MemoryTax-Res-world-model-mem

---

🔹 Pipeline (Section 5.2 & Appendix C.6)

Each baseline is run with its released checkpoint under one shared harness. Each sample is generated autoregressively in chunks of 81 frames from a first frame (+ text prompt) and a camera trajectory, then scored with PSNR / SSIM / LPIPS (AlexNet).

Context-as-Memory (CASM), test subset

Task	Input	Generation	Metric
A — revisit	first frame + scene prompt	camera turns away and back (yaw 0° → 90° → 0°), 4 chunks = 324 frames	first frame vs. last revisit frame (closure)
OV — open-domain revisit	image + object prompt	yaw 0° → 45° → 0°, 2 chunks = 162 frames	first frame vs. last revisit frame (closure)
D — trajectory replay	GT first frame + GT camera poses	3 chunks = 243 frames	every frame vs. the GT frame at the same pose, averaged

MemoBench

First frame + prompt + GT camera trajectory (per-clip FOV), 3 chunks = 243 frames. Generated frame i is paired with GT frame round(i·(F−1)/(N−1)) (the official "ALL" pairing); PSNR / SSIM / LPIPS are averaged over all frames.

---

🔹 File structure

res_casm_test_{method}/
├── summary.csv                    # psnr / ssim / lpips over all A + OV + D samples
├── A_{sample_id}/
│   ├── video.mp4                  # generated video
│   └── metrics.json               # sample metadata (scene, start frame, prompt, mode) + psnr / ssim / lpips
├── OV_{sample_id}/
└── D_{video}_start{frame}/
    ├── video.mp4
    └── metrics.json               # sample metadata + mean psnr / ssim / lpips over 243 frames

res_memobench_{method}/
├── per_sample.csv                 # one row per clip: hfov group, start frame, psnr / ssim / lpips
└── {clip_id}/
    ├── video.mp4
    └── metrics.json               # clip metadata (hfov, frames) + mean psnr / ssim / lpips

{method}  framepack, hydra, i3dm, magicworld, reworld, worldkv

summary.csv (CASM) and the per-clip means of per_sample.csv (MemoBench) give the numbers reported in the paper: PSNR / SSIM / LPIPS averaged over all samples (CASM: A, OV and D samples pooled); the paper reports 1 − LPIPS.

---

Citation

@misc{he2026memorytaxdissectingnet,
      title={Memory Tax: Dissecting the Net Value of Agent Memory},
      author={Xuan He and Sining Zhoubian and Ruozhen Yang and Yuhao Wen and Zhineng Wang and Yuyang Hu and Ke Yang and Yuyu Zhang and Ping Nie and Jiawei Han and ChengXiang Zhai},
      year={2026},
      eprint={2610.34073},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.34073},
}