Research map / Vision-language and multimodal / Multimodal LLMs

AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AdaptVision, a Qwen2.5-VL-7B agent that starts from a 1/4-resolution image and learns via RL (DTPO) to invoke a bounding-box crop tool only when needed, cutting visual tokens to 33% while beating prior efficient VLM methods on nine VQA benchmarks.

Problem

VLMs consume many visual tokens, and existing efficient methods use fixed compression ratios that cannot adapt per sample to task difficulty.

Approach

A coarse-to-fine VLM that answers from low-resolution tokens and calls a crop tool for high-resolution detail only when needed, trained with Decoupled Turn Policy Optimization, which normalizes tool and answer tokens separately and computes separate advantages for tool and outcome rewards.

Key results

• AdaptVision reaches 97.9% of vanilla Qwen2.5-VL-7B average performance using 33% of visual tokens, versus 92.1% for the 1/4-resolution down-sample model at 25% tokens.
• Compared with VisionThink-reproduced dynamic baseline, AdaptVision attains higher average (97.9% vs 95.8%) with fewer tokens (33% vs 52%).
• Reports 1.67x overall inference speedup over the vanilla model and VisionThink†.

Surprising findings

• Vanilla GRPO collapses to calling the tool on 100% of samples, while DTPO learns to call tools selectively by difficulty.
• Replacing GPT-4o with the open Qwen3-VL-4B judge for crop reward still gives competitive results (66.47 RealWorldQA).

Limitations

• Relies on a single bounding-box tool and a fixed 1/4-resolution initial compression.
• Reasoning limited to at most two turns, which may restrict deep visual reasoning.
• Crop reward depends on GPT-4o judge, adding API cost to training.

Experimental setup

Datasets: VisionThink-Smart-Train

Benchmarks: ChartQA OCRBench DocVQA MME MMVet RealWorldQA POPE MathVista MathVerse

Metrics: accuracy visual token ratio (#Token) average relative performance inference latency

Baselines: Vanilla Qwen2.5-VL-7B-Instruct Down-Sample FastV SparseVLM VisionZip VisionThink GRPO

Builds on: GRPO Qwen2.5-VL VisionThink veRL

Closest papers by meaning

• Rethinking Token Reduction for Large Vision-Language Models (CVPR 2026 · similarity 0.92)
• TTRV: Test-Time Reinforcement Learning for Vision Language Models (CVPR 2026 · similarity 0.91)
• Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow (CVPR 2026 · similarity 0.90)
• VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions (CVPR 2026 · similarity 0.90)
• LLMind: Bio-inspired Training-free Adaptive Visual Representations for Vision-Language Models (CVPR 2026 · similarity 0.90)
• Towards Self-Refinement of Vision-Language Models with Triangular Consistency (NeurIPS 2025 · similarity 0.90)