Tongyi DeepResearch: A New Era of Open-Source AI Researchers
a day ago
- #AI Agents
- #OpenSource
- #DeepResearch
- Tongyi DeepResearch is the first fully open-source Web Agent matching OpenAI’s DeepResearch performance across benchmarks.
- Achieves state-of-the-art results: 32.9 on HLE, 43.4 on BrowseComp, 46.7 on BrowseComp-ZH, and 75 on xbench-DeepSearch.
- Shares a complete methodology for advanced agents, including novel data synthesis, Agentic CPT, SFT, and RL stages.
- Introduces AgentFounder for scalable data synthesis, creating an entity-anchored open-world knowledge memory.
- Develops high-quality synthetic QA pairs through automated pipelines, enhancing AI agent performance.
- Bootstraps agent capabilities with SFT cold-starting using ReAct and IterResearch frameworks.
- Supports multiple rollout formats: Native ReAct Mode and Heavy Mode (IterResearch paradigm).
- Proposes Research-Synthesis framework for parallel research agent exploration and synthesis.
- Establishes an end-to-end agent training pipeline: Agentic CPT → Agentic SFT → Agentic RL.
- Uses on-policy Group Relative Policy Optimization (GRPO) for RL, ensuring stable and efficient training.
- Develops a synthetic training environment and stable tool sandbox for reliable agent training.
- Powers real-world applications like Gaode Mate (navigation agent) and Tongyi FaRui (legal research agent).
- Identifies limitations: 128k context length, scalability on larger models, and RL efficiency improvements.
- Part of an extensive deep research agent family with multiple published technical reports.
- Released Tongyi DeepResearch-30B-A3B model and plans for next-generation agentic models.