Model reports

PanoVideoBench

跟随 360 度视频中的人物与物体，把不同朝向、不同时间的证据连起来。能认出单帧内容，不等于能追踪事件的完整过程。

原论文官方项目官方数据集收录来源

AI Benchmark Radar / Equality Machine · 核对于 2026-11-03

2026 年 10 月 30 日提交的原论文 v1；11 月 3 日核查新增。覆盖全题集准确率及明确标注的输入、训练消融。

一道题要完成什么？

根据整个球面视野和时间线中的证据四选一。1,000 道人工标注题来自 223 段视频，覆盖外观、空间、动作、运动和事件。

PanoVideoBench original paper v1 · PanoVideoBench · author annotation dataset

固定观看预算

默认输入为原始等距柱状投影视频：均匀采样目标 1 fps、最多 128 帧、长边 1,280 像素，使用各模型默认温度，输出上限 32,768 token（含思考）。

PanoVideoBench · evaluation and input protocol

总分与分类成绩口径不同

总分对每道题等权。类别题量不同，分类成绩的简单平均不是总分。主成绩保留全部独立的全题集输入与训练设置，重复出现的基线只记一次。

PanoVideoBench · Table 3 · PanoVideoBench · Table 4 · PanoVideoBench · Table 5 · PanoVideoBench · author annotation dataset

在哪里取得评测资料

作者数据集提供标注和视频标识，原始视频需另外获取。标准答案和诊断标签用于评分，不应作为模型输入。

PanoVideoBench · author annotation dataset

表 3 · 完整对照

11 个基线和作者训练的系统采用默认视频输入。

表 3 · 完整对照
模型	总分（%）	外观（%）	空间（%）	动作（%）	运动（%）	事件（%）
Gemini 3.7 Flash	53.1	65.7	49.5	52.4	40.9	36.5
GPT-5.6-sol	49.6	59.9	46.0	51.9	33.5	41.9
Gemma-4-31B-it	47.0	58.4	48.5	44.6	34.8	27.0
Nemotron Omni 30B-A3B	46.3	59.9	46.0	41.6	31.1	35.1
InternVL3.5-8B	44.9	53.8	36.1	49.4	34.1	39.2
Qwen3.8-27B	44.8	54.7	44.1	46.8	31.1	27.0
Qwen3.5-9B	43.2	55.4	36.6	42.5	35.4	27.0
Eagle2.5-8B	39.6	50.5	35.6	38.6	32.3	21.6
MiniCPM-V-4.6	39.0	48.9	37.1	34.3	34.1	25.7
GLM-4.1V-9B-Thinking	38.0	46.5	38.6	32.2	31.1	32.4
PanoWorld-Hstar	42.9	51.7	45.0	36.5	34.1	37.8
PanoVideoReasoner	55.7	59.6	52.5	60.1	48.8	48.6

仅 Overall 进入主成绩；各类别成绩是子集。

PanoVideoBench · Table 3

表 4a · ERP 分辨率

所有单元格均为原文百分数。Qwen 指 Qwen3.8-27B，Gemma 指 Gemma-4-31B-it；训练变体以 Qwen3.5-9B 为基础。

表 4a · ERP 分辨率
Metric	Qwen 640	Qwen 1280	Qwen 1920	Gemma 640	Gemma 1280	Gemma 1920
Overall	42.3	44.8	46.2	46.0	47.0	47.2
APP	46.8	54.7	56.0	57.2	58.4	58.1
SPA	44.6	44.1	48.5	47.0	48.5	48.0
ACT	45.9	46.8	48.5	42.9	44.6	45.1
MOT	31.1	31.1	32.9	34.1	34.8	32.9
TEM	29.7	27.0	18.9	29.7	27.0	35.1

ERP 为默认投影；8V 为八视图拼图；Ovl 添加朝向标记。Unif./Q-g./EP 分别表示均匀、问题引导和证据特权教师视图。

PanoVideoBench · Table 4

表 4b · 输入形式

所有单元格均为原文百分数。Qwen 指 Qwen3.8-27B，Gemma 指 Gemma-4-31B-it；训练变体以 Qwen3.5-9B 为基础。

表 4b · 输入形式
Metric	Qwen ERP	Qwen 8V	Qwen Ovl	Gemma ERP	Gemma 8V	Gemma Ovl
Overall	44.8	41.0	44.5	47.0	40.9	44.5
APP	54.7	47.7	48.6	58.4	43.7	45.6
SPA	44.1	44.6	48.0	48.5	40.6	50.5
ACT	46.8	42.5	45.1	44.6	42.1	43.3
MOT	31.1	26.8	34.1	34.8	37.8	42.1
TEM	27.0	28.4	37.8	27.0	32.4	32.4

ERP 为默认投影；8V 为八视图拼图；Ovl 添加朝向标记。Unif./Q-g./EP 分别表示均匀、问题引导和证据特权教师视图。

PanoVideoBench · Table 4

表 5a · 监督方式

所有单元格均为原文百分数。Qwen 指 Qwen3.8-27B，Gemma 指 Gemma-4-31B-it；训练变体以 Qwen3.5-9B 为基础。

表 5a · 监督方式
Metric	Base	Answer	CoT
Overall	43.2	45.6	51.8
APP	55.4	55.0	57.2
SPA	36.6	41.1	45.0
ACT	42.5	45.5	57.1
MOT	35.4	36.0	45.1
TEM	27.0	37.8	44.6

