EVENT DOSSIER
The SimpleMemVLA team has developed a visual-language-action model that does not require a dedicated memory module. This model transmits sampled history to the backbone network in timestamped video format, using only the hidden states of generated sub-tasks as the sole pathway to the standard flow-matching action heads. Since continuous decision-making shares most of the history, pre-filling shared prefixes during action execution can maintain latency at the single-frame VLA level. With fixed backbone and training settings, SimpleMemVLA outperforms retrieval, compression, and cyclic state mechanisms significantly, and causal intervention confirms that its strategy indeed reads its history. This approach has set new records in four memory benchmark tests.
SimpleMemVLA 团队提出一种无需专用记忆模块的视觉 - 语言 - 动作模型，在四个记忆基准测试中刷新了最先进水平。该模型将采样历史以时间戳视频格式完整传递给骨干网络，仅利用生成子任务的隐藏状态作为通往标准流匹配动作头的唯一通道。由于连续决策共享大部分历史，在执行动作时预填充共享前缀可将延迟保持在单帧 VLA 水平。在固定骨干和训练设置的情况下，SimpleMemVLA 大幅优于检索、压缩和循环状态机制，因果干预证实策略确实读取了其历史。