Why Diversity Beats Data Scale: Insights from MiniMax & Fudan’s DIVE Paper

When training AI agents that can invoke external tools, the intuitive solution is to collect more data, but the DIVE paper (Dive: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use) provides a counter‑intuitive answer: the breadth and domain coverage of the tool pool is the primary driver of generalization.

Problem with existing datasets – Current static datasets such as APIGen and ToolBench suffer from three limitations: (1) tool combinations are narrow, typically using only 1‑2 tools in a fixed retrieve‑calculate‑answer pattern; (2) domain coverage is limited to generic web search, lacking specialized APIs in finance, biology, or medicine; (3) scaling data volume alone fails – expanding from 12k to 48k homogeneous examples actually degrades performance.

DIVE’s core insight – Models need diverse task types, not more of the same. The authors liken it to a chef who becomes a true “general‑purpose” cook only after practicing a variety of cuisines, not after repeating the same dish thousands of times.

Four‑stage closed‑loop synthesis pipeline (Collect → Verify → Instantiate → Generate):

Seed‑driven guidance : start from a small set of high‑quality tasks to generate initial tool‑call trajectories.

Multi‑domain synthesis : expand across four domains (finance, medicine, biology, academia) covering 373 tools, each annotated as Retrieve (R) or Process (P) to enable realistic R→P→R→P alternating patterns.

Graph‑preserving augmentation : deliberately create complex dependency graphs (trees, DAGs) to increase structural diversity.

Self‑correct filtering : automatically verify generated data and discard low‑quality samples.

Tool‑pool diversity – The 373‑tool pool is organized by R/P type, reflecting real‑world workflows where retrieval and processing alternate. This design supplies richer training signals than a simple search+browse pair.

Structural diversity – Beyond tool count, DIVE measures four dimensions of task structure. Existing datasets are homogeneous, with 1‑3 steps and linear graphs. DIVE’s graph‑preserving stage injects complex topologies, which the authors identify as crucial for generalization.

Key Experiment 1: Diversity vs. Quantity – Two groups were compared while keeping the number of training trajectories constant at 12k:

Diversity group : expanded from 1 to 4 domains (174 → 263 → 324 → 373 tools).

Quantity group : kept only two generic tools (search + browse) and increased trajectories from 12k to 48k.

Results: the diversity group showed monotonic gains across all benchmarks (e.g., FSC2 rose from 57.1 to 62.0, a ~5‑point increase). The quantity group hit an over‑fitting ceiling – FSC2 dropped from 55.1 to 51.3, and functional‑agent benchmark (FAB) fell from 18.0 to 8.5.

Key Experiment 2: Configuration expansion vs. Joint pool + configuration expansion – Adding more data within a single domain still helped (12k→48k raised FSC2 by +5.0), but the joint expansion (increasing both domains and data) produced a qualitative leap: with three domains and 36k examples, FAB jumped from 16.5 to 24.0 (+45 %), and XB‑DS rose from 42.8 to 57.0 (+33 %).

SFT → RL pipeline – DIVE first fine‑tunes on synthetic data (SFT) then applies GRPO‑based reinforcement learning (RL). RL consistently improves performance across all configurations, and also boosts structural diversity: during RL, the model discovers longer, more complex tool‑call chains, increasing graph diversity by 143 % in the academic domain.

Six core conclusions :

Diversity outweighs sheer data volume.

Joint expansion of tool pool and data yields the best gains.

RL enhances both accuracy and structural diversity.

Cross‑domain training leads to genuine generalization on unseen benchmarks.

SFT → RL is a robust, universally beneficial workflow.

Self‑correct filtering is essential for reliable synthetic data.

Practical recommendations for building AI agents :

Stop blindly increasing data size; first assess tool‑pool coverage.

Incorporate multi‑domain tools even for single‑domain applications.

Balance Retrieve and Process tool types to mimic realistic R/P alternation.

Adopt the SFT → RL two‑stage training as a standard practice.

Include a self‑validation step to filter low‑quality synthetic tasks.

Overall, DIVE demonstrates a systematic approach—from theoretical framework to experimental validation—that answers the central question of how to train agents capable of generalizing to new tools: expose them to diverse tool‑use scenarios rather than merely scaling homogeneous data.