acceptodds
Withdrawn from ICLR 2027

Rules Are Rare: When World Models Memorise Rather Than Acquire Rules

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

World models are trained to predict the next state, and by that objective they also learn the rules of their environment. We show that the world models we train learn them by memorising. On Craftax, a transformer world model (TWM) is as accurate on a rule before the episode has shown it as after (88% against 89%), and when the rules are resampled each episode it falls to the accuracy of predicting that nothing changes. Training on changing rules does not repair this: on XLand-MiniGrid, a Long-Context Transformer model that sees the whole episode memorises one rule table per training episode instead; updating the weights does, by fine-tuning, replay, hidden-parameter inference or online test-time adaptation (AdaJEPA), at three separable costs: interference, forgetting and erosion. The hidden parameter in these environments is not one vector: it splits by interaction type, and the interaction is observable while its effect is not. So we introduce PRISMO. It keeps one entry per interaction type in a store outside the frozen host, writes it from a single observation with no gradient step, and clears it when the rules may change. On two frozen Craftax hosts it raises rule-state accuracy from 29% to 87% (TWM) and from 24% to 79% (a long-context Transformer) while every other predicted token is unchanged, and on XLand two rules observed once each compose into a chain never seen executed.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
World Model, Memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_S5aUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that world models memorize environment rules in their weights rather than learn them from observations during an episode. When rules change between episodes, models fail to adapt their predictions, while fine-tuning and test-time adaptation introduce forgetting and interference. The authors propose PRISMO, an external memory that works with a frozen model. It stores each rule from a single observation, overrides predictions only for the variables governed by that rule, and resets between episodes. On two Craftax models, PRISMO improves rule-state accuracy from 29.4% to 87.2% and from 23.9% to 79.3%, without changing other predictions. On XLand-MiniGrid, it composes two individually observed rules into a two-step chain that the model never experienced, achieving 94.8% accuracy. With one rule withheld, accuracy drops to 0.6%, suggesting the model genuinely uses the stored rules.

Critical strengths
  1. The authors compare predictions before and after a rule is first observed, using predict-unchanged baselines and encoder ceilings to separate memorisation from actual rule learning. This makes the core claim convincing. They also acknowledge unfavourable results, including cases where apparent in-episode gains are simply due to a shifting baseline.

  2. The XLand experiment shows that rules learned from individual observations can be combined into a two-step chain the model has never executed. The negative control strengthens this result: accuracy drops to 0.6% when the second rule is withheld, compared with 93.1% when the full rule table is provided, suggesting the failure comes from the missing rule rather than the chain's complexity.

Critical weaknesses
  1. PRISMO assumes the interaction types and the variables each rule affects are known, so it is not clear how well it works when this structure must be discovered. On TWM, a frozen-host latent baseline achieves similar accuracy (86.8% vs. 87.2%), suggesting PRISMO’s main advantage is avoiding roughly 200 gradient steps per observation, rather than improving accuracy.

  2. The study uses two small Craftax models and XLand models trained on 2,700 episodes, with rule-event frequency deliberately increased. It remains unclear whether the findings hold for larger models, broader training data, or more natural environments. Other results also temper the claims: the fixed-mapping host achieves a longer rollout horizon than PRISMO (43 vs. 32 steps), while the decision-task advantage (61.6% vs. 45.5%) is based on just 36 episodes, with overlapping confidence intervals.

Questions for authors
  1. PRISMO assumes the interaction types and governed variables are known in advance. How much of its benefit remains if either must be inferred from observations? If this is out of scope, could the paper clarify that its contribution is focused on rule acquisition given a known rule structure?

  2. Appendix A.3 notes that default data collection produces too few rule events for evaluation. Could you vary the event density and report how the memorisation diagnosis and PRISMO’s gains change? This would clarify whether the findings hold across different densities or depend on the artificially increased frequency.

  3. The current findings are limited to the tested models, training setups, and corpus sizes. Could this limitation be stated more clearly in the abstract and introduction, rather than presenting memorisation as a general property of world models?

  4. Could PRISMO be tested with noisy or partially observed rule effects, or rules that change within an episode? These settings would help establish how well the approach works beyond the current assumption of observable rule effects and episode-level resets.

Reviewer_ZjFyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies changing rules in Craftax and XLand-MiniGrid. PRISMO freezes a world model and stores observed effects by interaction type. It corrects designated outputs, handles clipped observations, and resets memory when rules change. Experiments show improved rule prediction and two-rule composition.

