AdaLoop: Thinking with the Vision Encoder
We present AdaLoop, a framework for question-conditioned visual recurrence in VLMs. AdaLoop adapts visual computation to the question and reuses the encoder's transformations across multiple iterations, with the pretrained vision weights frozen.
ProofPaper ↗
Key points
- Most vision-language models (VLMs) treat the vision encoder as a static feature extractor: it encodes an image once, independently of the question, and leaves subsequent reasoning to the language model.
- Yet different questions may require different visual operations and processing depths.
- We also introduce VisRO, a procedural dataset and generation pipeline for training and evaluating vision-centric reasoning.
- Our findings suggest that pretrained vision encoders can contribute to reasoning through task-directed, iterative visual computation.
Sources (1)
- [1]AdaLoop: Thinking with the Vision EncoderarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 06:14 AM
We present AdaLoop, a framework for question-conditioned visual recurrence in VLMs. AdaLoop adapts visual computation to the question and reuses the encoder's transformations across multiple iterations, with the pretrained vision weights frozen.
Most vision-language models (VLMs) treat the vision encoder as a static feature extractor: it encodes an image once, independently of the question, and leaves subsequent reasoning to the language model.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026Noise-Space Reward Alignment for Few-Step Generative Models
- Oct 15, 2026Welcome FLUX.3 Image Klein in Diffusers
- Oct 15, 20263DCodeVerse: Exploring 3D Worlds as Code
- Oct 14, 2026Active Generative Simulation: Meta-Learning What to Simulate for Sim-and-Real Policy Co-Training
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
- Oct 13, 2026PairIT: Autoregressive Transformers for Low-Data Molecule Optimization