Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·20h ago
AI Research
arXiv CS.AI·Jun 11·4 min read
AVIS: Adaptive Test-Time Scaling for Vision-Language Models
TL;DR
Researchers identified a gap in LLM evaluation benchmarks. They built a synthetic dataset with 10k adversarial prompts targeting reasoning failures.
✦ Why It Matters
Use this benchmark to audit LLM robustness before deploying in production reasoning pipelines.
Key Takeaways
- Utilize AVIS for improved efficiency in VLMs.
- Implement KDV pruning to reduce redundant visual tokens.
- Leverage adaptive self-consistency for optimized reasoning.
- Achieve better accuracy-compute trade-offs in real-time tasks.
How It Works
AVIS operates by adapting both VCS and VRS for each query. VCS is achieved through KDV pruning, which eliminates unnecessary visual tokens before processing, thus streamlining the input for the language model.
VRS is enhanced by a learned difficulty predictor that determines how many reasoning rollouts are needed, allowing the model to focus computational resources where they are most effective.