Research map / Vision-language and multimodal / Multimodal LLMs
STEP-Nav: Spatial-Temporal Efficient Visual Token Pruning for Vision-and-Language Navigation with Large Language Models
Yantao Lu, Shiqi Sun, Ning Liu, Bo Jiang, Ying Zhang, Jinchao Chen, Chenglie Du
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 STEP-Nav, a spatial-temporal visual token pruning method to reduce the computational cost of LLM-based vision-and-language navigation.
Problem
LLM-based vision-and-language navigation encodes image sequences as dense visual tokens, incurring substantial computational overhead.
Approach
Prunes visual tokens across both spatial and temporal dimensions of the navigation image sequence to cut LLM input cost while preserving navigation accuracy (based on the title; details not available in the abstract provided).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use STEP-Nav instead of ToMe at 88.9% pruning | Navid-CE · VLN-CE R2R val-unseen | OS: 36.8 -> 46.2 | Improved |
| use STEP-Nav instead of ToMe at 88.9% pruning | Navid-CE · VLN-CE R2R val-unseen | SPL: 25.9 -> 33.2 | Improved |
| use STEP-Nav instead of VisionZip at 88.9% pruning | Navid-CE · VLN-CE R2R val-unseen | SPL and SR: +0.3% SPL, +2.5% SR | Improved |
| prune 66.7% of tokens with STEP-Nav vs full tokens | Navid-CE · VLN-CE R2R val-unseen | SR: 37.4 -> 36.4 | No clear effect |
| prune 88.9% of tokens with STEP-Nav vs full tokens | Navid-CE · VLN-CE R2R val-unseen | SPL: 35.9 -> 33.2 | Hurt |
| use STEP-Nav instead of ToMe at 88.9% pruning | NavGPT-2 FlanT5-XL · R2R val-unseen | SR: 52.6 -> 65.6 | Improved |
| use STEP-Nav instead of VisionZip at 88.9% pruning | NavGPT-2 FlanT5-XL · R2R val-unseen | SR: 64.2 -> 65.6 | Improved |
| use STEP-Nav instead of VisionZip at 88.9% pruning | Navid-CE · VLN-CE RxR val-unseen (zero-shot) | SR: 21.8 -> 22.4 | No clear effect |
| use STEP-Nav instead of ToMe at 88.9% pruning | Navid-CE · VLN-CE RxR val-unseen (zero-shot) | SR: 17.4 -> 22.4 | Improved |
| use STEP-Nav instead of VisionZip at 85% pruning | NavGPT-2 FlanT5-XL (1.5B) · R2R val-unseen | SR: 64.5 -> 66.3 | Improved |
| use STEP-Nav instead of VisionZip at 85% pruning | NavGPT-2 FlanT5-XXL (5B) · R2R val-unseen | SR: 70.1 -> 69.6 | Hurt |
| use STEP-Nav instead of VisionZip at 77.8% pruning | Navid-CE · Real-world (3 indoor environments) | SR: 78% -> 80% meeting room; 67% -> 69% bedroom; 84% -> 91% corridor | Improved |
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.
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 9 of 10 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.
Closest papers by meaning
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.90)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.90)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 2026 · similarity 0.90)
- Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models (CVPR 2026 · similarity 0.90)
- IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning (ICLR 2026 · similarity 0.90)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.90)
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.
