Research map / Large language models / Agents, code and math
Expanding the Capability Frontier of LLM Agents with ZPD-Guided Data Synthesis
Xuanzhong Chen, Zile Qiao, Guoxin Chen, Liangcai Su, Zhen Zhang, Xinyu Wang, Pengjun Xie, Fei Huang, Jingren Zhou, Yong Jiang, Ting Chen
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 |
|---|---|---|---|
| RFT on AgentFrontier instead of next-best dataset (MiroVerse) | Qwen3-8B · HLE text-only | accuracy (%): 15.00 -> 18.80 | Improved |
| RFT on AgentFrontier instead of next-best dataset (MiroVerse) | Qwen3-32B · HLE text-only | accuracy (%): 19.92 -> 23.82 | Improved |
| RFT on AgentFrontier instead of next-best dataset (MegaScience) | Qwen3-30B-A3B-Thinking-2507 · HLE text-only | accuracy (%): 20.15 -> 25.67 | Improved |
| RFT on any synthetic agent dataset versus tool-augmented base model | Qwen3-30B-A3B-Thinking-2507 · HLE text-only | accuracy (%): 10.17 -> 25.67 (AgentFrontier) | Improved |
| add continued pre-training on 50B tokens before RFT (ablation) | AgentFrontier-30B-A3B · HLE text-only | accuracy (%): 25.7 -> 28.6 | Improved |
| add continued pre-training before RFT (ablation) | AgentFrontier-30B-A3B · xBench-ScienceQA | accuracy (%): 54.0 -> 61.0 | Improved |
| add continued pre-training before RFT (ablation) | AgentFrontier-30B-A3B · ZPD Exam-v1 | accuracy (%): 91.4 -> 93.4 | Improved |
| add continued pre-training before RFT (ablation) | AgentFrontier-30B-A3B · RBench-T | accuracy (%): 74.4 -> 77.1 | Improved |
| RFT on AgentFrontier for tool-use accuracy | Qwen3-30B-A3B · HLE text-only | overall conditional tool accuracy (%): 20.5-21.0 -> 26.3 | Improved |
| sample 8 attempts instead of 1 | RFT Qwen3-30B-A3B · AgentFrontier validation (300) | accuracy (%): 21.7 -> 40.7 | Improved |
| give tools to base model | Qwen3-30B-A3B-Thinking-2507 · HLE text-only | accuracy (%): 9.2 -> 10.2 | No clear effect |
| give tools to base model | GPT-4o · ZPD Exam-v1 | accuracy (%): 4.8 -> 51.3 | Improved |
Closest papers by meaning
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.90)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.89)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.89)
- How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.89)
- AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.88)
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 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.
