acceptodds
Under review as a conference paper at ICLR 2027

Learning Lifted World Models with Transformers

Abstract

Recent work has studied whether transformers trained on sequential data learn reusable world models. In this paper, we address this question for lifted models, i.e., abstract world models that generalize across different numbers of objects. To this end, we leverage STRIPS, a logic-based formalism for specifying lifted, interpretable world models. We propose two transformer architectures for learning STRIPS models: the Lifted STRIPS (LS) Transformer, with built-in symbolic structure and parameters that map directly to STRIPS; and the Lifted Stick-Breaking (LB) Transformer, an architecture closer to a standard Transformer that replaces softmax with stick-breaking attention and uses randomized object embeddings. The transformers are trained in a supervised manner on action traces under two supervision settings. Under applicability supervision, each action in a sequence is labeled as applicable or inapplicable given the initial state and preceding actions, and no subsequent states are observed. Under final-state supervision, each trace contains an initial state, an applicable action sequence, and the resulting final state, while intermediate states remain unobserved. Experiments on five domains show that both transformers achieve high accuracy and generalize to longer traces and larger numbers of objects. Additionally, the extracted STRIPS models are fully interpretable and can be used for planning with off-the-shelf symbolic planners.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
world models, transformers, planning, logic-based representations, neuro-symbolic learning, generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qnW6OpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes two Transformer architectures for learning lifted STRIPS world models from symbolic action traces under action-applicability or final-state supervision. The author perform experiments on five planning domains to evaluate generalization to longer planning action sequences and larger object sets, as well as planning with the extracted symbolic models.

Critical strengths
  1. The problem is well defined. The author's problem formulation is clear, the presentation for architectures and supervision settings are well-designed.
  2. The experiments provides detailed insight by evaluating the models' generalization along two dimensions: action-sequence length and the number of objects.
  3. The extracted symbolic models enable inspection of learned preconditions and effects, and their evaluation with an off-the-shelf planner provides evidence beyond predictive accuracy.
Critical weaknesses
  1. Insufficient motivation. The paper explores whether Transformers can learn reusable symbolic world models, but does not clearly establish what this adds to existing symbolic induction methods or which limitations it addresses. Moreover, potential advantages of Transformers, such as learning stochastic transition or modelling multimodal inputs, remain unexplored. Thus, it is hard for me to understand the motivation and positioning of the paper.

  2. Limited justification of the incremental contribution. Prior work has already demonstrated that Transformers can learn transition dynamics and recover propositional STRIPS models, as acknowledged in the related work section. This paper extends the setting to lifted models, but it does not sufficiently explain what new insights or advantages this extension provides. In particular, LS is designed to share the same learned rules across different object sets, so its object-count generalization may largely reflect this pre-designed structure rather than provide a new insight into the learned Transformer generalization.

  3. No demonstrated advantage over symbolic induction methods. The experiments compare only against other Transformer architectures, without evaluating existing symbolic action-model learning methods under matched supervision. Consequently, it is hard for me to identify the relative advantages of the proposed approach.

  4. Unverified applicability to context-dependent dynamics. The experiments demonstrate generalization to longer traces and larger object sets within fixed STRIPS domains, but do not verify richer context-dependent action requirements. For example, stack(a,b) may require b to be clear when b is small, whereas a larger b may be able to support multiple objects without requiring b to be clear. Whether the proposed supervision and extraction procedure can capture such dependencies remains unverified, limiting the demonstrated generalization to clean symbolic data under the tested rule systems.

Questions for authors
  1. Can the authors clarify the motivation and incremental contribution discussed in Weaknesses 1 and 2? What new insights and advantages does the lifted setting provides beyond prior work and the structure built into the new architecture?

  2. Can the authors compare against existing symbolic action-model induction methods? What are the relative advantages of using the proposed transformer?

  3. Can the authors evaluate mixed traces governed by multiple context-dependent rules, with the relevant context provided as input, to test whether the models and extraction procedure recover the appropriate conditions rather than collapse them into a single rule? For example, the preconditions should be dependent on the object size and thus result in different planned traces.

