On a 100-question GSM8K timing test, a Llama-3.2-1B-Instruct-based system using Flow-based Latent Reasoning (FLaRe) nearly matched the accuracy of explicit chain-of-thought while taking about a quarter as long. With two denoising steps, stage-two FLaRe scored 57.4% and had a median latency of 68 milliseconds; the written-reasoning baseline scored 59.3% at 267 ms in the same sample.
Written solutions are compressed into eight latent slots during training. At inference, FLaRe does not receive or output a visible chain of thought: given a question, it starts with random noise and gradually refines it into a latent thought, or compact internal working state, then uses that state to produce the answer. A separate decoder can render the thought as text for analysis, but that written version is not part of the direct-answer path.
In a second training stage, the researchers had FLaRe propose several thoughts for new problems with known answers. They decoded those thoughts, kept the ones whose final answer matched the reference, and trained the model again on those accepted examples. The paper reports that this stage improved its scores across the three model sizes tested on GSM8K.
The timing result has a narrow basis: the authors measured one question at a time on a single RTX 3090 in bfloat16 and reported the median across 100 questions. In a separate GSM8K table for Llama-3.2-1B-Instruct, stage-two FLaRe scored 59.1%, below 62.5% for explicit chain-of-thought; on the SVAMP arithmetic set, the scores were 58.8% and 66.5%. The findings come from an arXiv preprint submitted Oct. 5, not a deployment test, and do not establish an advantage on general-purpose reasoning workloads.
Sources
- FLaRe preprint abstract and submission record
- Full paper, including Tables 8 and 20 and Appendix E.5
Image credit: Smorgasbord, “Math answers with a pen and calculator,” licensed under CC BY-SA 3.0.
