OneLatent: Single-Token Compression for Visual Latent Reasoning

TL;DR

OneLatent introduces a method to compress intermediate reasoning steps into a single latent token using image rendering and OCR supervision, significantly reducing output length and inference cost while maintaining high accuracy.

Contribution

It proposes a novel single-token latent reasoning framework that leverages rendered images and OCR supervision to efficiently condense reasoning processes.

Findings

Reduces output length by 11 times with minimal accuracy loss

Achieves up to 87.4 times compression on reasoning tasks

Maintains high accuracy (over 97%) with single latent tokens on logical reasoning benchmarks

Abstract

Chain-of-thought (CoT) prompting improves reasoning but often increases inference cost by one to two orders of magnitude. To address these challenges, we present \textbf{OneLatent}, a framework that compresses intermediate reasoning into a single latent token via supervision from rendered CoT images and DeepSeek-OCR hidden states. By rendering textual steps into images, we obtain a deterministic supervision signal that can be inspected and audited without requiring the model to output verbose textual rationales. Across benchmarks, OneLatent reduces average output length by $11\times$ with only a $2.21\%$ average accuracy drop relative to textual CoT, while improving output token contribution (OTC) by $6.8\times$. On long-chain logical reasoning, OneLatent reaches $99.80\%$ on ProntoQA and $97.80\%$ on ProsQA with one latent token, with compression up to $87.4\times$, supporting compression-constrained generalization.