OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs

2026-06-02

AI Summary

Existing benchmarks struggle to evaluate the dynamic spatial reasoning capabilities of multimodal large language models in continuous first-person video streams, particularly lacking assessments of temporal evolution and utilization of off-screen evidence. This work proposes OVO-S-Bench—the first human-annotated benchmark for streaming spatial intelligence—comprising 348 videos and 1,680 timestamped questions with annotated evidence intervals, hierarchically structured into four task categories: spatial perception, spatiotemporal tracking, spatial simulation, and allocentric mapping. Models are strictly constrained to reason using only the video prefix preceding each query timestamp. Data quality is ensured through dense manual annotation, blind cross-validation, and multi-round review. Evaluations across 38 state-of-the-art models reveal that even the best-performing model (Gemini-3.1-Pro) achieves only 59.2%, substantially below human experts (86.6%), with allocentric mapping being the weakest component; notably, streaming fine-tuning degrades performance, and unfounded chain-of-thought reasoning exacerbates spatial errors.

Abstract

Multimodal agents in robotics, AR, and autonomous driving must reason about places and layouts from continuous egocentric streams, often using evidence outside the current view. Existing benchmarks either evaluate offline over full videos or target events rather than spatial structure. We introduce OVO-S-Bench, a fully human-annotated benchmark for streaming spatial intelligence, comprising 1,680 questions over 348 source videos. Annotation involves 12 trained annotators, each also serving as a blind cross-reviewer, across roughly 804 person-hours of multi-round quality assurance. Each question carries a query timestamp and an evidence interval, and at evaluation, the model sees only the prefix preceding the query. Questions span four levels of increasing abstraction: instantaneous egocentric perception, spatiotemporal context tracking, spatial simulation and reasoning, and allocentric mapping. Across 38 proprietary and open-source MLLMs, Gemini-3.1-Pro trails human experts by 27 points, 59.2 vs. 86.6, with allocentric mapping as the dominant bottleneck. Notably, streaming and spatially fine-tuned MLLMs underperform their own backbones. We further find that chain-of-thought reasoning amplifies spatial errors when ungrounded in the stream. By exposing these limitations, OVO-S-Bench establishes a demanding testbed for next-generation streaming spatial MLLMs.

Problem

• streaming spatial intelligence
• multimodal LLMs
• spatial reasoning
• egocentric video
• benchmark

Innovation

• streaming spatial intelligence
• multimodal LLMs
• egocentric video benchmark
• allocentric mapping
• hierarchical reasoning