Continuous Perception Benchmark (CP-Bench)

Overview

The Continuous Perception Benchmark (CP-Bench) is a diagnostic dataset designed to evaluate whether modern vision-language and multimodal models can integrate continuous visual information over time—an ability that is central to human visual perception but largely absent in contemporary architectures. Inspired by the continuous, stream-based nature of human vision, CP-Bench isolates the core requirement of maintaining object permanence and building a stable scene representation from partial observations.

Where humans effortlessly accumulate evidence from a continuous flow of visual input, most current models treat videos as weakly ordered collections of frames. This frame-centric processing encourages reliance on texture and momentary cues rather than sustained, temporally consistent reasoning. CP-Bench is constructed to expose this gap by creating scenarios where appearance-based shortcuts are impossible and true continuous perception is required.

Dataset Design

CP-Bench is built using CLEVR-style rendering within Blender. Each video depicts a set of visually identical cubes placed on a flat plane—identical in shape, size, color, material, and texture. A virtual camera executes a smooth horizontal pan across the scene over 10 seconds, ensuring that only part of the environment is visible at any given moment. Because no cube is visually distinguishable from another and visibility is partial throughout the sequence, models must integrate information across the entire video to correctly determine the total number of cubes.

The primary benchmark includes five object-count configurations (3–7 cubes), with 20 unique scenes per configuration, resulting in 100 videos. All rendering parameters aside from cube positions are held constant to eliminate low-level confounds. The task for each video is simple yet demanding: “How many cubes are present in the scene?” Solving this requires accumulating evidence over time rather than relying on any single frame.

These videos constitute the split test_moving_cam_10s.

Control Test Set

To ensure that poor model performance on CP-Bench cannot be attributed to low-level visual processing limitations, the dataset includes a matched control test set. These control videos preserve the exact visual appearance of the main benchmark but eliminate the need for temporal integration. Instead of a panning camera, the viewpoint is static, and all cubes are simultaneously visible from the very first frame.

As with the main benchmark, the control set spans 3–7 cubes with 20 instances per count, totaling 100 videos. High performance here combined with low performance on the main set indicates that a model’s failure lies specifically in temporal integration rather than recognition or detection.

These videos constitute the split test_static_cam_10s.

Generalization Splits

To probe whether models can learn continuous perception, we include additional training and testing splits:

• train_moving_cam_10s
  Synthetic training set matching the 10-second main setting.
  Contains 1,000 instances (200 per ground-truth count).

• test_moving_cam_5s
  A 5-second version of the main test configuration, with a shorter camera traversal.

• train_moving_cam_5s
  Synthetic training set matching the 5-second short-duration setting.
  Contains 1,000 instances (200 per ground-truth count).

These splits support experiments in which a model is trained on 10-second videos but tested on 5-second videos, or vice versa. Models that have truly learned spatiotemporal correspondence and continuous evidence accumulation should generalize across these duration mismatches, rather than overfitting to a fixed traversal length.

---

We hope that CP-Bench serves as a foundation for future research toward models capable of genuinely continuous perception, i.e., systems that track persistent structure in the world, accumulate information over time, and reason about scenes the way humans do.

@article{cpbench2025,
  title={Continuous Perception Matters: Diagnosing Temporal Integration Failures in Multimodal Models},
  author={Zeyu Wang and Zhenzhen Weng and Serena Yeung-Levy},
  journal={arXiv preprint arXiv:2408.07867},
  year={2025}
}