AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Decomposition of Concept-Level Rules in Visual Scenes

Fan Shi, Yuxuan Liang, Xiaolei Chen, Haiyang Yu, Xu Li, Yi Zheng, Rui Zhu, Xiangyang Xue, Bin Li

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

Read the paperCheck your idea against 137,217 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 CRD, a framework that decomposes visual scenes into independent concepts and the rules that change them, aiming to do so without hand-crafted inductive biases or human-designed priors.

Problem

Vision-language systems process images holistically with limited explicit decomposition of concepts and concept-changing rules; prior decomposition methods depend on hand-crafted inductive biases or human-designed priors.

Approach

A framework (CRD) that parses visual scenes into independent concepts and their concept-changing rules, learned without hand-crafted inductive biases or human-designed priors, in contrast to holistic processing and prior prior-dependent decomposition methods.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CRD (concept-rule decomposition with LVLM-MH sampling) to the LVLMInternVL-3.5-8B · VSB-MAAvg. Similarity: 59.9 -> 64.0Improved
add CRD to Qwen2.5-VL-7BQwen2.5-VL-7B · I-RAVENaccuracy (%): 15.0 -> 89.3Improved
add CRD to Qwen2.5-VL-7BQwen2.5-VL-7B · RAVENaccuracy (%): 59.7 -> 89.4Improved
add CRD to InternVL-3.5-8BInternVL-3.5-8B · RAVEN / I-RAVENaccuracy (%): RAVEN 11.6 -> 31.6 (InternVL3-8B vs InternVL-CRD)Improved
remove the LVLM proposal ratio from the acceptance probability (ablation)CRD (InternVL-3.5-8B) · VSB-MAAvg. Similarity: 64.0 -> 61.0Hurt
remove the CRF score term from the acceptance probability (ablation)CRD (InternVL-3.5-8B) · VSB-MAAvg. Similarity: 64.0 -> 59.2Hurt
add patch-level scores on top of meta-attributes as prompt (ablation)Qwen2.5-VL-7B + CRD · SpatialEval overalloverall accuracy: 61.45 -> 63.37Improved
add CRD-meta prompts to the LVLMQwen2.5-VL-7B · SpatialEval overalloverall accuracy: 60.11 -> 61.45Improved
add CRD-meta prompts to the LVLMQwen2.5-VL-3B · SpatialEval SpatialMapaccuracy: 50.93 -> 50.93No clear effect
add CRD at inference timeInternVL-3.5-8B · VSB-MA latencylatency per token (ms): 77.3 -> 234.3Hurt
increase patch grid from 2x2 to 4x4 (ablation)CRD (InternVL-3.5-8B) · VSB-MA latencylatency per token (ms): 234.3 -> 239.5No clear effect
add CRD on human-annotated VSB-MA subsetInternVL-3.5-8B · VSB-MA (human-annotated)Avg. Similarity: 58.56 -> 69.22Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy