OneLatent: Single-Token Compression for Visual Latent Reasoning

2026-02-14

AI Summary

This work proposes an efficient compressed reasoning paradigm that addresses the substantial increase in output length and computational overhead associated with Chain-of-Thought (CoT) prompting. By compressing intermediate steps of complex visual reasoning into a single latent token, the method leverages supervision signals derived from text-to-image rendering and alignment with DeepSeek-OCR hidden states. Evaluated on ProntoQA and ProsQA, the approach achieves accuracies of 99.80% and 97.80%, respectively, while reducing average output length by 11× (up to 87.4×) with only a 2.21% accuracy drop. Furthermore, it yields a 6.8× improvement in reasoning throughput (OTC), enabling highly efficient, auditable, and low-redundancy inference.

Technology Category

• Computer Vision: Visual Reasoning & Symbolic Representations
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• Chain-of-thought
• reasoning compression
• inference cost
• latent token
• multimodal reasoning

Innovation

• OneLatent
• latent token compression
• chain-of-thought reasoning
• visual supervision
• output token efficiency