AVIS: New Method Optimizes Inference Efficiency for Vision-Language Models
Researchers introduced AVIS, a technique that adaptively scales both visual context and reasoning computation in vision-language models to reduce inference costs while maintaining accuracy. The method uses key diversity pruning to remove redundant visual tokens and a learned difficulty predictor to determine how much reasoning is needed per query. This addresses a key challenge in deploying large vision-language models by improving the accuracy-compute trade-off.
AVIS (Adaptive Visual Inference Scaling) is a lightweight policy designed to optimize how vision-language models allocate computational resources during inference. The technique operates along two coupled axes: Visual Context Scaling (VCS), which determines how much visual information reaches the language model, and Visual Reasoning Scaling (VRS), which controls the amount of inference-time reasoning performed. The method implements VCS through Key Diversity Visual pruning—a training-free algorithm that removes redundant visual tokens before processing—and realizes VRS through adaptive self-consistency guided by a learned difficulty predictor. Testing across image and video reasoning benchmarks shows AVIS improves the accuracy-compute trade-off compared to methods optimizing only one axis, and it remains effective when applied to reinforcement learning post-trained models while keeping computational and latency costs low.
What's missing
The paper does not discuss potential limitations of the difficulty predictor's generalization to out-of-distribution queries, nor does it provide detailed comparisons with other recent adaptive inference methods for vision-language models beyond VCS-only and VRS-only baselines.
What different sources said
- arXiv cs.AICenter
AVIS: Adaptive Test-Time Scaling for Vision-Language Models
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.