AdaReasoner, a 7B model, achieves superior performance on visual reasoning tasks like puzzles by dynamically learning tool selection, timing, and usage. It introduces 'Agentic Vision' with iterative think-act-observe loops, outperforming larger models without massive scaling. Open-source code, models, and paper available on arXiv and GitHub.
Key Points
- 1.Dynamic tool orchestration for visual reasoning
- 2.Beats GPT-5 on puzzles with 7B parameters
- 3.Agentic Vision: think-act-observe cycle
Impact Analysis
Demonstrates efficient small models can rival giants via smart tool use, lowering barriers for visual AI agents. Influences shift from static image processing to proactive investigation in multimodal AI.
Technical Details
Trains on what/when/how of tools as reasoning skill. Integrates with Gemini 3 Flash's Agentic Vision paradigm for iterative refinement.