Think in 360°: Benchmarking and Bootstrapping Panoramic Video Reasoning

Lidong Lu1Lidong Zou1Xinyi Mao1Jiaxin Lv1Zhenquan Li1Liangrui Dong1Tong Lu1Guo Chen2

1 Nanjing University2 NVIDIA

Paper Benchmark Dataset Model
PanoVideoBench evaluates how models find and connect evidence across direction and time. PanoVideoCoT turns verified sphere-time evidence and solution trajectories into reasoning supervision.

Abstract

PanoVideoBench

PanoVideoBench contains 1,000 manually annotated four-choice questions from 223 panoramic videos, spanning 23.06 h. The questions cover five categories and 15 question types, with diagnostic annotations for target search and panoramic capabilities.

The sunburst reports the five semantic categories and 15 question types. The heatmaps cross evidence-search settings and semantic categories with five panoramic capabilities; percentages indicate dataset-wide frequency, and cells give question counts.

PanoVideoCoT

PanoVideoCoT contains 13,465 QA–CoT pairs from 2,063 real panoramic videos, spanning 168.8 hours. The corpus includes verified sphere-time evidence and approximately 100K evidence nodes.

Source-video plots summarize duration and resolution. The remaining plots report angular evidence span, target-search setting, CoT length, and the co-occurrence of panoramic capabilities.

Main results

PanoVideoReasoner reaches 55.7% on PanoVideoBench, improving its Qwen3.5-9B backbone by 12.5 points and the previous SOTA by 2.6 points.

Accuracy (%) on PanoVideoBench.
ModelOverallAPPSPAACTMOTTEM

APP: appearance and detail; SPA: spatial and viewpoint reasoning; ACT: action and state; MOT: motion and identity tracking; TEM: spatio-temporal events. Best results in each column are bold.

Examples

    PanoVideoReasoner

    Qwen3.8-27B

    Reasoning

    BibTeX

    TBD