AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Efficient Multimodal Spatial Reasoning via Dynamic and Asymmetric Routing

Yixian Shen, Qi Bi, Zihan Wang, Zhiheng Yang, Changshuo Wang, Zhi Zhang, Prayag Tiwari, Andy D. Pimentel, Anuj Pathania

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

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 an efficient framework for multimodal spatial reasoning with visualization-of-thought that reduces the cost of lengthy, redundant visual reasoning tokens via dynamic and asymmetric routing.

Problem

Visualization-of-thought (VoT) improves spatial reasoning in MLLMs, but autoregressive accumulation of lengthy and redundant tokens greatly increases computation and memory costs.

Approach

Uses dynamic and asymmetric routing to make multimodal spatial reasoning more efficient, as named in the title; the abstract does not specify the mechanism in detail.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add DARE-LH hop- and layer-aware token routing to VolCanoVolCano · VSRaccuracy / FLOPs: 67.18 -> 68.09 acc; 19842.37 -> 11310.63 GFLOPsImproved
add DARE-LH routing to VolCanoVolCano · Winogroundaccuracy: 68.37 -> 68.31No clear effect
add DARE-LH routing to VolCanoVolCano · EmbSpatialaccuracy: 58.29 -> 68.09Improved
add DARE-LH routing to Anole-7B (vs VoT)Anole-7B · MAZEaccuracy / FLOPs: 86.56 -> 93.32; 25640.14 -> 12739.67 GFLOPsImproved
add DARE-L to VolCanoVolCano · GQAaccuracy: 64.40 -> 64.71No clear effect
add DARE-L to VolCanoVolCano · POPEaccuracy: 86.50 -> 87.27Improved
add hard visual pruning loss in deeper layers (DARE-L vs UniPrune) (ablation)DARE · VSRaccuracy: 63.71 -> 67.13Improved
add hop-aware retention on top of layer-wise (DARE-LH vs DARE-L) (ablation)DARE · VSRaccuracy: 67.13 -> 68.09Improved
add hop-aware retention (DARE-LH vs DARE-L) (ablation)DARE · MAZEaccuracy: 89.78 -> 93.32Improved
use DARE-LH KV-cache retention instead of full cacheVolCano · VSRKV cache: 5.45 GB -> 2.97GBImproved
reserve prefix of 2 tokens per hop instead of 0 (ablation)DARE-LH · VSRaccuracy: 64.12 -> 68.09Improved
increase prefix size beyond 2 (ablation)DARE-LH · VSRaccuracy / KV cache: 68.09 -> 68.09 acc; KV 2.97 -> 3.40 (kappa=4)No clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Builds on: visualization-of-thought

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