A new solver for the ARC-AGI-2 benchmark employs modality-driven search and holistic trace judging to improve selection among reasoning candidates. By generating diverse outputs across text, image, and code channels and using a judge model to compare them within a single prompt, the approach reliably identifies correct minority hypotheses even when the majority answer is wrong.
- The solver achieves 72.9% on the semi-private evaluation set at $38.99 per task, outperforming GPT-5.2 Pro (54.2%) and Gemini 3 Pro (54.0%).
- It reaches 76.1% on the public evaluation set at $19.69 per task.
- The author documents that prescriptive prompting templates and iterative refinement systematically reduce hypothesis diversity and degrade performance.
The work highlights the importance of preserving hypothesis diversity in abstract reasoning tasks, demonstrating that treating reasoning modalities as search operators is more effective than standard self-consistency methods.