Search papers, labs, and topics across Lattice.
The paper introduces DIVE, a method for synthesizing diverse agentic tasks to improve the generalization of tool-using LLMs. DIVE inverts the task synthesis order by first executing diverse, real-world tools and then reverse-deriving tasks from the resulting traces, ensuring executability and verifiability. Experiments show that training Qwen3-8B on DIVE-generated data significantly improves performance on out-of-distribution benchmarks, with diversity scaling proving more effective than quantity scaling.
Forget scaling data volume: scaling *diversity* in agentic task synthesis boosts tool-using LLMs' out-of-distribution generalization by +22 points, even with 4x less data.
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.
