- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 3(+1)
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper frames “understanding” as the ability to learn global constraints from local observations and to propagate their consequences to unseen facts. To study this, the authors construct monoid worlds: finite state spaces whose actions satisfy specified constraints. Held-out transitions are chosen so that they are logically entailed by the seen transitions together with the constraint. The central comparison is between next-state training, where intermediate states are given as input, and compositional training, where they are hidden, on identical training paths.
Across Transformer, GRU, LSTM, and SSM models, next-state training performs at chance on entailed held-out transitions, while compositional training reaches about 96% on inverse, commutativity, and composition constraints. Both objectives succeed on copy and both fail on periodicity. The authors then extend the comparison to two further settings. In an AI2-THOR world model, compositional training outperforms next-observation training. In Wiki data finetuned LLMs, compositional phrasing improves inverse and composition propagation, while converse relations are not propagated. Finally, they define a proof depth for held-out transitions and show that accuracy decays sharply with depth, that deeper transitions are learned later in a manner consistent with a constraint cascade, and that longer composition lengths enable deeper propagation.
Critical strengths
-
The paper asks a simple but interesting question, and its framing is convincing. Defining understanding as whether a model trained only on local transitions can recover the latent global constraints of its world, and predict their consequences, is intuitive and well motivated. Because held-out transitions are constructed to be entailed, and proof depth quantifies inference complexity, this notion becomes measurable rather than vague.
-
The experiments are run in a relatively well-controlled setting. The two objectives are compared on identical training data, so the observed differences can be attributed to the objective rather than to data or architecture. The pattern is also consistent across four architectures and across the synthetic, visual, and language settings. This makes the main empirical finding credible.
Critical weaknesses
-
Lack of explanation of the observed phenomena. The paper shows consistent patterns across tasks but does little to explain why they occur. Throughout the paper I kept wondering what makes compositional training succeed where next-state training fails. The authors suggest that intermediate states encourage models to “solve the task locally” (L435–436), but this hypothesis is not tested, for example by probing whether intermediate states are tracked internally. The failures on periodicity (App. G) is also left unexplained. Given that the title asks whether models learn global understanding, I think some attempt to explain why they do or do not is expected. Without it, the paper mainly reports phenomena, which considerably limits its contribution.
-
Positioning relative to prior work, and a risk of overclaiming. The core idea appears closely related to Guu et al. (2015) on compositional training and Bachmann & Nagarajan (2025) on teacher-forcing shortcuts. However, these works receive only a single sentence (L467–469), which makes it hard to tell which findings are new and which confirm earlier observations in a more controlled setting. In addition, while the terms “world models” and “global understanding” are defensible in a strict sense, most of the evidence comes from small models trained from scratch in finite, deterministic worlds. The paper’s own results suggest that determinism matters: propagation fails for one-to-many converse relations, and accuracy drops from 99% to 63% when only 10% of parents have a second child (App. E). I would suggest scoping the claims more explicitly, for example to constraint propagation in deterministic, action-conditioned sequence models.
Questions for authors
-
Positioning clarification: Could the authors clarify how this work relates to and differs from Guu et al. (2015) and Bachmann & Nagarajan (2025)? In particular, which findings are new, and which confirm earlier observations in a more controlled setting? A clear answer would directly affect my assessment of the contribution.
-
Interpretation clarification: Could the authors provide more intuition or evidence for why compositional training enables constraint propagation while next-state training does not? For instance, is there any support for the hypothesis that intermediate states lead models to “solve the task locally” (L435–436), such as probing whether intermediate states are tracked internally? Relatedly, do the authors have an explanation for why some constraints are not propagated at all? Periodicity fails even when training paths contain full loops (App. G), and it would help to know whether substantially longer training (beyond 150k steps) changes this, given the grokking-like delays observed elsewhere. Even partial evidence or a clear intuition here would materially improve my assessment of the contribution.
-
Future work: The framework assumes deterministic transitions. How might entailed held-out transitions and proof depth be defined in stochastic or partially observed environments? Given the sharp drop under branching (App. E), do the authors expect compositional training to remain beneficial there?
3 comments
Reply to Reviewer SEps (2/2): what is new, larger experiments, and probabilistic worlds
Positioning relative to prior work ... which findings are new, and which confirm earlier observations
That's a good point, we rewrote that part of related work, making more explicit what we confirm with earlier work and what is new.
We confirm that hiding intermediate states improves generalization. Guu et al. (2015) showed this with knowledge-graph embeddings on multi-hop queries and single-edge completion. Bachmann & Nagarajan (2025) showed it in path-finding where a model that sees previous steps takes shortcuts instead of planning.
What is new:
- Construction of worlds where held-out data is mathematically entailed by observed data and unseen constraints. This allows us to precisely explore whether constraints are learned and propagated and make statements like: compositional training propagates inverse, commutativity, and composition, but not periodicity or converse. Their work did not construct semantically entailed worlds.
- We measure not just if a constraint is propagated but also how far through proof depth. We find that propagation decays with depth and deeper transitions are learned later. We further find that composition length promotes deep propagation (Section 6).
- We explain the benefit of compositional training on propagating constraints through a shared state representation (Section 3.3).
- We connect constraint propagation in world models to control through transitions not seen in training and show the benefit of compositional training. We find that propagated constraints are encoded in the geometry of the state embeddings and can be used for downstream tasks (Section 5.2).
