Apple Boosts Looped Transformer AIME Scores to 73.3 With LoopCD

← Back to live feed · 1 stories across 1 day A research team at Apple has introduced a way to optimize AI models during the decoding process to increase their accuracy without requiring further training. The method, known as LoopCD, raised scores on the AIME 2024 benchmark for the Ouro-2.6B model from 61.9 to 73.3 and improved HumanEval results for the Huginn model from 22.6 to 31.7. These gains are achieved as an inference-time optimization that does not require additional models. The release of the paper, titled "Decoding Looped Transformers Better for (Almost) Free," comes amid speculation that frontier models including Gemini 4 and GPT-6 Astra utilize looped transformer architectures. Apple MLR's research demonstrates that such recursive structures can be enhanced without the computational cost of retraining.