New Solver Outperforms GPT-5.2 Pro on ARC-AGI-2 Benchmark

Large language models (LLMs) can generate fluent and internally consistent reasoning for abstract tasks, but they often produce confidently incorrect answers. This highlights that the primary challenge isn't just generating hypotheses, but effectively selecting the correct one from multiple candidates. A new solver for ARC-AGI-2, a few-shot visual reasoning benchmark, addresses this by employing two core principles.First, it treats different reasoning modalities—text, image, and code—as distinct search operators, allowing for the independent generation of diverse candidate solutions across these channels. This approach ensures a wide range of potential answers are considered. Second, it utilizes a context-preserving holistic judging mechanism. A dedicated judge model simultaneously compares all generated reasoning traces within a single, long-context prompt.This method, unlike simpler self-consistency or majority voting, has proven capable of identifying correct minority hypotheses even when the most common answer is wrong. On the ARC Prize semi-private evaluation set, the solver achieved an impressive 72.9% accuracy, significantly outperforming leading standalone models like GPT-5.2 Pro (54.2%) and Gemini 3 Pro (54.0%). The public evaluation set saw similar success at 76.1%. The researchers also shared extensive negative results, noting that prescriptive prompting and iterative refinement can reduce hypothesis diversity and degrade performance.