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

The paper introduces DIVE, an evidence-driven framework that prioritizes executing diverse real-world tools before reverse-deriving tasks to ensure grounding and structural variety, which significantly enhances the out-of-distribution generalization of tool-using LLMs compared to traditional quantity-focused scaling.

The Gist

Here is an explanation of the DIVE paper, translated into everyday language with some creative analogies.

The Big Problem: The "Robot Intern" Who Only Knows One Way to Do Things

Imagine you hire a brilliant robot intern to help you with your daily tasks. You train this robot by giving it thousands of examples of how to do one specific thing: searching the web for travel deals. The robot gets really, really good at that. It becomes a master of travel booking.

But then, you ask the robot to do something slightly different, like diagnosing a patient's illness or analyzing a complex stock portfolio. Suddenly, the robot freezes. It tries to "search the web" for medical advice or stock prices, gets confused, and fails.

Why? Because the robot was trained on a diet of only one type of food. It learned a rigid routine (Search → Click → Read) but never learned how to think flexibly or use different tools (like a calculator, a database, or a medical reference) in new combinations. It's like teaching a chef to only make toast, and then expecting them to bake a soufflé.

The Old Way vs. The DIVE Way

The Old Way (Query-First):
Most researchers try to fix this by asking an AI: "Hey, invent a new task for me!" The AI guesses a task (e.g., "Find the best protein for a rare disease"), and then the researchers try to see if the robot can actually do it.

• The Problem: It's like asking a student to write a math problem, and then hoping the answer actually works out. Often, the AI invents a task that is impossible to solve, or the "tools" it uses are fake. It's a lot of wasted effort checking if the task is even real.

The DIVE Way (Evidence-First):
The authors of this paper, DIVE, flipped the script. Instead of guessing a task first, they said: "Let's just go use the tools first and see what happens."

Think of it like cooking by accident:

1. The Old Way: You write a recipe for "Spicy Tacos," then try to find ingredients. If you can't find hot sauce, the recipe is useless.
2. The DIVE Way: You open the fridge, grab whatever random ingredients you have (a weird spice, some leftover chicken, a specific type of cheese), mix them together, and then you look at the delicious dish you made. You say, "Oh! This looks like a 'Spicy Chicken Surprise.' Let's write a recipe for that."

Because you started with real ingredients (real tools) and real results (real data), the recipe (the task) is guaranteed to be solvable.

How DIVE Works (The "Reverse-Engineered" Kitchen)

The DIVE system works in three simple steps:

1. The Tool Buffet: They gathered a massive buffet of 373 real-world tools. These aren't just "Google Search." They include tools for finance, biology, medicine, and coding. It's like having a kitchen with every appliance imaginable, not just a toaster.
2. The "Playground" Phase: The AI is sent into this kitchen. It picks random tools and starts "cooking" (executing them). It might look up a stock price, then calculate a percentage, then search for a news article about that stock. It creates a trail of breadcrumbs (evidence) of what it actually did.
3. The "Story" Phase: Once the AI has a trail of real actions, a second AI looks at the trail and says, "Wow, look at this interesting result! Let's write a question that leads to this exact answer."
   • Example: The AI calculates that a specific drug dosage works for a patient. DIVE then writes a question: "Which drug allows a patient to take 400mg in a single sip?"
   • The Magic: The question is guaranteed to be answerable because the answer was created first by the tools.

Why This is a Game-Changer

The paper tested this on a model called Qwen3-8B (a smart but relatively small robot). They trained it on 48,000 of these "reverse-engineered" tasks.

The Results:

• Generalization: When they tested this robot on tasks it had never seen before (like medical diagnosis or complex financial analysis), it didn't just do okay; it skyrocketed.
• The "Diversity" Secret: The researchers found that variety matters more than volume.
  • Analogy: Imagine training a student for a test.
    • Method A: Give them 48,000 practice questions, but they are all variations of "What is 2+2?"
    • Method B: Give them only 12,000 questions, but they cover math, history, science, and art.
  • DIVE found that Method B wins. Even with 4 times less data, training on diverse, real-world tools made the robot much smarter at solving new problems than training on a huge pile of repetitive, fake tasks.

The Bottom Line

DIVE is a new recipe for training AI agents. Instead of forcing AI to guess what it should learn, it lets the AI play with real tools first, creates a record of what happened, and then builds a lesson plan based on that reality.

It turns the AI from a parrot (who just repeats what it's been told) into a tinkerer (who knows how to mix and match tools to solve any problem, even ones it's never seen before). And the best part? It does this by proving that quality and variety of training data are far more important than just having more data.

Technical Summary

Here is a detailed technical summary of the paper "DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use."

1. Problem Statement

Current approaches to training tool-using Large Language Models (LLMs) rely heavily on synthesized agentic tasks. However, these models often fail to generalize when faced with shifts in task distributions or toolsets (Out-of-Distribution or OOD scenarios). The authors identify the root cause as insufficient diversity in synthesized training data.

Existing synthesis methods suffer from a fundamental tension:

• Structural Diversity: To generalize, models need exposure to heterogeneous tool types, diverse toolset combinations, and complex multi-step usage patterns (e.g., retrieval followed by analysis).
• Grounded Validity: For effective training (Supervised Fine-Tuning and Reinforcement Learning), tasks must be verifiable (have a deterministic correct answer) and executable (solvable with the provided tools).

