types). StructureScale ParamsBinsClasses Singlenone13× 1 = 3 Indepnode count53× 5 = 15 Phaindepth only43× 4 = 12 Forkdepth× width4× 43× 16 = 48 Joindepth× width4× 43× 16 = 48 DAGdepth× width4× 43× 16 = 48 Mixdepth× width4× 43× 16 = 48 Total222 Naming convention: Format varies by structure type: • Single: R/P/Single, e.g., PureR/Single • Indep: R/P/Indep/n, e.g., R+P/Indep/n4-6 20 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use • Phain: R/P/Phain/d, e.g., PureP/Phain/d3-4 • Fork/Join/DAG/Mix: R/P/Structure/d/w, e.g., R+P/DAG/d5-7/w3-5 DIVE covers 153 of 222 possible classes (69%); Gen-DR covers only 65 (all PureR, 29%). E. Additional Experimental Results 0255075100 0.2 0.4 0.6 0.8 Accuracy +20% Academic 0255075100 0.2 0.4 0.6 0.8 +39% Biological 0255075100 0.2 0.4 0.6 0.8 +44% Financial 0255075100 0.2 0.4 0.6 0.8 +20% Medical 0255075100 0 10 20 Calls / Task +160% 0255075100 0 10 20 +46% 0255075100 0 10 20 +42% 0255075100 0 10 20 +64% 0255075100 0 100 200 Tool-call Graphs +144% 0255075100 0 100 200 +28% 0255075100 0 100 200 +40% 0255075100 0 100 200 +2% 0255075100 Steps 0 100 200 R/C Topologies +143% 0255075100 Steps 0 100 200 +28% 0255075100 Steps 0 100 200 +39% 0255075100 Steps 0 100 200 +1% Figure 6. RL training dynamics by domain over 100 steps. 4 rows (metrics: Accuracy, Calls/Task, Tool-call Graphs, R/P Topologies)× 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. E.1. Scaling Analysis Raw Data The following tables provide raw experimental data for the scaling analysis in Figure 2. F. Training Details SFT. We fine-tune Qwen3-8B in bf16 precision using AdamW (훽 1 =0.9,훽 2 =0.95,휖=1e-8, weight decay 0.1) with a cosine learning rate schedule (warmup 5% of total steps, peak lr 1e-5, min lr 0). RL.We use GRPO with entropy loss enabled, gradient clipping at 1.0, and an off-policy filter (threshold 12.0) to discard stale samples. Rollouts are generated with SGLang (TP=4, group size 8) and a memory fraction of 0.7. Training uses TP=4 and context parallelism (CP=4) with dynamic micro-batch sizing and token-level loss. 21 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use 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→4 domains, 174→373 tools). Quantity Scaling: Fixed general-purpose search/browse tools (2 tools), expanding data (12k→48k). Condition SettingFSC 2 FSC 3 FAB MAB GAIABCHLE XB-DS SWE Tool. Scaling Diversity (12k fixed, tool pool expanding) 1 domainfin (174 tools)57.119.014.038.832.06.58.531.210.90.9 2 domainfin_med (263 tools)60.723.415.543.645.59.811.544.211.21.5 3 domainfin_med_bio (324 tools)61.227.417.045.247.110.812.348.611.82.5 4 domainfin_med_bio_aca (373 tools)62.029.818.047.650.512.213.450.112.52.9 Scaling Quantity – Gen-DR (search/browse only, 2 tools fixed) 12k1 config55.115.65.522.938.111.511.850.14.40.3 24k1 config60.119.08.530.940.911.112.449.24.30.9 36k1 config57.617.06.333.242.611.312.750.04.10.6 