Critical strengths
  1. The evaluation measures errors on rule-related events separately from unchanged parts of the state. This is useful because high overall accuracy can hide failures on rare events. Testing both fixed and changing rules also helps identify what the models have learned.
  2. PRISMO improves predictions across two base models and can combine observed rules in XLand. Its treatment of bounded observations is sensible: when a value reaches its maximum, the observed change need not equal the full rule effect. The comparisons with simpler memory and latent-variable baselines are also useful.
Critical weaknesses
  1. The methodological advance is narrow. The paper adopts the HiP-MDP setting and acknowledges that external memory is established. WALL-E 2.0 [1] learns executable rules and uses their feedback in world modeling and planning. NeSyS [2] combines neural predictions with executable rules through candidate reranking. PRISMO differs by directly storing interaction-specific effects and replacing designated outputs during an episode. These differences deserve a clear comparison with prior work; the broad benefit of external rules is already established.
    PRISMO assumes that interaction types can be identified and the affected variables are specified in advance. Under these assumptions, one clean observation can identify an effect. Table 3 does show an improvement over plain memory with the same keys and output interface: 87.2% versus 72.7% on the common TWM events. Section 5.3 attributes this 14.5-point gain to handling clipped observations as bounds. This is useful evidence, but the added capability remains closely tied to the assumed effect structure. Unchanged predictions outside the designated variables also follow from the design.

  2. The negative results do not fully explain why the tested models fail to learn rules. Appendix A.16 reports 2,700 XLand training episodes with about 1.7 TileNear events per episode, and acknowledges that more data or training could change the result. Prior work [3] already distinguishes recognizing environments from learning new dynamics, and studies the role of context and diversity. More evidence across training conditions would help establish what the present diagnosis adds. Applying the AdaJEPA protocol to these hosts also tests an adaptation rule, rather than the complete original system. The authors qualify their conclusions, but these limits still reduce the explanatory value of the negative findings.

References:
[1] WALL-E 2.0: World Alignment by NeuroSymbolic Learning Improves World Model-based LLM Agents. 2025. https://arxiv.org/abs/2504.15785
[2] Neuro-Symbolic Synergy for World Modeling. 2026. https://arxiv.org/abs/2602.10480
[3] Context and Diversity Matter: The Emergence of In-Context Learning in World Models. 2025. https://arxiv.org/abs/2509.22353

Questions for authors
  1. Have you tested whether the benefit over plain memory extends to noisy effects or effects that depend on context beyond the stored interaction type? The current ablation establishes a benefit for handling clipped observations.
  2. Do the baseline failures persist with more training, more varied rule sets, or more repeated interactions within each episode? How do these results relate to the context and diversity effects studied in [3]?
  3. What supervision is needed to train g and h and specify the affected variables? How sensitive is PRISMO to errors in identifying the interaction type or the variables to correct?
Reviewer_CowPOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether world models acquire episode-specific rules from observations or memorize training environments. Experiments on Craftax and XLand-MiniGrid show that the evaluated Transformers struggle to predict resampled rules, even when relevant evidence is available in context. The paper introduces PRISMO, an interaction-indexed external memory that records observed effects and overrides the corresponding predictions of a frozen world model. It reports improved rule-state prediction, reduced adaptation costs, and successful two-step composition of observed rules.

Critical strengths

The paper targets an important problem: whether world models can acquire new environment rules from interaction rather than rely on memorized training patterns. This distinction matters for deploying world models in environments whose dynamics vary across episodes.

Critical weaknesses
  • The diagnosis is not thorough enough to support the general claim that world models memorise rather than acquire rules. The evidence comes from a few small, shallow transformers (2–3 layers, although at most ~2M parameters), all trained on small corpora (2700 episodes, 20K steps) with e.g., early stopping or short schedules (instead of grokking). All these settings favor memorization over generalization: the XLand Transformer already scores 90.8% on its own training episodes before any in-context evidence exists (A.10), and prior work shows that in-context learning emerges only above data-burstiness and task-diversity thresholds
  • PRISMO relies on substantial prior structure: identifiable interaction types, known rule-governed variables, and an interface for extracting and applying effects. The evaluated rules are scalar increments with clipping or discrete object mappings. These assumptions make direct estimation and isolated updates particularly effective, but leave unclear how the method extends to unknown preconditions, stochastic effects, or more complex transitions. The results therefore establish a useful solution for restricted rule families rather than a broadly applicable approach to rule acquisition.
Questions for authors
  • Are there results with deeper transformers (even w/ nanogpt) trained on larger corpora?
  • How to apply PRISMO in practical, general settings?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.