TikArt: Stabilizing Aperture-Guided Fine-Grained Visual Reasoning with Reinforcement Learning

TL;DR

TikArt introduces an aperture-guided agent that enhances fine-grained visual reasoning in multimodal models by sequentially focusing on regions of interest, improving interpretability and accuracy in complex visual tasks.

Contribution

It proposes a novel aperture-guided reasoning framework with a TAO loop, integrating reinforcement learning and a new reward to stabilize long-horizon evidence collection.

Findings

Significant improvements in high-resolution reasoning tasks.

Enhanced performance in multimodal understanding and segmentation.

Effective transfer to pixel-level grounding tasks.

Abstract

Fine-grained visual reasoning in multimodal large language models (MLLMs) is bottlenecked by single-pass global image encoding: key evidence often lies in tiny objects, cluttered regions, subtle markings, or dense charts. We present \textbf{TikArt} (\textbf{T}h\textbf{i}n\textbf{k}ing \textbf{A}pe\textbf{rt}ure), an aperture-guided agent that formulates multimodal reasoning as sequential evidence acquisition over regions of interest. TikArt follows a Think--Aperture--Observe (TAO) loop that interleaves language reasoning with two aperture actions: Zoom, which extracts rectangular crops, and Segment, which invokes an off-the-shelf segmenter to produce object-centric mask-based views for irregular targets. A mandatory Observation step after every aperture action writes local evidence back into text, yielding interpretable aperture trajectories and persistent linguistic memory. Built on Qwen3-VL-8B, TikArt is trained with GRPO-style reinforcement learning under a two-stage curriculum. To stabilize long-horizon tool-integrated learning, we introduce Relative Uncertainty Reduction (RUR), a dense reward computed by a frozen evaluator that favors evidence-building trajectories and mitigates degenerate tool use. Experiments on high-resolution reasoning, general multimodal understanding, and both referring and reasoning-oriented segmentation show consistent gains over the backbone, demonstrating that aperture-guided observation improves fine-grained visual reasoning and transfers naturally to pixel-level grounding.