48k1 config51.314.34.729.747.811.612.650.14.20.3 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→48k) and configurations. Pool+Config Scaling: Jointly expanding tool pool diversity (174→373 tools) and data quantity, all paths starting from fin. Condition SettingFSC 2 FSC 3 FAB MAB GAIABCHLE XB-DS SWE Tool. Config Scaling (fin domain fixed, 174 tools) 12kfin57.119.014.038.832.06.58.531.210.90.9 24kfin60.622.915.041.341.68.99.837.611.31.2 36kfin61.624.216.041.845.59.810.540.511.51.5 48kfin62.125.316.542.047.010.311.042.811.51.5 Pool+Config Scaling (all paths starting from fin, tool pool expanding) 12k (1d)fin57.119.014.038.832.06.58.531.210.90.9 24k (2d) fin_med60.525.016.048.350.58.810.848.011.93.1 fin_bio62.226.210.041.849.511.613.657.011.51.4 fin_aca56.327.412.041.948.511.111.040.012.42.5 36k (3d) fin_med_bio60.530.424.050.147.511.314.247.012.22.5 fin_med_aca62.327.522.049.649.911.312.454.013.03.1 fin_bio_aca60.526.218.043.352.412.212.746.013.64.4 48k (4d)fin_med_bio_aca62.133.327.050.250.513.213.850.213.24.7 22 DIVE : Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use Table 13. Scaling diversity & quantity (SFT→RL). Each cell shows category score after SFT and after RL (SFT→RL).ΔAvg denotes Avg 푅퐿 -Avg 푆퐹푇 . L2-A averages GAIA/XB-DS/BC/HLE; L2-B averages FSC 2 /FSC 3 . DomainsComboL2-AL2-BL3-AL3-BL3-CL3-DAvgΔAvg 1 domain 12k fin19.6→29.0 38.0→44.8 14.0→22.0 38.8→49.9 10.9→11.9 0.9→3.7 20.4→26.9+6.5 med20.8→31.1 37.0→44.0 12.0→16.0 47.1→52.1 12.2→12.4 2.3→4.4 21.9→26.7+4.8 aca30.1→31.0 37.6→46.5 10.0→16.0 42.3→44.8 14.2→14.7 1.5→2.5 22.6→25.9+3.3 bio25.3→30.1 38.2→37.16.0→8.039.9→50.3 11.0→15.9 1.0→2.5 20.2→24.0+3.7 Avg23.9→30.3 37.7→43.1 10.5→15.5 42.0→49.3 12.1→13.7 1.4→3.3 21.3→25.9+4.6 2 domain 24k fin-med29.5→31.8 42.8→45.5 16.0→26.0 48.3→56.4 11.9→14.2 3.1→4.0 25.3→29.7+4.4 fin-bio32.9→32.3 44.2→45.3 10.0→20.0 41.8→55.3 11.5→13.7 1.4→4.4 23.6→28.5+4.9 med-bio28.1→32.9 43.5→43.4 14.0→16.0 49.8→56.6 11.8→12.8 2.5→5.3 25.0→27.8+2.9 aca-med28.6→31.0 41.0→41.2 14.0→16.0 46.6→53.4 14.2→16.2 3.1→5.3 24.6→27.2+2.6 aca-fin27.6→31.1 41.8→47.5 12.0→14.0 41.9→51.3 12.4→13.3 2.5→3.7 23.1→26.8+3.8 aca-bio31.8→34.4 43.0→42.6 12.0→10.0 43.2→52.2 13.5→17.3 1.4→3.7 24.1→26.7+2.6 Avg29.8→32.2 42.7→44.3 13.0→17.0 45.3→54.2 12.6→14.6 2.3→4.4 24.3→27.8+3.5 3 domain 36k fin-med-bio30.0→33.6 45.5→50.9 24.0→32.0 50.1→56.6 12.2→18.0 2.5→6.2 27.4→32.9+5.5 fin-bio-aca30.8→35.0 43.4→49.7 18.0→30.0 43.3→55.7 13.6→17.2 4.4→6.5 25.6→32.4+6.8 fin-med-aca31.9→33.6 44.9→45.9 22.0→28.0 49.6→54.8 13.0→15.5 3.1→4.0 27.4→30.3+2.9 med-bio-aca29.1→34.5 43.9→45.3 18.0→18.0 48.3→53.9 12.6→14.3 3.7→5.3 25.9→28.5+2.6 Avg30.5→34.2 44.4→48.0 20.5→27.0 47.8→55.3 12.8→16.2 3.4→5.5 26.6→31.0+4.4 4 domain 48k aca-fin-med-bio 31.9→38.4 47.7→52.3 27.0→34.0 50.2→57.3 13.2→18.3 4.7→8.3 29.1→34.8+5.6 Avg31.9→38.4 47.7→52.3 27.0→34.0 50.2→57.3 13.2→18.3 4.7→8.3 29.1→34.8+5.6 23