Research map / Vision-language and multimodal / Multimodal LLMs

CitySeeker: How Do VLMs Explore Embodied Urban Navigation with Implicit Human Needs?

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Introduces CitySeeker, a benchmark of 6,440 urban navigation trajectories across 8 cities testing VLMs on implicit human needs, evaluates 27 VLMs (best 21.1% TCP), and explores backtracking, spatial cognition enrichment, and memory retrieval strategies.

Problem

VLMs are underexplored for embodied urban navigation driven by implicit human needs (e.g., "I am thirsty") requiring long-horizon spatial reasoning over dynamic street-view environments.

Approach

A demand-driven benchmark built on a Neo4j graph of street-view panoramas and POIs maps implicit needs to POI categories and evaluates a ReAct-style VLM navigation loop, plus three human-inspired strategies (BCR) that add backtracking, cognitive-map cues, and graph memory retrieval.

Key results

• Qwen2.5-VL-32B-Instruct achieves the best overall TCP at 21.1% versus 30.1% for human participants.
• Human participants reach 5.7% TCE and 30.1% TCP, outperforming all 27 VLMs on TCP and TCC.
• Memory-based retrieval R1 raises Qwen2.5-VL-32B-Instruct TCP from 19.9% to 26.9%, the largest gain among the tested BCR strategies.

Surprising findings

• Providing a global 2D map to agents surprisingly degraded task completion.
• Some VLMs performed below the random-choice baseline on overall TCP and TCC.

Limitations

• Street-view imagery cannot be redistributed under API terms, so users must re-fetch images via public APIs.
• BCR strategies were explored on a 650-sample mini subset rather than the full test set.
• Performance gaps between cities and models are not fully explained; GPT-4o underperforms in Beijing.

Experimental setup

Datasets: CitySeeker

Benchmarks: CitySeeker

Metrics: TCE TCP TCC SPL SPD nDTW AS

Baselines: GPT-4o Gemini-2.5-Pro Qwen2.5-VL-32B Human baseline Random Choice baseline Forward Direction baseline

Builds on: ReAct Touchdown nDTW

Closest papers by meaning

• CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos (CVPR 2025 · similarity 0.88)
• Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.87)
• Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.87)
• SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.87)
• Run, Ruminate, and Regulate: A Dual-process Thinking System for Vision-and-Language Navigation (AAAI 2026 · similarity 0.87)
• Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 2025 · similarity 0.87)