6 comments
AuthorsOpenReview ↗

Answer to Questions (1/3)

Thank you for your feedback. We answer each specific question (Q) below.

Q1. Can the authors clarify the motivation and incremental contribution discussed in Weaknesses 1 and 2? What new insights and advantages does the lifted setting provides beyond prior work and the structure built into the new architecture?

We agree that LS is designed to share rules across object sets. This is an intentional inductive bias, much as convolutional architectures exploit spatial symmetries. But this does not make successful learning or generalization automatic. The architecture defines a hypothesis class of lifted models, including many incorrect ones. Representing the correct model does not guarantee that it can be identified from limited supervision or recovered through gradient descent. We also believe that making these relational inductive biases explicit provides a useful reference for future work into whether and how less structured Transformers learn similar computations without such biases. We will briefly discuss this broader motivation in the revised paper.

Our second proposed LB architecture investigates precisely whether transformers can learn object-independent relational dynamics without explicitly encoding STRIPS rules or variable binding. Here we propose randomized object embeddings, to encourage such computations to emerge through learning and study whether similar generalization can be achieved with weaker symbolic inductive biases.

Thus, our main motivation is to investigate whether transformers can learn lifted world models that generalize to longer traces and larger object sets, a topic of great interest for the Deep Learning community [4,5], and to understand the inductive biases that enable this generalization. This is related to [3], where authors study transformer capabilities to verify plans, whereas we study transformer capabilities to learn lifted world models instead.

We now clarify our contribution, when compared to prior work on learning propositional STRIPS models [1]. In the propositional setting of [1], tokens represent ground actions from a fixed vocabulary shared between training and testing. In our lifted setting, action schemas remain fixed, but test instances contain new objects and previously unseen ground actions, with potentially many more objects that during training.

The main technical challenge is thus variable binding: the transformer must identify when preconditions and effects of different ground actions refer to the same atom. For example, if move(x,y,z) requires at(x,y) and adds at(x,z), the effect of move(a,b,c) makes true the precondition of move(a,c,d), despite involving different argument positions.

Addressing this challenge requires substantial modifications. First, the B-RASP construction (Appendix B) is considerably more complex than for the propositional case, as it needs to consider patterns when matching action preconditions and effects (see Eqs. 6-11). Second, the lifted STRIPS (LS) Transformer (see Section 5.1, Appendix C) learns lifted rules (using patterns) which are grounded on the particular objects in the traces. Doing so requires encoding objects with one-hot vectors (Eq. 16) which are permutted according to patterns (Eqs. 25-33) when computing attention. Both operations allow learned parameters to be independent of object identities and number of objects, and are absent in the propositional STRIPS Transformer. Third, the lifted Stick-Breaking (LB) Transformer leverages randomized object embeddings, consistent within each trace but resampled across presentations, to encourage object-independent reasoning without explicitly encoding symbolic rules, which is also missing in the propositional SB Transformer.

We agree that the contribution and technical challenges introduced by the lifted case were not sufficiently emphasized. We will revise Sections 4--6 to clarify this.

AuthorsOpenReview ↗

Answer to Questions (2/3)

Q2. Can the authors compare against existing symbolic action-model induction methods? What are the relative advantages of using the proposed transformer?

To the best of our knowledge, there exists no Deep Learning-based approach for learning symbolic action models which does not require state transitions, as we do in our applicability setting (where each trace contains its initial state but not its final state).

The only method applicable to this setting is SIFT [2], which is symbolic in nature. However, SIFT lacks model alignment, i.e., the extracted STRIPS model is encoded using different predicates P’ to the set of predicates P of the ground-truth domain. This means we cannot directly use the model learned by SIFT to perform planning, since problems are encoded in terms of ground-truth predicates P instead of learned predicates P’. To perform planning, we would first need to manually translate each problem to the encoding learned by SIFT, after interpreting the meaning of its learned predicates P’. In contrast, the models extracted from our transformers are aligned and can be directly used for planning (see paragraph Direct extraction via symbolic alignment in Section 6, and extracted models in Appendix G).

