continuous-reasoning

The paper examines why pretrained language models fail to preserve continuous mixtures of token embeddings in reasoning, identifying architectural distortions and dynamical system effects as primary failure sources.

MUX proposes a method for lossless continuous reasoning by distilling discrete reasoning steps into multiplexed latent tokens that encode a superposition of subwords, achieving higher bandwidth and enabling parallel exploration in language model reasoning tasks.

Proposes Asymmetric Mutual Variational Learning (AMVL) to resolve train-inference mismatch in multimodal continuous reasoning by using bidirectional calibration to prevent answer leakage and improve latent-space stability, achieving significant gains on the BLINK benchmark.