AdaReasoner, a 7B model, excels in puzzle reasoning over GPT-5 via dynamic tool orchestration. Implements Agentic Vision with think-act-observe loops for visual tasks. Accepted to ICLR 2026 with open code and models.
Key Points
- 1.Dynamic tool selection
- 2.7B outperforms larger models
- 3.Agentic vision loops
Impact Analysis
Shows small models can rival giants through smart tool use, inspiring efficient vision agents.
Technical Details
Learns what/when/how to use tools iteratively; supports visual reasoning benchmarks.