acceptodds
Under review as a conference paper at ICLR 2027

Do World Models Learn Global Understanding?

Abstract

AI systems often feel frustratingly brittle and fragmented. A large language model (LLM) may correctly explain a concept but fail to apply it, or follow safety instructions in one context but not another. This behavior suggests a general failure to lift local information to a global understanding. To gain fundamental insight, we frame "understanding" as learning constraints and propagating their consequences. We construct learning tasks on monoid worlds, sets of states connected by action transitions (state, action, next state), where observed training transitions and an underlying unseen constraint jointly determine held-out transitions. Measuring generalization tests whether models can learn global constraints from local transitions and propagate their consequences. We consider inverse, commutativity, composition, and periodicity constraints relevant to spatial and semantic structure. Across attention, recurrent, and state-space architectures, next-state training on paths of observed transitions fits the data but fails to propagate non-trivial constraints. Compositional training, which uses identical paths but hides intermediate states from the input, achieves 96% accuracy on inverse, commutativity, and composition constraints across architectures, yields corresponding improvements in geometric generalization of world models trained on embodied environments (+42% on propagating commutative constraints) and relational generalization in Wikidata-finetuned LLMs (+73% on composition propagation). How far do models propagate constraints when inferring an unseen fact may depend on first inferring others? We define proof depth of a held-out transition, measuring the minimum number of inference rounds to infer the transition, and find that model generalization decreases sharply with increasing proof depth. A compositionally trained transformer on length paths generalizes at 95% for transitions derivable directly from observations () but only at 49% for those requiring an additional inference round (). Increasing compositional path length improves generalization, and compositional models achieve 95% accuracy on transitions. These results provide a formal way to investigate global understanding in language and world models and demonstrate that compositional training promotes information propagation and integration.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
World Models, Constraint Propagation, Algebraic Structure, Monoid Actions, Systematic Generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SEpsOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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
  1. 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.

  2. 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.

  3. 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
AuthorsOpenReview ↗

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.

Lack of explanation of the observed phenomena ... what makes compositional training succeed where next-state training fails

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 nn states are dd-dimensional vectors xsx_s, actions are d×dd \times d matrices MaM_a, and an n×dn \times d matrix UU 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 UMaxsU M_a x_s which is invariant if we replace MaM_a by GMaG M_a and UU by UG−1U G^{-1} for any invertible GG. A constraint like inverse, MRML=IM_R M_L = I, does change, so the next-state objective is blind to it, hence it is not learned.

Compositional training fits UMbMaxsU M_b M_a x_s and UMbxa(s)U M_b x_{a(s)} to the same target, incentivizing Maxs=xa(s)M_a x_s = x_{a(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 LL and back by RR then give MRMLxu=xuM_R M_L x_u = x_u for many states uu and since n≫dn \gg d that suggests (under mild conditions) MRML=IM_R M_L = I which is the global constraint. In Appendix M we find trained linear models satisfy MRML=IM_R M_L = 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)(s, R, t) follows from its seen reverse xs=MLxtx_s = M_L x_t, thus MRxs=MRMLxt=xtM_R x_s = M_R M_L x_t = x_t.

Lack of explanation ... failures on periodicity ... whether substantially longer training (beyond 150k steps) changes this

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=1A^4=1 as opposed to relating two actions as in other constraints. Without a second action, if we hold out a transition (s0,A,s1)(s_0, A, s_1) the input state s0s_0 and output state s1s_1 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⟩\langle A,B \mid A^4=1 \rangle we have a second arbitrary action BB. However, since it is unstructured, a proper shared representation would need about n−1n-1 dimensions, far more than dd, which prevents propagation. By giving BB additional structure like commutativity (even the trivial identity) ⟨A,B∣A4=1,AB=BA⟩\langle A,B \mid A^4=1, AB=BA \rangle 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.
AuthorsOpenReview ↗

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:

  1. 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.
  2. 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).
  3. We explain the benefit of compositional training on propagating constraints through a shared state representation (Section 3.3).
  4. 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 PaP_a, constraints become equalities between kernel compositions, like PAPB=PBPAP_A P_B = P_B P_A 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.

