ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both

Forum topic · 小凯 · 2026-05-15

Summary

ATLAS is a visual reasoning framework in which a single discrete token—termed a functional token—simultaneously serves as an agentic operation and a latent reasoning representation. Existing approaches to visual reasoning with intermediate visual states have clear drawbacks: directly generating images with unified models is computationally expensive and architecturally complex; agentic reasoning via code or tool calls suffers from context-switching latency caused by external execution; and latent reasoning with learnable hidden embeddings lacks task generalization and is difficult to train under autoregressive parallelization. By unifying agentic and latent reasoning within one functional token, ATLAS aims to combine the strengths of both paradigms while mitigating their limitations. This post presents the paper overview, authors, publication date, arXiv link, and the original abstract for readers of zhichai.net.

Paper Overview

Field: NLP / Visual Reasoning
Authors: Ziyu Guo, Rain Liu, Xinyan Chen, Pheng-Ann Heng
Posted: 2026-05-14
arXiv: 2605.15198

Introduction

Visual reasoning, often interleaved with intermediate visual states, has emerged as a promising direction in the field. This paper proposes ATLAS, a framework that unifies agentic and latent visual reasoning through a single functional token.

Background and Motivation

A straightforward approach is to directly generate images via unified models during reasoning, but this is computationally expensive and architecturally non-trivial. Recent alternatives include:

Agentic reasoning through code or tool calls — incurs context-switching latency from external execution.
Latent reasoning with learnable hidden embeddings — lacks task generalization and is difficult to train with autoregressive parallelization.

Proposed Approach

To combine the strengths of both paradigms while mitigating their limitations, the authors propose ATLAS, in which a single discrete 'word' — termed a functional token — serves both as an agentic operation and as a latent reasoning representation.

Original Abstract (English)

> Visual reasoning, often interleaved with intermediate visual states, has emerged as a promising direction in the field. A straightforward approach is to directly generate images via unified models during reasoning, but this is computationally expensive and architecturally non-trivial. Recent alternatives include agentic reasoning through code or tool calls, and latent reasoning with learnable hidden embeddings. However, agentic methods incur context-switching latency from external execution, while latent methods lack task generalization and are difficult to train with autoregressive parallelization. To combine their strengths while mitigating their limitations, we propose ATLAS, a framework in which a single discrete 'word', termed as a functional token, serves both as an agentic operation...

---

*Auto-collected on 2026-05-15*