A new arXiv preprint, identified as arXiv:2610.20642, introduces CoTS (Competitive Token Selection) for latent reasoning in large language models. The paper positions CoTS against recent latent reasoning methods that move beyond discrete Chain-of-Thought (CoT) prompting.
According to the abstract, these methods replace sampled tokens with continuous soft embeddings or hidden states. The authors argue this enables richer intermediate representations and preserves alternative options, though the sentence is cut off in the provided text.
The abstract does not detail how the competitive selection mechanism works or present experimental results, as the text is truncated. The claims about richer representations and preserved alternatives are presented as motivations for the approach.