What Matters for Latent Reasoning with Flow Matching
FLaRe is presented as a flow-matching recipe for latent reasoning that nearly matches explicit chain-of-thought accuracy with much lower latency.
The paper argues latent thoughts should be useful, diverse, explainable, refinable with more compute, and efficient. It says many current methods miss those marks by learning shortcuts, absorbing explicit CoT into weights, or imitating CoT token by token. Its proposed Flow-based Latent Reasoning method specifies how to encode and shape the latent space, train the flow, read out answers, and refine on verified model-generated thoughts. The authors report gains over prior latent methods on all five probes and 97% of explicit CoT accuracy at one quarter of the latency. HF Daily Papers' note
The paper argues latent thoughts should be useful, diverse, explainable, refinable with more compute, and efficient. It says many current methods miss those marks by learning shortcuts, absorbing explicit CoT into weights, or imitating CoT token by token. Its proposed Flow-based Latent Reasoning method specifies how to encode and shape the latent space, train the flow, read out answers, and refine on verified model-generated thoughts. The authors report gains over prior latent methods on all five probes and 97% of explicit CoT accuracy at one quarter of the latency. HF Daily Papers' note
score 5