Current solutions fail to balance this:

• Fixed-toolset synthesis (e.g., deep-research with only search/browse) lacks diversity.
• Simulated tools often produce unverifiable or unsolvable tasks.
• Query-first synthesis on real tools requires heavy manual verification to ensure solvability, which is not scalable.

2. Methodology: DIVE Framework

The authors propose DIVE (Diverse, Verifiable, and Executable), an evidence-driven synthesis recipe that inverts the traditional synthesis order. Instead of generating a query and hoping it is solvable, DIVE executes tools first and derives tasks from the resulting traces.

A. Core Mechanism: Evidence-First Synthesis

1. Inverted Order: The system executes real-world tools to generate execution traces (evidence) before formulating the task query. This guarantees grounding by construction: if the evidence exists, the task is solvable and verifiable.
2. Resource Pools: DIVE constructs three decoupled pools to maximize diversity:
   • Tool Pool: 373 validated tools across 5 domains (General, Finance, Medicine, Biology, Academia), categorized as Retrieval (fetching data) or Processing (transforming data).
   • Seed Pool: ~20,000 specific entity seeds (e.g., "Erlotinib") mined from Wikipedia, PubMed, NCBI, and stock exchanges to prevent topic collapse.
   • Exemplar Pool: Query-only examples from diverse benchmarks to provide structural priors for task phrasing and implicit tool-use patterns.
3. Synthesis Loop:
   • Configuration Sampling: Randomly samples a seed, a toolset (15–50 tools), and exemplars.
   • Evidence Collection: An agent agent performs multi-step reasoning, interleaving tool calls with the real toolset to accumulate a grounded evidence set ($E_k$).
   • Task Derivation: A generator LLM observes the accumulated evidence and reverse-engineers a specific Query-Answer pair ($Q_k, A_k$) strictly entailed by the traces.
   • Iteration: This loop runs $K$ times (typically 3), progressively increasing task complexity and diversity while maintaining executability.

B. Training Pipeline

The synthesized dataset is used for a two-stage post-training process on the Qwen3-8B model:

1. Agentic SFT (Cold Start): A teacher model generates trajectories for the synthesized tasks. Trajectories are accepted only if they match the reference answer (Rejection Sampling), creating a high-quality SFT dataset (48k tasks).
2. Agentic RL: Reinforcement Learning (GRPO) is applied to frontier tasks (those where the model is near the learning boundary) to improve robustness and exploration (3.2k tasks).

3. Key Contributions

• Identification of the Diversity Bottleneck: The paper argues that scaling quantity of data is insufficient; scaling structural diversity (tool types, combinations, and patterns) is critical for OOD generalization.
• Evidence-Driven Synthesis Recipe: DIVE solves the "verifiability vs. diversity" trade-off by deriving tasks from real execution traces, ensuring 100% executability and verifiability without manual filtering.
• Scalable Resource Construction: The creation of a massive, heterogeneous resource bank (373 tools, 5 domains) that enables the generation of 48k diverse training tasks.
• Scaling Laws for Diversity: The authors demonstrate that diversity scaling (expanding tool pools) outperforms quantity scaling (increasing data volume with fixed tools) for generalization, even with 4x less data.

4. Experimental Results

The model was evaluated on 9 benchmarks spanning In-Distribution (L1), General OOD (L2), and Specialized OOD (L3) settings.

• Performance Gains:
  • DIVE-8B (RL) improved by +22.2 average points across 9 OOD benchmarks compared to the base model.
  • It outperformed the strongest 8B baseline by +68%.
  • It achieved competitive performance against much larger frontier models (e.g., Gemini-3-Pro, Claude-4-Sonnet) on specialized benchmarks like Finance (FAB) and Medical (MAB).
• Generalization vs. Specialization: Unlike specialist models (e.g., WebExplorer-8B) that excel in one domain but fail in others (negative transfer), DIVE maintains robust performance across all domains.
• Scaling Analysis:
  • Diversity > Quantity: A model trained on 12k diverse tasks (4 domains) outperformed a model trained on 48k tasks with fixed tools (1 domain) on OOD benchmarks.
  • RL Amplification: RL further amplified the benefits of diversity, allowing the model to explore complex tool-use patterns (e.g., Retrieval-Processing topologies) that SFT alone could not fully master.
• Structural Diversity Metrics: DIVE trajectories showed significantly higher diversity in tool-call graphs, unique tool combinations, and Retrieval/Processing (R/P) topologies compared to baseline datasets.

5. Significance

• Paradigm Shift: DIVE challenges the prevailing "Query-First" synthesis paradigm, proving that "Evidence-First" synthesis is superior for training generalizable agents.
• Scalable Generalization: It provides a blueprint for training small-to-medium models (8B) to perform at the level of much larger models by focusing on data quality and diversity rather than just model size.
• Real-World Applicability: By using real, validated APIs and tools across critical domains (Medicine, Finance), DIVE bridges the gap between synthetic training and real-world deployment, reducing the risk of hallucination and unexecutable actions.
• Efficiency: The finding that diversity scaling yields better results with less data suggests a more compute-efficient path to robust agentic AI.

In conclusion, DIVE demonstrates that the key to generalizable tool use is not merely more data, but structurally diverse, grounded data synthesized through an inverted, evidence-driven process.