Research map / Large language models / Agents, code and math
TaskCraft: Automated Generation of Agentic Tasks
Dingfeng Shi, Jingyi Cao, Qianben Chen, Weichen Sun, Weizhen Li, Hongxuan Lu, Fangchen Dong, Tianrui Qin, King Zhu, Minghao Liu, Yuchen Eleanor Jiang, Jian Yang and 5 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 |
|---|---|---|---|
| add tool context (input index and relation) to atomic task generation (ablation) | TaskCraft atomic task generation · Atomic task generation | pass rate (%): 18.5% -> 43.0% | Improved |
| add tool context to atomic task generation (ablation) | TaskCraft atomic task generation · Atomic task generation | generation time: 119.7s -> 86.7s | Improved |
| add tool context to atomic task generation (ablation) | TaskCraft atomic task generation · Atomic task generation | average tool use: 2.8 -> 2.1 | Improved |
| add tool context to atomic task generation (ablation) | TaskCraft atomic task generation · Atomic task generation | variance of tool use: 1.2 -> 0.4 | Improved |
| add bootstrap few-shot prompt optimization (ablation) | atomic task generation · Atomic task generation | pass rate (%): 54.9% -> 68.1% | Improved |
| add prompt optimization (ablation) | atomic task generation · Atomic task generation | time per document: 29.1s -> 23.5s | Improved |
| add prompt optimization (ablation) | depth-wise extension · Depth-wise@6 | pass rate (%): 41.0% -> 51.2% | Improved |
| add prompt optimization (ablation) | depth-wise extension · Depth-wise@6 | time per task: 31.5s -> 30.2s | Improved |
| add SFT on synthetic trajectories | Qwen2.5-3B-Base base workflow · HotpotQA, Musique, Bamboogle average | accuracy: 0.034 -> 0.174 | Improved |
| add SFT on synthetic trajectories | Qwen2.5-3B-Instruct base workflow · HotpotQA, Musique, Bamboogle average | accuracy: 0.113 -> 0.173 | Improved |
| add SFT on synthetic trajectories | Search-R1 Qwen2.5-3B-Base · Bamboogle | accuracy: 0.088 -> 0.280 | Improved |
| add SFT on synthetic trajectories | Search-R1 Qwen2.5-3B-Base · Musique | accuracy: 0.049 -> 0.111 | Improved |
Closest papers by meaning
- AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents (ICLR 2026 · similarity 0.90)
- TaskBench: Benchmarking Large Language Models for Task Automation (NeurIPS 2024 · similarity 0.89)
- A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks (NeurIPS 2026 · similarity 0.89)
- AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering (NeurIPS 2026 · similarity 0.89)
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.89)
- AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 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.
