HF Paper 'What Matters for Latent Reasoning' Introduces FLaRe, Hits 97% CoT Accuracy at 4x Speedup

October 6 paper introduces FLaRe (Flow-based Latent Reasoning), a method for LLMs to reason in continuous latent space rather than token-by-token chains of thought. The authors identify five properties — useful, diverse, explainable, refinable, efficient — and show FLaRe reaches 97% of explicit CoT accuracy on GSM8K at one-quarter of the latency, roughly a 3.9x speedup through a two-stage VAE-plus-flow-matching recipe.