AI Paper Atlas

Research map / Large language models / Alignment and preferences

SMAP: Semantic Route Planning with Map-Grounded Multimodal Alignment

Wenjie Zhang, Chen Yang, Xin Lu, Zhen Wang, Yue Liu, Bobo Xi, Pengbo Zhang

CVPR 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add SFT plus HDPO post-training to the base modelQwen2.5-VL-32B · MM-Route testRDR: 0.667 -> 0.831Improved
add SFT plus HDPO post-training to the base modelQwen2.5-VL-32B · MM-Route testSHR (%): 31.9 -> 14.0Improved
add SFT plus HDPO post-training to the base modelQwen2.5-VL-32B · MM-Route testTSPR (%): 34.5 -> 77.5Improved
add HDPO on top of SFT (ablation)Qwen2.5-VL-32B-SFT · MM-Route testRDR: 0.789 -> 0.831Improved
add HDPO on top of SFT (ablation)Qwen2.5-VL-7B-SFT · MM-Route testORS: 3.00 -> 3.37Improved
use standard DPO (ground-truth accepted, SFT output rejected) instead of verifier-refined HDPO pairs (ablation)Qwen2.5-VL-7B-SFT · MM-Route testRDR: 0.763 -> 0.733Hurt
use standard DPO instead of SFT (ablation)Qwen2.5-VL-7B-SFT · MM-Route testSHR (%): 22.1 -> 27.6Hurt
use standard DPO instead of SFT (ablation)Qwen2.5-VL-32B-SFT · MM-Route testRDR: 0.789 -> 0.765Hurt
use map tile (text-image) instead of text-only input with HDPO (ablation)Qwen2.5-VL-32B-HDPO · MM-Route testRDR: 0.789 -> 0.831Improved
use map tile instead of text-only input with HDPO (ablation)Qwen2.5-VL-32B-HDPO · MM-Route testTSPR (%): 73.5 -> 77.5Improved
use map tile instead of text-only input with HDPO (ablation)Qwen2.5-VL-7B-HDPO · MM-Route testORS: 3.39 -> 3.37No clear effect
use map tile instead of text-only input with SFT (ablation)Qwen2.5-VL-7B-SFT · MM-Route testRTPR (%): 84.8 -> 83.8Mixed

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