AVIS paper accepted at NeurIPS 2026: jointly scaling visual context and reasoning per query
CSProfKGD · x · 2026-09-25
AVIS: Adaptive Visual Inference Scaling for Vision-Language Models has been accepted at NeurIPS 2026, by a team of 11 authors including Ahmadreza Jeddi.
Key idea: existing test-time scaling for VLMs optimizes a single axis of compute — either scaling visual context or doing more reasoning rollouts. AVIS adaptively allocates compute across both coupled axes on a per-query basis:
- Visual Context Scaling (VCS) via Key Diversity Visual (KDV) pruning — a training-free O(N) key-based rule that removes redundant visual tokens before prefilling;
- Visual Reasoning Scaling (VRS) via adaptive self-consistency, with a learned difficulty predictor choosing the number of reasoning rollouts.
It's deployment-friendly: all rollouts share a single prefill pass and reuse the KV cache. Across diverse image and video reasoning benchmarks, AVIS achieves a better accuracy-compute trade-off than single-axis scaling methods. Paper, code and website are public.
More from Models
- OpenAI reportedly prepping GPT-6 Cyber security model for DevDay — emmanuelvivier · 2026-09-25
- Anthropic Publishes Official Prompting Guide for Claude Opus 5.5: Calibrate Effort, Don't Max It — CodeByPoonam · 2026-09-25
- OrcaSAQ-2-27B Trends on Hugging Face: 3-Bit Mixed-Precision Qwen3-Based Reasoning Model — orcarouter · 2026-09-25
- A 1978 logic puzzle book is catching chatbots that memorize instead of reason — JafarNajafov · 2026-09-25
- "I feel the AGI": developer says Opus 5.5 cracked video editing, and he's the bottleneck — LeviTurk · 2026-09-25
- Not every task needs frontier models: local Qwen 4 27B is pulling users away — haider1 · 2026-09-25