Apple LoopCD Lifts Ouro-2.6B AIME to 73.33% Training-Free

The gains are large and the recipe is simple: run a looped transformer, then pick tokens by contrasting the final loop against an earlier one. In an Apple paper on Hugging Face, Weihao Liu and co-authors call this LoopCD and report that it "raises Ouro-2.6B-Thinking's AIME 2024 pass@1 from 61.88% to 73.33%" and "lifts Huginn's HumanEval pass@1 from 22.56% to 31.71%".

The authors frame it as free signal that recurrence already produced. "Each loop yields an intermediate representation decodable for the same next token, yet standard decoding discards earlier states," they write. Because earlier passes embody less computation, they stand in as weak references for the strong final prediction without any auxiliary model or extra training.

Two variants are described. LoopCD-Logits does the contrast in logit space and costs one extra output pass; LoopCD-Hidden operates on hidden states with "zero output overhead." The paper says gains hold across four looped Transformer families.

The headline claim for operators is compute. Halving the number of recurrent loops with LoopCD still matches or exceeds full-depth unguided baselines, "reducing forward FLOPs by 22.5% to 48.2%." Those two task numbers are the abstract's advertised wins; per-family tables are not in the abstract itself.

It is the second recurrence-and-reasoning paper on our open-source tracker this week, alongside the "Sharpening Tax" result on RL post-training surfaced the same day.