Research map / Large language models / Agents, code and math
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 and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale tool-pool coverage 1 to 4 domains (diversity-only, 12k) rather than quantity 12k to 48k with fixed tools (ablation) | SFT Qwen3-8B · OOD benchmarks | success rate | Improved |
| scale data quantity 12k to 48k with fixed deep-research tasks and tools (ablation) | SFT Qwen3-8B · OOD benchmarks | success rate | Hurt |
| expand tool pool across domains in addition to toolset variety (ablation) | SFT Qwen3-8B (Finance start) · OOD benchmarks | success rate | Improved |
| only vary toolsets with fixed tool pool (ablation) | SFT Qwen3-8B · OOD benchmarks | success rate | Mixed |
| add RL after SFT on DIVE data | DIVE-8B · 9 OOD benchmarks | average success points gain over base: +16.2 (SFT) -> +22.2 (RL) | Improved |
| add RL stage to SFT with more domains (ablation) | DIVE-8B · average over benchmarks (24 paths) | mean score: 29.1 -> 34.8 | Improved |
| RL-SFT gap grows with diversity (ablation) | DIVE-8B · average over benchmarks | AvgRL - AvgSFT: +4.6 at 1 domain vs. +5.6 at 4 domains | Improved |
| train on fixed-toolset deep-research specialist data (WebExplorer-8B) vs DIVE | Qwen3-8B · L3 specialized-tool benchmarks | success rate: drops 5.8 points below base model | Hurt |
| train on DIVE (RL) instead of Qwen3-8B base | Qwen3-8B · Toolathlon | success rate: 0.9 -> 8.3 | Improved |
| train on DIVE (RL) instead of EnvScaler-8B (query-first simulated) | Qwen3-8B · 9 OOD benchmarks | OOD lift: 3.2x larger OOD lift | Improved |
| DIVE (RL) vs DIVE (SFT) on SWE-bench Verified (ablation) | Qwen3-8B · SWE-bench Verified | success rate: 13.2 -> 18.3 | Improved |
Closest papers by meaning
- A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks (NeurIPS 2026 · similarity 0.91)
- ScaleEnv: Scaling Environment Synthesis from Scratch for Generalist Interactive Tool-Use Agent Training (ICML 2026 · similarity 0.90)
- AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents (ICLR 2026 · similarity 0.90)
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.89)
- SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (NeurIPS 2026 · similarity 0.89)
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
