Source-linked AI summary
What Should World Models Forget? Stratified Retention for Continual Adaptation
Nishit Anand, Ramani Duraiswami, Dinesh Manocha
TL;DR
World models face changing prediction targets, so standard forgetting metrics cannot distinguish harmful knowledge loss from necessary revision of outdated facts. The paper proposes stratifying retention by invariance timescale and evaluating invariant regression alongside revision latency, separately.
Problem
Continual-learning metrics treat degradation on earlier tasks as failure, but cannot distinguish catastrophic forgetting from warranted revision when a world model's environment changes.
Source: §1 · p. 1, §4 · pp. 3–4Method
The paper proposes differential retention: rerun held-out invariant probes after each adaptation round and report their retention separately from revision latency.
Source: §5 · p. 4Results
Standard forgetting metrics give their best score to a model that never updates and cannot distinguish warranted revision from catastrophic forgetting.
Source: §4 · pp. 3–4Takeaways & Limitations
World-model retention should be stratified by invariance timescale, with invariant preservation and revision latency reported jointly but not aggregated.
Source: §3 · p. 3, §5 · p. 4The strata are a starting proposal with domain-dependent boundaries, and probes currently exist only for physical invariants; invariant localization in learned models is unknown.
Source: §7 · p. 5
Abstract
from arXiv · showhide
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.
1 Introduction
World models predict changing environments, so knowledge that was once accurate can become false and need revision. The paper distinguishes this obsolescence from catastrophic forgetting and proposes stratifying retention by how long knowledge remains true.
1 Introduction
- World models approximate environment dynamics, but deployed systems must revise their account as environments change.A prediction target that changes makes deterioration on previously learned material an unreliable signal of failure.Source: §1 · p. 1
- Knowledge that becomes false is obsolete and must be revised, unlike still-true knowledge lost through catastrophic forgetting.Deliberate unlearning instead removes knowledge that remains true but is unwanted.Source: §1 · p. 2
- World models combine revisable facts with physical, geometric, and causal structure that should survive every update.The paper argues this combination distinguishes world models from concept drift and language-model factuality settings.Source: §1 · p. 2
- The paper proposes stratifying knowledge by invariance timescale and evaluating invariant preservation alongside revision latency.It frames obsolescence as distinct from catastrophic forgetting and identifies differential retention as its evaluation protocol.Source: §1 · p. 2, Figure 1 · p. 2
2 World models break the stationary-target assumption
A world model can be wrong about the past because it never learned something or because the environment changed after it learned it. Existing continual world-model benchmarks generally hold ground truth fixed, leaving the second case untested.
2 World models break the stationary-target assumption
- A past prediction error can reflect either missing acquired knowledge or a changed environment, but only the former is a learning failure.Observed behavior alone does not distinguish these cases.Source: §2 · p. 2
- Continual-Dreamer, WMAR, and DRAGO benchmarks use fixed ground truth within tasks, so subsequent degradation represents forgetting rather than obsolescence.DRAGO's task sequences share dynamics and differ only in reward.Source: §2 · p. 2
- With an unbounded replay buffer, DreamerV3 retains essentially all measurable earlier-task knowledge while the policy degrades, suggesting failure need not reflect lost stored information.Wan et al. also report poor adaptation by PlaNet and DreamerV2 to local environmental changes.Source: §2 · pp. 2–3
3 Stratification by invariance timescale
The paper organizes world-model knowledge by invariance timescale, a property of the environment that determines how long knowledge remains true. This differs from protecting knowledge according to past task importance or memory update frequency.