Reviewer_X6bmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies whether world models and language models acquire global understanding. The authors define "understanding" as learning constraints and propagating their consequences. To study this, they construct a simple environment with controllable constraints and compare the "understanding" of models trained with next-state and compositional objectives. Next-state training is largely unable to infer unseen transitions. Compositional training can infer them under inverse, commutativity, and composition constraints, but fails under periodicity constraints. The authors observe similar trends in visual world models and fine-tuned LLMs. Also the authors define the proof depth of a test transition, and their experiments show that accuracy drops sharply as proof depth increases, while a longer composition length T during training allows models to infer transitions at greater depths.

Critical strengths
  1. The experimental setup and environments are rigorous and clean, with well-controlled variables, which makes the conclusion that "the training objective determines whether constraints are propagated" convincing.
  2. Proof depth is well defined. The experiments show that accuracy decreases as proof depth increases and that models learn transitions in order of depth, and the authors test the cascade hypothesis with a control that shuffles the dependency structure.
Critical weaknesses
  1. The paper does not explain several key failure cases. For example, periodicity constraints are not propagated even under compositional training; converse constraints are not propagated under either training objective, and the authors only show that accuracy drops as branching increases, without explaining why.
  2. Although the experimental setting is rigorous, it is somewhat toy and far from world models and llms used in practice, and the experiments are also small in scale.
Questions for authors
  1. Could the authors provide some experiments and explanations, even preliminary ones, for the failure cases? For example, why are periodicity constraints not propagated even under compositional training, and why does branching break the propagation of converse constraints?
  2. Could the authors add experiments at a larger scale and in settings closer to real-world world-model environments?
  3. The figures could use more polish. For example, in Figure 1d(1), next-state training is plotted only as a single point at T=1T = 1, so the training objective and the composition length share the same x-axis, which makes the labeling somewhat confusing. I would suggest further revising all figures in the paper for clarity.
2 comments
AuthorsOpenReview ↗

Reply to Reviewer X6bm (1/2): why periodicity and converse fail

Thank you for reading our paper and your thoughtful comments. We added additional sections for explanation and additional experiments.

The paper does not explain several key failure cases ... periodicity constraints are not propagated even under compositional training

Agreed, we added explanations for why compositional training works (new Section 3.3 and Appendix K-M) and where periodicity fails and how to fix it (Appendix N).

Intuitively, compositional training incentivizes learning a shared state representation where applying an action aa to a state vector xsx_s yields the same representation as the next state xa(s)x_{a(s)}. Constraints propagate through this representation. We show this using linear analysis in Appendix L which we motivate in Appendix K. We find a shared state representation in compositionally trained models (cosine 0.89) and not in next-state trained models (0.09) (Appendix M).

Periodicity fails because this representation is not learned (Appendix N). Periodicity A4=1A^4 = 1 is the only constraint in our worlds that involves a single action. The held-out transition (s3,A,s0)(s_3, A, s_0) is the only transition by AA out of s3s_3 and into s0s_0, so these two states are trained only through the second action BB. Since BB is arbitrary in our periodicity world, a shared representation would need about N−1N-1 dimensions, far more than the model width dd, which prevents propagation. By giving BB additional structure like commutativity (even the trivial identity) ⟨A,B∣A4=1,AB=BA⟩\langle A,B \mid A^4=1, AB=BA \rangle 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}$$ We can instead also explicitly "fix" state embeddings to a proper shared state representation for just $A$ and we get 100% propagation in linear models and 95% in transformers. > converse constraints are not propagated under either training objective, and the authors only show that accuracy drops as branching increases, without explaining why We added an explanation for why branching breaks propagation (Section 4.2 and Appendix E) which connects to a new probabilistic generalization of monoid worlds (Appendix H). Intuitively, branching breaks the inverse constraint. Without branching every parent has one child so parent and child are inverses which compositional training propagates at 99%. With branching, the child of a person's parent may be a sibling, so the inverse constraint only holds sometimes. We took a line world without branching and added random transitions that break the inverse constraint by the same amount (L2 deviation) and got similar drops: $$\begin{array}{lcccc} \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}$$ This suggests that branching can be understood as noise on the inverse constraint. Randomness itself is not the problem. In our probabilistic worlds transitions are noisy but constraints still hold. Even with 50% noise, compositional training recovers 99-100% of held-out transitions under commutativity and composition while next-state training fails. So noise that preserves constraints leaves propagation intact as opposed to noise that breaks them like branching.
AuthorsOpenReview ↗

Reply to Reviewer X6bm (2/2): larger experiments and cleaner figures

Although the experimental setting is rigorous, it is somewhat toy and far from world models and llms used in practice, and the experiments are also small in scale.

