DeepSeek V3.2 reaches 67% accuracy on the ARC-AGI abstract reasoning test

Researchers have published a method on arXiv that lets the open DeepSeek V3.2 model solve tasks from the ARC-AGI-1 abstract reasoning test with 67% accuracy at pass@2. It relies only on agentic architectures, runs without thinking tokens and uses no special training on ARC data.

• DeepSeek V3.2 is an open model, run in non-thinking mode
• Explorer-Definer Pipeline: 57.50% pass@2 at $0.25 per task
• Reflective Orchestrator: 67.25% pass@2 at $0.62 per task
• A 52-point improvement over the baseline one-shot result (15.50%)
• Finding: the quality of candidate generation matters more than candidate selection

How the two-stage agent works

The researchers split the task into two stages: pattern discovery and program synthesis. The Explorer-Definer Pipeline consists of two sequential agents. In the first stage, the model finds regularities in the examples; in the second, it turns those patterns into executable transformation code. This split lets the system break the task down methodically instead of trying to solve it "from scratch."

Result: the pipeline reached 57.50% pass@2 accuracy on the public set of 400 tasks, spending $0.25 per task. For comparison, the one-shot baseline (a single example) scored 15.50%.

Reflective Orchestrator: iterative re-exploration

Building on the pipeline, the authors created the Reflective Orchestrator, a system that autonomously re-explores a task when a hypothesis fails on the training examples. When the transformation it found does not work, the Orchestrator looks for new candidates rather than just cycling through those it already has. This raised accuracy to 67.25% pass@2, but cost rose to $0.62 per task, a trade-off between accuracy and budget.

A critical observation: the analysis showed the model is generation-bound, not selection-bound. Picking the best answer from those proposed captures ~95% of the potential; the remainder has to come from expanding generation itself. The Orchestrator confirmed this: unbiased pass@1 improved by 9.81 points thanks to new candidates, not re-ranking of old ones.

Why this challenges frontier models

The authors avoided two popular approaches: heavy test-time compute (evolutionary search, chain-of-thought) on top of GPT-5/Claude, and benchmark-specific fine-tuning of small models. Instead, they showed that the open DeepSeek V3.2 reaches 67% through an agentic architecture alone, without fine-tuning. The model's think tool is a notable component: disabling it cut the result by 5.75 points, which points to hidden reasoning even in non-thinking mode.

What this means

The study offers an alternative to scale: splitting the task properly and re-exploring iteratively yield bigger gains than simply using a larger model with more compute. For developers, it is a signal that a well-built system can get around scale through architecture engineering.

Frequently asked questions

How does ARC-AGI-1 differ from ordinary capability tests?

ARC-AGI-1 consists of 400 tasks about discovering hidden patterns. The model sees 3–5 examples (input-output pairs) and must propose the transformation logic for new inputs. It is closer to "learning from examples in a handful of steps" than to knowledge questions.

Why is this cheaper than fine-tuning?

Fine-tuning requires dataset preparation, training and validation on GPUs. Here there is only inference with agentic calls: three to four model calls per task. Everything happens at inference time, with no weight updates.

When will this show up in real products?

For now, this is a study on arXiv dated July 6, 2026. Adoption in products depends on whether the method generalizes beyond ARC-AGI, which is still unclear.