What Should World Models Forget? Stratified Retention for Continual Adaptation
世界模型該遺忘什麼？以「分層保留」實現持續適應環境的能力
Traditional continual learning penalizes all forgetting, assuming ground truth is stationary. However, world models face non-stationary environments where outdated facts must be discarded. This paper introduces 'stratified retention,' categorizing knowledge by invariance timescale. While invariants like physics must never be revised, instance-specific facts must update dynamically. To evaluate this properly, the authors propose 'differential retention,' a metric pairing invariant regression testing with revision latency.
Key points
Redefining Forgetting
In non-stationary environments, actively discarding outdated information is a required capability, not a system defect.
Stratified Retention Mechanism
Stratifies knowledge by invariance timescale, separating permanent laws (physics) from fast-changing, instance-level facts.
Limitations of Traditional Metrics
Standard forgetting metrics fail to distinguish correct knowledge revision from catastrophic forgetting, often favoring completely frozen models.
Differential Retention Metric
Introduces a new metric that jointly reports invariant regression testing and revision latency without aggregating them.
How it works
| 不變要素 (Invariants) | 實例級事實 (Instance-level Facts) | |
|---|---|---|
| Definition | 物理規律、物體恆存等永久不變的法則 | 特定環境中會隨時間與空間改變的動態資訊 |
| Update Strategy | 絕對不可修改 (Never revise) | 隨環境變動立即修正 (Revise dynamically) |
| Effect of Forgetting | 導致災難性遺忘 (Catastrophic forgetting) | 實現正確的知識適應 (Required adaptation) |
Why it matters
Previously, embodied AI and world models relied on frozen weights due to fear of catastrophic forgetting, limiting adaptability. By debunking the 'forgetting is a failure' myth, this research provides the theoretical and evaluation foundation for building next-generation AI agents and robotic systems that can continuously adapt to dynamic physical realities.
Who it affects
- AI Researcher
- AI Developer
- Student & Learner
How to use it
- 1Autonomous robotic navigation and dynamic mapping
- 2Continuous adaptation of physics-based simulators to environmental shifts
- 3Real-time road condition and rule updates for autonomous driving systems
Limitations & caveats
- Does not yet present a specific large-scale network architecture for perfect stratified retention, focusing primarily on the theoretical framework and evaluation.
- Automatically and precisely identifying which features are 'permanent invariants' within complex multimodal data remains challenging.
Related
Less Decoder is More Encoder: Extracting Robust 3D Geometric Representations via Novel View Synthesis
減少解碼器反而增強編碼器：從新視角合成中提煉強大三維幾何表徵
This paper reveals how expressive decoders dilute geometric learning in Novel View Synthesis, and proposes SNAP—a self-supervised framework that restricts decoders to force encoders to learn robust 3D representations.
4DCodeBench: Benchmarking AI Agents on 4D Inverse Graphics and Dynamic Scene Code Generation
4DCodeBench：評估 AI Agent 動態場景 4D 反向圖形學與程式碼生成能力的全新基準
4DCodeBench is a new benchmark designed to evaluate AI agents' ability to reconstruct 4D dynamic scenes from videos by generating executable graphics and physics code.
RNADyn: A Unified Benchmark and Model for Predicting RNA Molecular Dynamics
RNADyn：預測與理解 RNA 動態變化的統一基準與 AI 模型
This study introduces RNADynBench, a standardized RNA molecular dynamics benchmark, and RNADynNet, a unified model predicting all-atom trajectories from a single conformer.