Agreed! We did four things. For embodied world models we scaled up to 174M parameters (comparable in size to DreamerV3) trained on 12 scanned real homes, with turns and heading-relative steps (Appendix J). For language we scaled to 4B-14B LLMs trained on Wikipedia-style articles about thousands of people instead of templated sentences and tested in QA format (Appendix I). For control, which is what world models are used for in practice, we added goal reaching in AI2-THOR and a Habitat scan of a real home (new Section 5). We also took three world-model benchmarks in physics prediction, planning, and weather forecasting, where we change only the training objective (Appendix P). Compositional training improves all of them:

\hline \text{Experiment} & \text{Next-state} & \text{Compositional} \\\\ \hline \text{Scanned homes (held-out accuracy)} & 13\\% & \mathbf{66\\%} \\\\ \text{Wikipedia-style articles (held-out questions)} & 0\\% & \mathbf{81\\%} \\\\ \text{Goal reaching in a real home (SPL, search)} & 0.34 & \mathbf{0.93} \\\\ \text{3-body physics (10-step prediction)} & 74\\% & \mathbf{90\\%} \\\\ \text{Two-Room (planning success)} & 87\\% & \mathbf{94\\%} \\\\ \text{Weather (5-day forecast error, lower is better)} & 1519 & \mathbf{874} \\\\ \hline \end{array}$$ > The figures could use more polish. For example, in Figure 1d(1), next-state training is plotted only as a single point ... We fixed Figure 1d(1) so it now plots both objectives at every path length with a legend, so the objective and the path length no longer share an axis. Thanks for pointing that out. We also cleaned up the other figures: legends for held-out transitions (Figure 3), color scale (Figure 5), and consistent labels and wording throughout. We rewrote captions and sections for clarity. We hope this was helpful and we would appreciate reconsidering your score. Thanks!
Reviewer_ikToOpenReview ↗
Rating
2
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper explore whether next state prediction can achieve the global understanding. The authors use controlled monoid worlds to show that next-state training memorizes observed transitions but fails to generalize non-trivial constraints, whereas compositional training achieves about 96% accuracy on inverse, commutative, and compositional constraints. Similar improvements appear in AI2-THOR visual environments and Wikidata relational reasoning tasks. However, performance decreases sharply as proof depth increases, indicating that current models still struggle with deep constraint propagation.

Critical strengths
  1. The paper formalizes global understanding as learning constraints.
  2. The controlled experiments are designed well. Next-state and compositional training use identical paths, isolating the effect of the training objective.
  3. The findings are consistent across Transformers, GRUs, LSTMs, and SSMs.
  4. The authors introduce “Proof depth” to evaluate how far models can propagate learned constraints.
  5. The framework is evaluated in synthetic worlds, AI2-THOR visual environments, and Wikidata relational tasks.
Critical weaknesses
  1. Constraint propagation captures only one aspect of global understanding. There are still planing, reasoning and so on.
  2. The experiments are in synthetic datasets, not the real downstream tasks. AI2-THOR uses simple movements, while the Wikidata experiment relies on only hundreds of renamed entities and templated sentences.
  3. Since compositional training directly practices action-sequence composition, it is more matching the evaluation task than the next state prediction.
  4. Accuracy declines rapidly with proof depth and approaches chance beyond a few inference rounds.
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer ikTo: we added planning and control and larger experiments

Thanks for reading our paper and for the thoughtful review. These were great points and we made many additions.

Constraint propagation captures only one aspect of global understanding. There are still planing, reasoning and so on.

Agreed! So we added a new section on planning and control with world models (Section 5). We look at three control methods: search in the model, policy learning in its imagination, and policy learning on its representations. The task is goal reaching in monoid and embodied worlds (AI2-THOR, and a newly added Habitat scan of a real home with a larger state and action space). Every shortest path to a test goal crosses a held-out transition, so reaching it efficiently requires inference by the world model. Compositionally trained world models beat next-state trained ones with every method in every world.

