CMU and Oxford Paper Shows Looped Flows Hit 58.8% on ARC-AGI-1 Without Long Chains of Thought
rohanpaul_ai · x · 2026-09-18
A Carnegie Mellon and Oxford paper, "Thinking with Looped Flows," shows models can "think longer" by repeatedly refining a hidden state instead of writing longer chains of thought — test-time compute doesn't have to mean more tokens.
Key points:
- Standard looped models backprop through only one or a few recurrent updates, making early updates hard to train for future ones. Looped flows train the recurrence with local denoising objectives: progressively decreasing noise levels and shared noise create temporal association, incentivizing recurrent states that transfer useful computation over time.
- Inference integrates the velocity of a probability flow parameterized by the learned denoiser, coupled with recurrent states; harder problems are solved by spending more compute on a finer temporal grid, and different initial noise samples yield multiple valid predictions.
- Across six reasoning benchmarks including two multi-solution ones, looped flows beat prior looped-model SOTA: 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2.
Related event: CMU and Oxford: Looping Hidden States Can Replace Long Chain-of-Thought(2 posts)→
More from Research
- New Paper: Contrastive Noise Alignment Cuts FID Over 50% in Few-Step Generation — burny_tech · 2026-09-18
- ICLR26 Paper Defines 'Interpretive Equivalence': Comparing Neural Network Algorithms Without Full Interpretation — burny_tech · 2026-09-18
- DeepSWE-mini: A 16-Instance Subset That Replicates the DeepSWE Leaderboard Rankings — asankhs · 2026-09-18
- Tsinghua and ByteDance Seed unveil SMELT, the first fair comparison of Looped Transformers — jiqizhixin · 2026-09-18
- Gary Marcus: Judea Pearl's causality challenges remain unsolved in the LLM paradigm — GaryMarcus · 2026-09-18
- TPA: Tensor Product Attention unifies MHA/GQA/MLA, NeurIPS 2025 Spotlight — burny_tech · 2026-09-18