Latent Dynamics · 2026-10-05
What Should World Models Forget? Stratified Retention for Continual Adaptation
Nishit Anand, Ramani Duraiswami, Dinesh Manocha
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors argue continual world models must treat “obsolescence” (knowledge becoming false as the environment changes) as required behavior, not forgetting failure. They propose stratifying knowledge by invariance timescale τ, separating invariants (e.g., physics, object permanence) that must never be revised from instance-level facts that should be revised promptly. They introduce differential retention with invariant regression testing plus revision latency, reported separately.
Why it matters
Standard forgetting metrics and many benchmarks evaluate only frozen checkpoints or reward high retention and slow revision, so they cannot distinguish correct revision of outdated knowledge from catastrophic forgetting. The authors propose an evaluation protocol that separates invariant retention from revision latency by using held-out invariant probes and measuring how quickly changed facts are reflected, addressing a structural evaluation gap for continual adaptation in changing environments.
Method
- Stratify world-model knowledge by invariance timescale τ (environment property): invariants (τ=∞) must never be revised; instance configuration and transient state should be revised promptly with no retention expected for transient state.
- Use differential retention evaluation with (a) invariant regression testing: fix a probe set from the τ=∞ stratum, hold it out, and re-run after every adaptation round.
- Use (b) revision latency: measure how many observations are needed before predictions reflect a changed fact; report invariant retention and revision latency jointly but separately (no aggregation).
Limitation
Probes of sufficient rigor exist today only for the physical sub-case; realizing the full protocol requires comparable probes for geometric consistency and causal structure. Also, how τ can be estimated from observation or supplied as domain knowledge remains unclear.
Abstract (from arXiv)
Continual learning treats degradation on previously seen data as evidence of failure, a convention inherited from settings with a stationary prediction target, where a correct label remains correct indefinitely. World models do not satisfy this condition. Their prediction target is the environment, which changes, so knowledge that was accurate when acquired may later become false, and discarding it is required behavior rather than a defect. Non-stationary ground truth is well studied in the concept drift literature and in the temporal factuality of language models, but has not been formulated for world models, which are distinctive in that they also encode knowledge that must never be revised. We argue that continual world models require retention stratified by invariance timescale, separating invariants such as physics and object permanence, which must never be revised, from instance-level facts that should be revised as soon as the environment changes. Standard forgetting metrics cannot distinguish a world model that has correctly revised outdated knowledge from one that has suffered catastrophic forgetting, and consequently rank a frozen model highest, while existing physical-reasoning benchmarks evaluate only frozen checkpoints. We propose differential retention, which reports invariant regression testing across the adaptation stream jointly with revision latency, without aggregation.
Related papers
- MemoWM: How World Models Change What Agents Need to Remember
- Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models
- Shaping Persistent Representations from Independent Interactions
- Do World Models Learn Global Understanding?
- When Low Prediction Error Misleads Planning: Diagnosing Representation, Dynamics, and Decision Failures in Latent World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-02 · Indexed by Pub-AI: 2026-10-05 UTC
What Should World Models Forget? Stratified Retention for Continual Adaptation
Preprint posted: 2026-10-02 · Original record ↗
- arxiv
- 2610.03713
Explicit source versions (1)
- v1 · 2026-10-02
- Preprint available
Source date: 2026-10-02 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Continual learning treats degradation on previously seen data as evidence of failure, a convention inherited from settings with a stationary prediction target, where a correct label remains correct indefinitely. World models do not satisfy this condition. Their prediction target is the environment, which changes, so knowledge that was accurate when acquired may later become false, and discarding it is required behavior rather than a defect. Non-stationary ground truth is well studied in the concept drift literature and in the temporal factuality of language models, but has not been formulated for world models, which are distinctive in that they also encode knowledge that must never be revised. We argue that continual world models require retention stratified by invariance timescale, separating invariants such as physics and object permanence, which must never be revised, from instance-level facts that should be revised as soon as the environment changes. Standard forgetting metrics cannot distinguish a world model that has correctly revised outdated knowledge from one that has suffered catastrophic forgetting, and consequently rank a frozen model highest, while existing physical-reasoning benchmarks evaluate only frozen checkpoints. We propose differential retention, which reports invariant regression testing across the adaptation stream jointly with revision latency, without aggregation.