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

2026-05-14 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

ATLAS is a novel framework designed for visual reasoning that integrates both agentic and latent reasoning capabilities using a single discrete "functional token." This token acts as both an agentic operation and a latent visual reasoning unit, associated with an internalized visual operation without requiring explicit visual supervision. The framework avoids the computational expense and architectural complexity of directly generating intermediate images, and mitigates context-switching latency of external execution in agentic methods, while improving task generalization over latent methods. ATLAS is compatible with standard SFT and RL training, requiring no architectural or methodological modifications. To address functional token sparsity during RL, it introduces Latent-Anchored GRPO (LA-GRPO), which stabilizes training with a statically weighted auxiliary objective for stronger gradient updates. Experiments show ATLAS achieves superior performance on challenging benchmarks with clear interpretability.

Key takeaway

For research scientists developing visual reasoning models, ATLAS presents a compelling alternative to computationally expensive image generation or context-switching agentic methods. You should investigate integrating functional tokens and Latent-Anchored GRPO into your next-token prediction models to achieve superior performance and interpretability without complex architectural changes.

Key insights

ATLAS unifies agentic and latent visual reasoning via a single functional token, enhancing efficiency and interpretability.

Principles

• Unify agentic and latent reasoning.
• Avoid explicit intermediate visual generation.
• Stabilize RL with auxiliary objectives.

Method

ATLAS uses a functional token for both agentic and latent visual operations, trained with standard SFT/RL, and stabilized by Latent-Anchored GRPO (LA-GRPO) for functional token sparsity.

In practice

• Integrate functional tokens into existing tokenizers.
• Apply LA-GRPO for sparse token training.
• Leverage ATLAS for complex visual reasoning tasks.

Topics

• ATLAS Framework
• Visual Reasoning
• Functional Tokens
• Agentic Reasoning
• Latent Reasoning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Spatial Atlas: Compute-Grounded Reasoning for Spatial-Aware Research Agent Benchmarks — Compute-grounded reasoning enhances AI agent reliability by resolving deterministic sub-problems before LLM generation.
• 😺 Watch: Elorian wants to fix AI's toddler vision — Current AI lacks human-like visual reasoning, hindering agentic software development and physical world applications.
• Unlocking the Working Memory of Large Language Models for Latent Reasoning — Large language models can be trained to use working memory for efficient, latent reasoning, decoupling internal computation from external communication.
• Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning — Visual-OPSD distills reasoning from expensive visual thought generation pathways into efficient text-only models, achieving substantial speedup and performance gains.
• Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning — Gains in latent visual reasoning stem from boundary markers and format, not visual memory encoded in latent slots.
• A Mathematical Forum Platform for Collaborative Problem Solving and Dataset Generation for AI Reasoning — Integrating image-to-LaTeX conversion into forums creates a seamless user experience and generates valuable AI training data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.