CMBench: Context Memory Benchmark

CMBench evaluates whether a video generator can recall and reproduce visual evidence from an approximately one-minute context. Given a context video and a continuation instruction, the model must bring back a specific person, object, or view observed earlier. The benchmark contains 58 context videos and 116 continuation tasks, combining synthetic and real-world footage.

The central question is whether information from the video history remains available when it is needed again. A continuation can look natural while failing to reproduce the requested target. CMBench therefore measures recall against visual references from the context itself. It is designed to evaluate context memory in autoregressive video generation, particularly under controlled compression of historical key-value (KV) caches.

Benchmark design

CMBench combines three design choices:

• Recall beyond the recent context. Target events occur earlier in a minute-scale video, and continuation instructions request information that is absent from the recent view. Multiple events within a context serve as distractors for one another.
• Reference-grounded evaluation. Each task has an annotated reference frame and target bounding box. Evaluation compares the generated target with the entity actually visible in the context, so it tests preservation of specific visual evidence rather than prompt adherence alone.
• Controlled and natural contexts. Synthetic videos support control over event timing, target visibility, and camera transitions. Real-world videos provide complementary natural footage evaluated with the same task definitions and scoring protocol.

Synthetic contexts consist of six 10-second clips constructed with first–last frame conditioning. A boundary frame is shared between adjacent clips to maintain continuity. Each clip contains at most one complete target event, keeping that event within a single clip. Pairing a target event with a continuation instruction defines one task; one context video can support multiple tasks.

Tasks and dataset composition

Reappear asks a person or object observed earlier to appear again. Success requires reproducing the same entity and appearance, including its distinguishing visual details.

Revisit starts from a context that shows a camera transition from scene A to scene B and back to A. The continuation must return to B. A salient object in B serves as the evaluation anchor. In synthetic contexts, the complete A → B → A event occurs within one 10-second clip, providing a consistent reference for the requested return.

Video type	Videos	Reappear tasks	Revisit tasks	Total tasks
Synthetic	50	67	35	102
Real	8	10	4	14
Total	58	77	39	116

Both tasks test recall of previously observed evidence. The release contains 58 MP4 context videos and 116 task records, with six ordered clip descriptions, a continuation instruction, target labels and aliases, and a reference annotation for each task.

Evaluation protocol

For each task, provide the context video and continue_prompt to the generator, then evaluate the resulting continuation against references:

1. Extract the annotated reference frame and target bounding box from the context video.
2. Use OWL-ViT to localize the target in generated frames, and SAM 2 to segment the reference and generated targets.
3. Compare DINOv2 embeddings of the resulting target crops using cosine similarity.

For reference crop $r$ and generated target crop $g_t$ at continuation frame $t$, the task score is

$$S_{\mathrm{D}\mathrm{I}\mathrm{N}\mathrm{O}}=\underset{t}{\max }\cos \text{\hspace{0.17em}⁣}(f_{\mathrm{D}\mathrm{I}\mathrm{N}\mathrm{O}}(r),f_{\mathrm{D}\mathrm{I}\mathrm{N}\mathrm{O}}(g_t))\mathrm{.}S\_\{\backslash mathrm\{DINO\}\}\; =\; \backslash max\_t\; \backslash operatorname\{cos\}\backslash !\backslash left(f\_\{\backslash mathrm\{DINO\}\}(r),\; f\_\{\backslash mathrm\{DINO\}\}(g\_t)\backslash right).$$

Taking the maximum allows the requested event to occur at any point in the continuation. A task receives zero if the target is never detected. For Revisit, the salient object in scene B is the target, so both task types use the same scoring protocol. The benchmark score is the mean over all tasks; higher scores indicate better recall of the annotated targets.

Evaluating memory compression

For historical-KV compression experiments, report recall alongside the effective pruning ratio (PR) and continuation-generation throughput. For each task,

$$\mathrm{P}\mathrm{R}=1-\frac{\sum _{i,\mathrm{\ell },h}{k}_{i,\mathrm{\ell },h}}{\sum _{i,\mathrm{\ell },h}{k}_{i,\mathrm{\ell },h}^{\mathrm{f}\mathrm{u}\mathrm{l}\mathrm{l}}},\backslash mathrm\{PR\}\; =\; 1\; -\; \backslash frac\{\backslash sum\_\{i,\backslash ell,h\}\; k\_\{i,\backslash ell,h\}\}\{\backslash sum\_\{i,\backslash ell,h\}\; k^\{\backslash mathrm\{full\}\}\_\{i,\backslash ell,h\}\},$$

where $k_{i,\ell,h}$ counts historical KV token positions visible to attention head $h$ in layer $\ell$ when denoising generated chunk $i$, and $k^{\mathrm{full}}_{i,\ell,h}$ is the corresponding count when retaining the full history. Exclude the current noisy chunk from both sums and report the arithmetic mean of per-task PR values. This measures cumulative historical-token reduction; peak memory usage and computation require separate measurements. FPS and speedup measure continuation generation, excluding context-prefix processing. General video-quality metrics can complement the reference-grounded recall score.

Files and identifiers

CMBench/
├── README.md
├── metadata.jsonl
└── videos/
    ├── synthetic/
    │   ├── syn_001.mp4
    │   └── ...
    └── real/
        ├── real_001.mp4
        └── ...

metadata.jsonl contains one JSON object per task. Multiple tasks can share one context video. Video paths are relative to the dataset root.

Video identifiers are syn_001–syn_050 and real_001–real_008. Task identifiers follow {video_id}_{task_type}_{task_number}, for example syn_001_reappear_01. Task numbers start at 01 within each video and task type.

Metadata

Field	Description
type	synthetic or real.
task_id	Unique task identifier.
scene	One- or two-word English scene description. Scene descriptions need not be unique.
video	Relative path to the context video.
context_clip_prompts	Six descriptions in chronological order. Each entry contains clip_id (01–06) and clip_prompt. These describe consecutive portions of the context video.
continue_prompt	Instruction for continuing the context video from its final frame.
task_type	reappear or revisit.
target_label	English description of the evaluation target.
target_aliases	Alternative English labels for the target.
references	Reference timestamps and target bounding boxes in the context video.

References

Each references entry has the same schema:

Field	Description
timestamp_seconds	Reference time in seconds from the start of the context video.
frame_index	Zero-based reference frame index when explicitly annotated; otherwise null.
bbox_xyxy_normalized	[x_min, y_min, x_max, y_max], normalized by the full frame width and height. Coordinates range from 0 to 1, with the origin at the top-left corner.

For a frame of width W and height H, multiply the box coordinates by [W, H, W, H] to obtain pixel coordinates. Use floor for the top-left corner and ceil for the bottom-right corner when an integer crop is needed.

Use frame_index to select the reference frame when it is present; otherwise seek by timestamp_seconds. Synthetic references include explicit frame indices. Real references are annotated by timestamp. Each of the 116 tasks has one reference.

Reading the benchmark

import json
from pathlib import Path

root = Path("CMBench")
with (root / "metadata.jsonl").open(encoding="utf-8") as f:
    tasks = [json.loads(line) for line in f if line.strip()]

for task in tasks:
    video_path = root / task["video"]
    instruction = task["continue_prompt"]
    references = task["references"]