Latent Reasoning with Normalizing Flows

Hugging Face Daily Papers Papers
latent-reasoning normalizing-flows chain-of-thought large-language-models code-generation probabilistic-modeling

Summary

Proposes NF-CoT, a latent reasoning framework using normalizing flows to model continuous thoughts in LLMs, preserving autoregressive advantages and achieving better code generation performance with lower cost.

Large language models often improve reasoning by generating explicit chain-of-thought (CoT), demonstrating the importance of intermediate computation. However, textual CoT forces this computation through a discrete, serial, and communication-oriented token stream: each reasoning step must be verbalized before the model can proceed, even when the underlying update is semantic, uncertain, or only partially formed. Latent reasoning offers a higher-bandwidth alternative by performing intermediate computation in compact continuous states before committing to text. Yet existing latent-reasoning methods often sacrifice key advantages that make CoT effective in autoregressive language models, including native left-to-right generation, probabilistic sampling, compatibility with KV-cache decoding, and tractable likelihood estimation. We propose NF-CoT, a latent reasoning framework that preserves these advantages by modeling continuous thoughts with normalizing flows. NF-CoT instantiates a TARFlow-style normalizing flow inside the LLM backbone, defining a tractable probability model over compact continuous thoughts distilled from explicit CoT. Continuous-thought positions are generated by an NF head, while text positions are generated by the standard LM head within the same causal stream. This design provides exact likelihoods for latent thoughts, enables probabilistic left-to-right decoding with the original KV cache, and supports direct policy-gradient optimization in the latent reasoning space. On code-generation benchmarks, NF-CoT improves pass rates over explicit-CoT and prior latent-reasoning baselines while substantially reducing intermediate-reasoning cost.
Original Article
View Cached Full Text

Cached at: 06/05/26, 06:07 AM

Paper page - Latent Reasoning with Normalizing Flows

Source: https://huggingface.co/papers/2606.06447

Abstract

Latent reasoning framework using normalizing flows preserves autoregressive generation advantages while enabling efficient, probabilistic intermediate computation in large language models.

Large language models often improve reasoning by generating explicitchain-of-thought(CoT), demonstrating the importance of intermediate computation. However, textual CoT forces this computation through a discrete, serial, and communication-oriented token stream: each reasoning step must be verbalized before the model can proceed, even when the underlying update is semantic, uncertain, or only partially formed.Latent reasoningoffers a higher-bandwidth alternative by performing intermediate computation in compact continuous states before committing to text. Yet existing latent-reasoning methods often sacrifice key advantages that make CoT effective in autoregressive language models, including native left-to-right generation,probabilistic sampling, compatibility withKV-cache decoding, and tractablelikelihood estimation. We propose NF-CoT, alatent reasoningframework that preserves these advantages by modeling continuous thoughts withnormalizing flows. NF-CoT instantiates aTARFlow-style normalizing flow inside the LLM backbone, defining a tractable probability model over compact continuous thoughts distilled from explicit CoT. Continuous-thought positions are generated by an NF head, while text positions are generated by the standard LM head within the same causal stream. This design provides exact likelihoods for latent thoughts, enables probabilistic left-to-right decoding with the original KV cache, and supports directpolicy-gradient optimizationin thelatent reasoningspace. Oncode-generation benchmarks, NF-CoT improves pass rates over explicit-CoT and prior latent-reasoning baselines while substantially reducing intermediate-reasoning cost.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2606\.06447

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.06447 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.06447 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.06447 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces

arXiv cs.CL

Introduces ReasoningFlow, a framework to capture discourse structures of large language model reasoning traces as directed acyclic graphs, enabling fine-grained analysis of reasoning behaviors like self-reflection and backtracking. Based on manual and automatic annotation of thousands of traces, it reveals structural similarities across models and that most erroneous steps do not contribute to final answers.

Adaptive Latent Agentic Reasoning

arXiv cs.CL

This paper introduces Adaptive Latent Agentic Reasoning (ALAR), a dual-mode framework for LLM agents that uses compact latent reasoning for routine turns and selectively escalates to explicit chain-of-thought for harder decisions, achieving up to 84.6% token reduction while maintaining task accuracy.

Tools as Continuous Flow for Evolving Agentic Reasoning

arXiv cs.AI

This paper introduces FlowAgent, a novel framework that reconceptualizes tool chaining as continuous trajectory generation using conditional flow matching to improve robustness in long-horizon agentic reasoning.

NoisyCoconut: Counterfactual Consensus via Latent Space Reasoning

arXiv cs.LG

This paper introduces NoisyCoconut, an inference-time method that improves LLM reliability by injecting noise into latent trajectories to generate diverse reasoning paths. The approach enables models to abstain when uncertain, significantly reducing error rates in mathematical reasoning tasks without requiring retraining.