Research map / Vision-language and multimodal / Multimodal LLMs
On Locality and Length-Generalization in Visual Reasoning
Pulkit Madan, Sanjay Haresh, Reza Ebrahimi, Apratim Bhattacharyya, Sunny Panchal, Roland Memisevic
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local high-res glimpses to global-view VLM (ablation) | Qwen-VL · Finding Zeros OOD-numroots | accuracy (%): 32.63 -> 67.12 | Improved |
| add local glimpses to global-view VLM (ablation) | Qwen-VL · Finding Zeros In-distribution | accuracy (%): 57.24 -> 82.26 | Improved |
| add local glimpses (multi-function split) (ablation) | Qwen-VL · Finding Zeros multi-function OOD-numroots | accuracy (%): 1.06 -> 63.66 | Improved |
| raise global glimpse resolution from 600x400 to 1200x800 with local glimpses (ablation) | Foveated Qwen-VL · Finding Zeros OOD-subplots+numroots | accuracy (%): 59.70 -> 57.46 | Hurt |
| replace global full image with local+peripheral glimpses (recurrent model) (ablation) | Recurrent foveation model · Visual Parity / State Machine OOD | success rate | Improved |
| use nonlinear recurrent backbone instead of transformer, Mamba, xLSTM on local glimpses (ablation) | Recurrent foveation model · State Machine OOD | success rate | Improved |
| fine-tune global-view Qwen-VL on Visual Parity/State Machine | Qwen-VL · Visual Parity / State Machine OOD | success rate | Hurt |
| vary peripheral glimpse resolution and size (ablation) | Recurrent foveation model · Visual Parity | accuracy / steps | Mixed |
| use global model instead of foveated model on conjunctive search (recall) | Qwen-VL · Conjunctive Search (recall) | accuracy | Improved |
| add black-dot marks on visited switches as external memory (ablation) | Recurrent foveation model · Visual Parity without arrows | trainability | Improved |
| uniformly raise global resolution at matched compute (ablation) | Qwen-VL · Finding Zeros | accuracy (%) | No clear effect |
Closest papers by meaning
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.88)
- Mechanisms of Object Localization in Vision-Language Models (CVPR 2026 · similarity 0.88)
- VLMs have Tunnel Vision: Evaluating Nonlocal Visual Reasoning in Leading VLMs (NeurIPS 2025 · similarity 0.87)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.87)
- In-Context Learning Can Help Vision Language Models Overcome Training Prior (NeurIPS 2026 · similarity 0.87)
- Understanding the Limits of Vision Language Models Through the Lens of the Binding Problem (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
