ATLAS: Attention-Guided Test-Time Latent Selection for Compositional Visual Autoregressive Generation
Visual autoregressive models produce convincing images but can misassign attributes and spatial relations.
ProofPaper ↗
Key points
- We introduce ATLAS, a training-free framework that uses the frozen generator's internal attention to guide search before final images are available.
- Within-prompt diagnostics identify attention-derived scores that predict final compositional preferences on Infinity, with category-dependent transfer to Switti.
- ATLAS uses these on-the-fly signals to progressively concentrate computation on promising trajectories.
- Together, these results support scalable compositional search: additional sampling expands exploration, while attention-guided selection limits continuation cost without parameter updates or gradient-based optimization.
Sources (1)
- [1]ATLAS: Attention-Guided Test-Time Latent Selection for Compositional Visual Autoregressive GenerationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 04:42 PM
Visual autoregressive models produce convincing images but can misassign attributes and spatial relations.
We introduce ATLAS, a training-free framework that uses the frozen generator's internal attention to guide search before final images are available.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
- Oct 17, 2026Estimating the Theoretical Parameter Capacity of a Hypothetical GPT-6-Class Model [D]
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 16, 2026[Research] 2,788 tokens/s on a single RTX 5090 with a diffusion LLM — Introducing ORL (fully open source)
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint