AI4SGI/ExoMind
ExoMind by AI4SGI/Shanghai Artificial Intelligence Laboratory is a 35.1 billion parameter extended-mind-inspired agentic system, fine-tuned from Qwen3.5-35B-A3B, designed for scientific reasoning and research. It integrates a general-purpose model with specialized interaction objects and autonomous processes to discover sources, ground evidence, and verify scientific problems. The model achieves an average score of 68.3 across eight scientific benchmarks, outperforming other evaluated models, and is optimized for complex scientific workflows.
Loading preview...
ExoMind: Extended-Mind-Inspired Agentic System for Scientific Intelligence
ExoMind, developed by AI4SGI and Shanghai Artificial Intelligence Laboratory, is a 35.1 billion parameter model fine-tuned from Qwen3.5-35B-A3B. It represents the first extended-mind-inspired agentic system specifically engineered for scientific reasoning and research. The system unifies a large language model with specialized interaction objects and autonomous processes, enabling it to perform tasks like source discovery, evidence grounding, executable verification, and iterative reasoning updates for scientific problems.
Key Capabilities and Innovations
- Extended-Mind-Inspired Intelligence: Integrates the LLM, interaction objects, and autonomous interaction processes into a cohesive scientific agentic system.
- Training-Value-Aware Data Engineering: Employs a sophisticated data engineering approach to identify challenging problems and route them for pure-reasoning or interaction-reasoning data processing.
- Scientific Interaction Framework: Facilitates composable objects for source discovery, evidence grounding, executable verification, and observation integration.
- Progressive CoI Training: Utilizes a two-stage progressive Chain-of-Interaction (CoI) training method with high-quality trajectories to develop both intrinsic reasoning and autonomous interaction.
- Efficient Frontier Performance: Achieves significant performance improvements, raising the average score across eight scientific benchmarks from 36.2 to 68.3, and ranks first on six of these benchmarks, all while maintaining efficient training on 8 NVIDIA H200 GPUs.
Performance Highlights
ExoMind demonstrates superior performance in scientific research and reasoning tasks. In evaluations, it achieved an average score of 68.3 across eight scientific benchmarks, surpassing the next-best representative model's average of 54.2. It notably ranks first in benchmarks such as FrontierScience-Research, CMT-Benchmark, AMO-Bench, IMO-AnswerBench, HiPhO, and FrontierScience-Olympiad.
Intended Use Cases
ExoMind is designed for research and development in areas including:
- Scientific question answering
- Literature-grounded investigation
- Mathematical and computational reasoning
- Code-assisted verification
- Agentic scientific workflows