Tag
The paper proposes TAPS (Trajectory Adaptive Progress–Fluctuation Scheduler), which adaptively adjusts the step size of recurrent updates in looped transformers by decomposing terminal loss sensitivity into persistent progress and centered fluctuation, improving accuracy without retraining and delivering up to 1.56× wall-clock speedup when applied during training.
DART proposes a distributional adversarial training framework for recurrent reasoning models, improving their robustness and performance on structured problems by using a local target distribution instead of single-point supervision.
The article introduces BDH-CQ, a 150M parameter recurrent model that combines in-context learning with latent reasoning, achieving 29.5% on ARC-AGI-1 at a cost of $0.0007 per task, setting a new standard for cost efficiency.