A pair of agentic architectures reached near-state-of-the-art scores on a standard AI reasoning benchmark without expensive training or brute-force compute.
Researchers tested DeepSeek V3.2 in non-thinking mode on the ARC-AGI-1 public evaluation set, which contains 400 pattern-reasoning tasks. Rather than fine-tuning on ARC data or running exhaustive sampling, they built two agent harnesses on top of the base model. The first, an Explorer-Definer Pipeline, splits the work into two stages: one agent discovers patterns, another synthesizes executable transformations. The second, a Reflective Orchestrator, adds a feedback loop that triggers fresh exploration when prior hypotheses fail. The pipeline scored 57.50% pass@2 at $0.25 per task; the orchestrator reached 67.25% pass@2 at $0.62 per task. Both beat a 15.50% one-shot baseline by roughly 52 points.
Most published ARC-AGI-1 progress has come from one of two expensive routes: frontier models burned through with extended chain-of-thought and evolutionary search, or small models fine-tuned specifically on ARC data. This work carves out a third lane - structured agent architecture - and shows it can close most of the gap at a fraction of the cost. The finding also has a diagnostic edge: the analysis suggests the pipeline is bottlenecked by generation diversity, not by which candidates it selects, and the orchestrator's re-exploration loop confirms that prediction directly.
One ablation worth noting: removing the pipeline's internal think tool dropped pass@2 by 5.75 points, a reminder that even "non-thinking" setups benefit from explicit reasoning scaffolding. Whether these results hold outside a controlled benchmark is, as always, the open question.