ERP 为默认投影；8V 为八视图拼图；Ovl 添加朝向标记。Unif./Q-g./EP 分别表示均匀、问题引导和证据特权教师视图。

PanoVideoBench · Table 5

表 5b · 教师视图

所有单元格均为原文百分数。Qwen 指 Qwen3.8-27B，Gemma 指 Gemma-4-31B-it；训练变体以 Qwen3.5-9B 为基础。

表 5b · 教师视图
Metric	CoT	Unif.	Q-g.	EP
Overall	51.8	50.9	53.9	55.7
APP	57.2	57.8	58.4	59.6
SPA	45.0	46.5	49.0	52.5
ACT	57.1	55.4	58.8	60.1
MOT	45.1	41.5	47.0	48.8
TEM	44.6	39.2	47.3	48.6

ERP 为默认投影；8V 为八视图拼图；Ovl 添加朝向标记。Unif./Q-g./EP 分别表示均匀、问题引导和证据特权教师视图。

PanoVideoBench · Table 5

原始发布

2026 年 10 月 30 日提交的原论文 v1；11 月 3 日核查新增。覆盖全题集准确率及明确标注的输入、训练消融。

核对于 2026-11-03

已收录的评测资料

这条来源记录收录了 24 条数值成绩。来源区分了 16 个模型身份或配置。同一模型可能有多条记录；比较前请核对具体版本、工具和运行预算。查看成绩及逐条证据 →

报告指标: Overall choice accuracy (all 1,000 questions)
原始单位: percent
指标方向: 数值越高越好

从这些资料开始

收录来源 · arxiv.org

引用文档及日期（1）

PanoVideoBench paper v1 · reported benchmark results · 2026-10-30

评测信息与证据

来源: Model reports
发布者: 未收录
评测发布日期: 2026-10-30
采集日期: 未收录
模态: 未收录

本站保留来源报告的结果，不代表独立复测。请先核对任务版本、评测协议、分数单位和报告日期；不同运行条件不能直接合并为统一排名。

报告成绩 24

按来源中的观察顺序展示，顺序不代表排名。日期保留原始精度。

模型	数值	单位	协议	报告日期与口径	来源
Gemma-4-31B-it	44.5	percent	Full 1,000-question choice accuracy; ERP with fixed question-independent orientation overlay; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 4b Ovl input; see Appendix C.3	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.8-27B	44.5	percent	Full 1,000-question choice accuracy; ERP with fixed question-independent orientation overlay; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 4b Ovl input; see Appendix C.3	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Gemma-4-31B-it	40.9	percent	Full 1,000-question choice accuracy; eight-view compass perspective mosaic; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 4b 8V input; see Appendix C.3	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.8-27B	41.0	percent	Full 1,000-question choice accuracy; eight-view compass perspective mosaic; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 4b 8V input; see Appendix C.3	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Eagle2.5-8B	39.6	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
GLM-4.1V-9B-Thinking	38.0	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
GPT-5.6-sol	49.6	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Gemini 3.7 Flash	53.1	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Gemma-4-31B-it	47.0	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
InternVL3.5-8B	44.9	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
MiniCPM-V-4.6	39.0	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Nemotron Omni 30B-A3B	46.3	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
PanoVideoReasoner	55.7	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
PanoWorld-Hstar	42.9	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.5-9B	43.2	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.8-27B	44.8	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 3 main comparison	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.5-9B · CoT SFT	51.8	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 5a CoT SFT; trained by paper authors on PanoVideoCoT	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.5-9B · answer-only SFT	45.6	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 5a answer-only SFT; trained by paper authors on PanoVideoCoT	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.5-9B · question-guided OPSD	53.9	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 5b CoT SFT plus question-guided teacher views; trained by paper authors on PanoVideoCoT	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.5-9B · uniform-view OPSD	50.9	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1280px; model default temperature; max 32768 output tokens incl thinking; Table 5b CoT SFT plus uniform teacher views; trained by paper authors on PanoVideoCoT	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Gemma-4-31B-it	47.2	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1920px; model default temperature; max 32768 output tokens incl thinking; Table 4a resolution variant	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.8-27B	46.2	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 1920px; model default temperature; max 32768 output tokens incl thinking; Table 4a resolution variant	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Gemma-4-31B-it	46.0	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 640px; model default temperature; max 32768 output tokens incl thinking; Table 4a resolution variant	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results
Qwen3.8-27B	42.3	percent	Full 1,000-question choice accuracy; original ERP video; uniform whole-video 1 fps, at most 128 frames; long edge 640px; model default temperature; max 32768 output tokens incl thinking; Table 4a resolution variant	2026-10-30 · document_publication	PanoVideoBench paper v1 · reported benchmark results

一道题要完成什么？

固定观看预算

总分与分类成绩口径不同

在哪里取得评测资料

表 3 · 完整对照

表 4a · ERP 分辨率

表 4b · 输入形式

表 5a · 监督方式

表 5b · 教师视图

原始发布

已收录的评测资料

从这些资料开始

评测信息与证据

报告成绩 24

相关评测探索

继续阅读

五个新评测详情：Agent 安全、云运维与视频理解

六个新评测详情：上下文、Agent 与可执行任务

Navers Lab / Einsia：六个评测分别测什么

AI 评测与排行榜，分别告诉你什么

怎样比较 AI 评测成绩，避免混用评测口径

四个新 Agent 评测：适应变化、安全执行与视频交付

继续探索