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

Hugging Face Daily Papers Papers
visual-reasoning functional-tokens latent-representation agentic-reasoning reinforcement-learning next-token-prediction

Summary

ATLAS presents a visual reasoning framework that combines agentic operations and latent representations using functional tokens, enabling efficient training via next-token prediction and reinforcement learning while avoiding intermediate image generation.

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 and a latent visual reasoning unit. Each functional token is associated with an internalized visual operation, yet requires no visual supervision and remains a standard token in the tokenizer vocabulary, which can be generated via next-token prediction. This design avoids verbose intermediate visual content generation, while preserving compatibility with the vanilla scalable SFT and RL training, without architectural or methodological modifications. To further address the sparsity of functional tokens during RL, we introduce Latent-Anchored GRPO (LA-GRPO), which stabilizes the training by anchoring functional tokens with a statically weighted auxiliary objective, providing stronger gradient updates. Extensive experiments and analyses demonstrate that ATLAS achieves superior performance on challenging benchmarks while maintaining clear interpretability. We hope ATLAS offers a new paradigm inspiring future visual reasoning research.
Original Article
View Cached Full Text

Cached at: 05/15/26, 04:23 AM

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

Source: https://huggingface.co/papers/2605.15198

Abstract

ATLAS presents a visual reasoning framework that combines agentic operations and latent representations using functional tokens, enabling efficient training and improved performance on complex benchmarks.

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 includeagentic reasoningthrough 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 withautoregressive 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 and alatent visual reasoningunit. Each functional token is associated with an internalized visual operation, yet requires no visual supervision and remains a standard token in the tokenizer vocabulary, which can be generated vianext-token prediction. This design avoids verbose intermediate visual content generation, while preserving compatibility with the vanillascalable SFTandRL training, without architectural or methodological modifications. To further address the sparsity offunctional tokensduring RL, we introduceLatent-Anchored GRPO(LA-GRPO), which stabilizes the training by anchoringfunctional tokenswith a statically weightedauxiliary objective, providing strongergradient updates. Extensive experiments and analyses demonstrate that ATLAS achieves superior performance on challenging benchmarks while maintaining clear interpretability. We hope ATLAS offers a new paradigm inspiring futurevisual reasoningresearch.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2605\.15198

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/2605.15198 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.15198 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.15198 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

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.

Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Hugging Face Daily Papers

This paper introduces Beacon, an agentic visual reasoning model that improves multimodal LLMs' ability to decide when to use tools and benefit from tool use, using Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion in reinforcement learning.

Agentic Graph Token Reasoning

arXiv cs.LG

Introduces agentic graph token reasoning, a method that recasts graph tokenization as part of an LLM's step-by-step reasoning process, allowing the model to dynamically choose graph views and granularity during inference. The method outperforms baselines across seven graph domains and transfers zero-shot to unseen domains.