Compact flow models improve diverse reasoning over long math problems
Rationale Distribution Distillation via Flow Matching for Distributional Continuous Latent Reasoning
Computation and LanguageArtificial Intelligence
Summary
When language models try to solve hard math problems by explaining their thinking step-by-step, they usually generate long answers that are slow to produce multiple times. The authors propose a way to teach a smaller, faster model to imitate many different detailed explanations all at once. This model uses a special flow method to quickly create varied summaries of reasoning steps, which helps find more correct answers. It works better than older methods that only rely on one explanation at a time.
What this means in practice
- •For ai developers: Create AI that generates multiple diverse reasoning paths quickly, improving answer accuracy on complex questions like advanced math.
- •For automated tutoring platforms: Provide diverse step-by-step explanations generated efficiently to help students understand complex math problems from AI tutors.
Authors
Mingyeong Kim, Gyusang Cho, Juho Lee, Giung Nam
Abstract
Sampling multiple chain-of-thought rationales from a language model can uncover correct solutions missed by a single attempt. However, these rationales are typically lengthy, making repeated autoregressive generation costly. Continuous latent reasoning offers a compact alternative, but point-valued methods condition repeated answer sampling on a single rationale, which can limit exploration of diverse paths and answer coverage. We thus propose rationale distribution distillation (RDD), training a conditional flow over fixed-length latent sequences by distilling from multiple textual reasoning paths. Furthermore, we expand the distillation targets through within-query mixup, interpolating between complete latent rationales for the same query. The flow nonautoregressively generates the entire latent sequence, and a frozen language model autoregressively decodes answers from sampled rationale contexts. Empirically, the resulting distributional reasoner achieves higher answer coverage than the evaluated point-valued latent reasoning baselines, particularly on advanced mathematical tasks.