Risk of overclaiming ... most of the evidence comes from small models trained from scratch in finite, deterministic worlds
We added experiments that go beyond those limitations! We finetune 4B-14B pretrained base LLMs on Wikipedia-style articles about thousands of people, finding that compositional training answers 81% of held-out questions against 0% under next-state (Appendix I). We trained 174M-parameter world models (comparable in size to DreamerV3) on 12 scanned real homes; compositional training recovers 66% of held-out transitions compared to 13% for next-state training (Appendix J). We generalized our monoid worlds and experiments to probabilistic transitions that we will talk about in the next paragraph.
How might entailed held-out transitions and proof depth be defined in stochastic or partially observed environments? Given the sharp drop under branching (App. E), do the authors expect compositional training to remain beneficial there?
This turned out to be very interesting and led to a probabilistic generalization of monoid worlds (Appendix H). Actions become Markov kernels , constraints become equalities between kernel compositions, like for commutativity, and a held-out row is entailed if every kernel that satisfies the constraint and has the seen rows agrees on the held-out row. Proof depth carries over as the number of rounds of applying the constraint needed to derive a held-out row. Inverse and periodicity don't have probabilistic generalizations and their constraints force deterministic kernels which will be important for branching.
Even with 50% noise, compositional training recovers 99-100% of held-out transitions under commutativity and composition while next-state training fails. Noise that preserves constraints in probabilistic worlds leaves propagation intact.
The drop in performance from branching can be understood as breaking the inverse constraint since the child of a person's parent may be a sibling. Adding random transitions that break the constraint by the same amount (L2 deviation) leads to similar drops (Appendix E):
\hline \text{Chance of a second child} & 0.05 & 0.1 & 0.2 & 0.5 \\\\ \hline \text{Branching} & 75\\% & 63\\% & 35\\% & 13\\% \\\\ \text{Random transition, same violation} & 71\\% & 53\\% & 41\\% & 15\\% \\\\ \hline \end{array}$$ These results suggest that determinism is not required and constraints can be propagated by compositional training in stochastic environments. We have not worked out partially observed environments yet and leave them to future work. We hope you found this interesting and would reconsider your score. Thanks!Thank you for the detailed response and the revision. The positioning of the paper compare to the prior works is now much clearer to me. Some aspects of the novelty still remain unclear in my own judgment, but most of my main concerns have been addressed. In particular, the explanation of why compositional training works, the scale-up experiments, and the experiments in probabilistic worlds make the paper stronger. I have raised my score accordingly.
Reply to Reviewer SEps (1/2): why compositional training works and periodicity fails
Thanks for the careful read and the insightful questions. Working on them helped us explain why compositional training works and why periodicity fails, as well as develop a probabilistic generalization of monoid worlds with connections to the branching failure. We hope you find them interesting.
Totally agreed, an explanation was needed and we have added Section 3.3 and analysis in Appendices K-N exactly for this.
The intuition is that next-state training leaves open degrees of freedom which compositional training fixes by hiding intermediate states. This is easier to see in a linear model where the n states are d-dimensional vectors xs, actions are d×d matrices Ma, and an n×d matrix U reads out the logits. In Appendix K we show that the linear model reproduces the behavior of our nonlinear models.
Next-state training only fits UMaxs which is invariant if we replace Ma by GMa and U by UG−1 for any invertible G. A constraint like inverse, MRML=I, does change, so the next-state objective is blind to it, hence it is not learned.
Compositional training fits UMbMaxs and UMbxa(s) to the same target, incentivizing Maxs=xa(s), which we call a shared state representation. In Appendix M we probe representations of trained transformer, GRU, LSTM, and SSM models and find a shared state representation for compositional training (cosine 0.89) but not for next-state training (0.09). Seen paths out by L and back by R then give MRMLxu=xu for many states u and since n≫d that suggests (under mild conditions) MRML=I which is the global constraint. In Appendix M we find trained linear models satisfy MRML=I only under compositional training (cosine 0.87, against 0.00 for next-state training).
Constraint propagation follows since a held-out transition (s,R,t) follows from its seen reverse xs=MLxt, thus MRxs=MRMLxt=xt.
The shared state explanation above also gives us insight into the periodicity failure. Full details are in Appendix N, but in short things break down due to the periodicity constraint involving a single action A4=1 as opposed to relating two actions as in other constraints. Without a second action, if we hold out a transition (s0,A,s1) the input state s0 and output state s1 are never trained, so a shared state representation is not learned. By explicitly "fixing" state embeddings to a proper shared state representation we get 100% propagation in linear models and 95% in transformers.
In our periodicity monoid world ⟨A,B∣A4=1⟩ we have a second arbitrary action B. However, since it is unstructured, a proper shared representation would need about n−1 dimensions, far more than d, which prevents propagation. By giving B additional structure like commutativity (even the trivial identity) ⟨A,B∣A4=1,AB=BA⟩ a shared representation is restored and constraints propagate (the held-out transitions are not entailed by the commutative structure, only by periodicity).
\hline \text{Second action } B & \text{Shared representation (cosine)} & \text{Held-out accuracy} \\\\ \hline \text{Arbitrary (original world)} & 0.73 & 0\\% \\\\ \text{Identity} & 0.96 & 99\\% \\\\ \text{Commutes with } A & 0.95 & 96\\% \\\\ \hline \end{array}$$ We also ran your longer-training test. At 1.5M steps (ten times longer) periodicity shows no sign of propagating for all four architectures (Appendix G), in agreement with the shared state explanation.