adaptive-compute

LEV 是一种通过循环重跑层而非生成文本来实现可变推理深度的决策模型，在相同权重下随深度提升准确率，并通过按需深度机制节省 15%-24% 计算量，在低延迟下显著优于同等规模的稠密模型。

RecurTrace introduces loop-time memory and adaptive halting to improve latent reasoning in language models, achieving higher accuracy on MathQA with optimized compute compared to fixed-loop methods.

CAPS introduces a cascaded adaptive selection framework for efficient parallel reasoning, reducing verifier compute costs by over 75% while outperforming existing pairwise verification methods across multiple LLM benchmarks.

GridProbe is a training-free inference paradigm for Long-Video VLMs that adaptively selects relevant frames using posterior probing, achieving sub-quadratic attention costs with minimal accuracy loss.