Paper Overview
Field: NLP Authors: Ziyu Guo, Rain Liu, Xinyan Chen, Pheng-Ann Heng Published: 2026-05-14 arXiv: 2605.15198
Abstract
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, the authors propose ATLAS, a framework in which a single discrete "word", termed a functional token, serves both as an agentic operation and a latent reasoning carrier.
Key Ideas
- Problem space: intermediate visual states during reasoning can be produced via image generation, agentic tool/code calls, or latent embeddings—each with distinct costs.
- Agentic reasoning: powerful but suffers latency from external execution and context switching.
- Latent reasoning: efficient but limited in task generalization and difficult to train in an autoregressive parallel setting.
- ATLAS proposal: a functional token—one discrete word—that plays a dual role, unifying agentic operation and latent reasoning in a single framework.
*Auto-collected on 2026-05-15. Original abstract truncated in source; see the arXiv link above for the full paper.*