Papers

DART 方法修复潜空间世界模型丢失运动信息的问题

Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model

Why we picked it

一篇很扎实的世界模型诊断论文:冻结的 flow 学不会物体运动,DART 只重训 flow 加解码监督就修好了,还顺带指出了评估指标的坑。

论文指出,在冻结自监督潜空间中训练的 flow 世界模型虽然训练稳定且成本低,但预训练 flow 不会移动被操控的物体,潜空间损失重训只会产生类似瞬移的运动。

Body

作者将问题归因于训练信号:anchor 稀疏的纯潜空间监督没有说明时间轴上变化发生的位置。提出的 decode-augmented rollout training(DART)在保持表征冻结的情况下,仅用 decode 路径监督重训 flow,恢复了运动的时序结构。

DART 在完整协议上超越纯潜空间基线,较大规模下接近 oracle-informed 插值参考近一半的剩余差距,论文还发现仅用像素误差评估会奖励冻结的预测。

Original · arXiv cs.AI

Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model

Latent world models that integrate a flow in a frozen self supervised latent space train stably and cheaply, yet silently lose the property manipulation depends on most: motion. The pretrained flow never moves the manipulated object; retraining it with latent-only losses only trades stillness for teleport-like motion. We trace the failure to the training signal, not the representation: anchor-sparse, latent-only supervision never says where along the horizon change belongs. Decode-augmented rollout training (DART) repairs this while keeping the representation frozen, retraining only the flow with decode-path supervision. DART outperforms its latent only parent on the full protocol, restores the temporal structure of motion, and re-couples predicted motion to the scene; at larger scale it further improves prediction quality, closing nearly half the remaining gap to an oracle-informed interpolation reference. Finally, we report an unexpected finding about evaluation: pixel error alone rewards frozen predictions.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include world model, flow model, DART, video generation.

Source attribution: curated from arXiv cs.AI (@Xiwen Chen, Rigaudiere Z. Li, Zhiruo Zhou, Xiaojun Zhu, Houde Liu).