SPL (success weighted by path length), next-state / compositional, better in bold:

\hline \text{Method} & \text{Inverse} & \text{Commutative} & \text{AI2-THOR} & \text{Habitat} \\\\ \hline \text{Search} & 0.05 / \mathbf{0.97} & 0.25 / \mathbf{0.99} & 0.48 / \mathbf{0.80} & 0.34 / \mathbf{0.93} \\\\ \text{Imagination} & 0.44 / \mathbf{0.96} & 0.34 / \mathbf{0.97} & 0.40 / \mathbf{0.72} & 0.32 / \mathbf{0.78} \\\\ \text{Representation} & 0.08 / \mathbf{0.94} & 0.10 / \mathbf{0.77} & 0.38 / \mathbf{0.64} & 0.37 / \mathbf{0.64} \\\\ \hline \text{Optimal control, seen transitions only} & 0.01 & 0.14 & 0.40 & 0.50 \\\\ \hline \end{array}$$ More results, including success and optimal rates, are in Appendix O if you're curious. Interestingly, the Representation results suggest that propagated constraints are encoded in the geometry of the state representations and are transferable. > The experiments are in synthetic datasets, not the real downstream tasks. AI2-THOR uses simple movements, while the Wikidata experiment relies on only hundreds of renamed entities and templated sentences. Also agreed. We did four things. For embodied world models we scaled up to 174M parameters (comparable in size to DreamerV3) trained on 12 scanned real homes, with turns and heading-relative steps (Appendix J). For language we scaled to 4B-14B LLMs trained on Wikipedia-style articles about thousands of people instead of templated sentences and tested in QA format (Appendix I). For downstream tasks we added goal reaching in AI2-THOR and Habitat (Section 5, table above) and three world-model benchmarks in physics prediction, planning, and weather forecasting, where we change only the training objective (Appendix P). Compositional training improves all of them: $$\begin{array}{lcc} \hline \text{Experiment} & \text{Next-state} & \text{Compositional} \\\\ \hline \text{Scanned homes (held-out accuracy)} & 13\\% & \mathbf{66\\%} \\\\ \text{Wikipedia-style articles (held-out questions)} & 0\\% & \mathbf{81\\%} \\\\ \text{3-body physics (10-step prediction)} & 74\\% & \mathbf{90\\%} \\\\ \text{Two-Room (planning success)} & 87\\% & \mathbf{94\\%} \\\\ \text{Weather (5-day forecast error, lower is better)} & 1519 & \mathbf{874} \\\\ \hline \end{array}$$ > Since compositional training directly practices action-sequence composition, it is more matching the evaluation task than the next state prediction. Main evaluation is on single transitions (given a state and an action, what is the next state?), not on action-sequence composition. Next-state training practices exactly this task, so it is not at a disadvantage. We revised Section 2.3 to make this clearer. > Accuracy declines rapidly with proof depth and approaches chance beyond a few inference rounds. Agreed. This is a limitation of current models and we view it as a key finding. We introduce proof depth to formalize how far information propagates (Section 6). Compositional training on longer paths does help and appears to prevent full collapse to chance: Held-out accuracy (%) by proof depth $d$, for composition length $T$ (chance is 0.1%): $$\begin{array}{lcccccccc} \hline & d=1 & d=2 & d=3 & d=4 & d=5 & d=6 & d=7 & d=8 \\\\ \hline T=2 & 95 & 49 & 15 & 5 & 1 & 0 & 0 & 0 \\\\ T=3 & 100 & 92 & 50 & 32 & 21 & 13 & 11 & 11 \\\\ T=4 & 98 & 95 & 77 & 48 & 35 & 25 & 21 & 18 \\\\ \hline \end{array}$$ Designing new learning methods that don't suffer from proof depth decay is something we are actively interested in and a topic of future work. Hope this helps and was interesting. We would also appreciate reconsidering your score. Thanks!

Thanks for the clarifications and new experiments. However, some aspects of the novelty remain unclear to me. What's more, the language model task relies on hundreds of renamed entities, and there are no results for the other downstream tasks, such as GSMK 8k or MMLU. Thus, I’ll keep my overall score, but will increase the Informativeness scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.