NeuralStackStanford researchers published a new architecture on July 16 2026 that achieves 98 percent accuracy on the ARC-AGI benchmark. The model combines dynamic program synthesis with a transformer backbone and trains on just 400 examples. It completes tasks in an average of 1.8 seconds on A100 GPUs.
The paper details a novel memory-augmented controller that generates executable code snippets on the fly. Lead author Dr. Elena Vargas presented the work at the International Conference on Machine Learning in Vienna. The team open-sourced the code under an MIT license on July 17.
ARC-AGI was introduced in 2019 to measure abstract reasoning beyond pattern matching. Previous top scores hovered around 79 percent using much larger models and datasets. The Stanford approach uses 92 percent fewer parameters than the prior leader.
Why this matters
The breakthrough demonstrates that targeted architectural innovations can outperform brute-force scaling on reasoning tasks. It may redirect research funding toward hybrid neuro-symbolic systems in the next funding cycle. Companies building automated software engineering tools are already testing the released code.
Follow-up papers are expected at NeurIPS 2026 with extensions to video reasoning tasks.