3 Stratification by invariance timescale
- Invariance timescale τ measures how long a piece of knowledge remains true and belongs to the environment, not the model or its update schedule.The proposed organization is by τ rather than task or modality.Source: §3 · p. 3
- Existing consolidation methods protect knowledge important to past performance, a criterion unrelated to how long that knowledge stays true.They may therefore strongly protect facts the environment has invalidated.Source: §3 · p. 3
- Standard pipelines can under-protect physical or geometric invariants while replay over-protects outdated scene layouts.Evicting outdated exemplars addresses the latter problem in classification, but not the world-model need to protect invariants.Source: §3 · p. 3
- Unlike multi-timescale memory's update-frequency strata, this proposal tracks how quickly ground truth changes.The top stratum includes object permanence, geometric consistency, causal action–effect structure, and the agent's model of its own body.Source: §3 · p. 3
4 What current evaluation cannot measure
Standard continual-learning metrics treat any decline on earlier tasks as forgetting, so they cannot tell warranted revision from damage. In changing environments, this makes a frozen model appear optimal despite the need to adapt.
4 What current evaluation cannot measure
- Backward transfer and forgetting measures assign the same score to degradation from broken knowledge and degradation from correctly revised knowledge.They do not ask whether the decline was warranted.Source: §4 · pp. 3–4
- The best possible standard forgetting score belongs to a model that never updates, making the frozen model a failure mode in changing environments.In stationary benchmarks, such a model also performs poorly on subsequent tasks.Source: §4 · p. 4
- Updated-target metrics can reward scene revision without detecting damage to invariants such as object permanence.Matching a refreshed target alone says nothing about whether invariants survived the update.Source: §4 · p. 4
5 Differential Retention
Differential retention evaluates continual world models on invariant preservation and timely revision, reported separately. The proposed protocol exposes trade-offs that a single score cannot capture, but comparable probes currently cover only physical invariants.
5 Differential Retention
- Invariant regression testing reruns a held-out τ = ∞ probe set after every adaptation round, revealing which update damages invariants.Existing concept-isolated physical probes can be reused across the adaptation stream without new data.Source: §5 · p. 4
- Revision latency counts observations until predictions reflect a changed fact and separately counts facts the model never revises.Source: §5 · p. 4
- Frozen models score well on invariant retention but fail revision latency, while overwrite-adapting models show the opposite pattern; neither axis alone suffices.Source: §5 · p. 4
- Comparable probes currently make invariant regression testing implementable only for physical invariants, not geometric consistency or causal structure.Source: §5 · p. 4
- Current evaluation cannot tell whether online recovery reflects correct appearance-prior revision or dynamics-model degradation; Maes et al. [17] report Push-T success falling from 50.8% to 6.0% for both LeWM and PLDM after background recoloring.Source: §5 · p. 4
6 Related Work
Related work spans continual world-model replay, non-stationary ground truth, physical evaluation, and memory-plasticity constraints. The paper distinguishes its timescale-based framing from approaches that assume fixed targets or stratify memory by update frequency.
Related work
- Continual-Dreamer adopts reservoir sampling over past experience to robustly mitigate forgetting, Yang et al. [23] make the buffer memory-efficient by matching past-experience distributions, and Fu et al. rehearse synthetic trajectories; all three assume a fixed target.Source: §6 · p. 4
- Concept-drift work supplies the framing, while Jiang et al. offer the closest metric for language-model factuality.Source: §6 · p. 4
- WorldBench and PhyGround [16] isolate physical concepts, Maes et al. [17] document distribution-shift brittleness, and Xing et al. discuss world-model purpose.Source: §6 · p. 4
- Behrouz et al. stratify memory by update frequency, while Dohare et al. show that networks lose learning capacity over long streams.Source: §6 · p. 4
7 Limitations and open questions
The proposed framing and evaluation protocol leave open questions about defining timescale strata, extending probes beyond physics, and whether invariants are localized within learned models.
Limitations and open questions
- The proposed strata are a starting point with likely domain-dependent boundaries, and it remains unclear whether τ can be estimated from observation or requires domain knowledge.Source: §7 · pp. 4–5
- Comparable probes currently exist only for physical invariants, leaving geometric consistency and causal structure without the probes needed for full protocol coverage.Source: §7 · p. 5
- Whether invariants are localized in a learned world model is unknown; if they are entangled with other knowledge, stratified retention may need architectural support.Source: §7 · p. 5