Another advantage of our approach compared to SIFT is that SIFT runs in time that is exponential in the number of action patterns, which are linear in the number of action names and exponential in the arity of the actions. This computational bottleneck does not arise in our approach where gradient descent is doing all the work (our LS Transformer scales quadratically in the number of action patterns, and the LB Transformer does not depend on it).

Still, we want to emphasize that our goal is not to compete with SIFT, but to study whether transformers can learn symbolic world models that generalize to different and larger sets of objects.

AuthorsOpenReview ↗

Answer to Questions (3/3)

Q3. Can the authors evaluate mixed traces governed by multiple context-dependent rules, with the relevant context provided as input, to test whether the models and extraction procedure recover the appropriate conditions rather than collapse them into a single rule? For example, the preconditions should be dependent on the object size and thus result in different planned traces.

Our current scope (lifted STRIPS models) is capable of encoding the multiple context-dependent rules you mention, e.g., your size-dependent example “stack(a,b) may require b to be clear when b is small, whereas a larger b may be able to support multiple objects without requiring b to be clear.”

Below, we provide a lifted STRIPS model that adapts Blocksworld to your example, where large blocks can have two blocks on top, and small blocks only one. As it is lifted STRIPS, both our LS and LB Transformers are capable of learning such a model:

(define (domain blocksworld-size-dependent)
(:requirements :strips)

(:constants zero one two)

(:predicates
(on ?x ?y)
(ontable ?x)
(holding ?x)
(handempty)
(occupancy ?x ?n) ;; number of blocks currently on top of ?x (replaces *clear* predicate)
(next ?n ?m)
(capacity ?x ?n) ;; number of blocks ?x can *eventually* have on top
)

(:action pick-up
:parameters (?x)
:precondition (and
(ontable ?x)
(occupancy ?x zero)
(handempty)
)
:effect (and
(holding ?x)
(not (ontable ?x))
(not (occupancy ?x zero))
(not (handempty))
)
)

(:action put-down
:parameters (?x)
:precondition (holding ?x)
:effect (and
(ontable ?x)
(occupancy ?x zero)
(handempty)
(not (holding ?x))
)
)

(:action stack
:parameters (?x ?y ?before ?after)
:precondition (and
(holding ?x)
(occupancy ?y ?before)
(next ?before ?after)
(capacity ?y ?after)
)
:effect (and
(on ?x ?y)
(occupancy ?x zero)
(handempty)
(not (holding ?x))
(not (occupancy ?y ?before))
(occupancy ?y ?after)
)
)

(:action unstack
:parameters (?x ?y ?before ?after)
:precondition (and
(on ?x ?y)
(occupancy ?x zero)
(occupancy ?y ?after)
(next ?before ?after)
(handempty)
)
:effect (and
(holding ?x)
(not (on ?x ?y))
(not (occupancy ?x zero))
(not (handempty))
(not (occupancy ?y ?after))
(occupancy ?y ?before)
)
)
)

References

[1] Núñez-Molina, C., Gómez, V., & Geffner, H. (2026). From Next Token Prediction to (STRIPS) World Models. In Proceedings of the 23rd International Conference on Principles of Knowledge Representation and Reasoning.

[2] Gösgens, J., Jansen, N., & Geffner, H. (2025, September). Learning lifted strips models from action traces alone: A simple, general, and scalable solution. In Proceedings of the International Conference on Automated Planning and Scheduling.

[3] Sarrof, Y., Du, Y., Stein, K., Koller, A., Thiébaux, S., & Hahn, M. (2026). On the Ability of Transformers to Verify Plans. arXiv preprint arXiv:2603.19954.

[4] Battaglia, P., Pascanu, R., Lai, M., & Jimenez Rezende, D. (2016). Interaction networks for learning about objects, relations and physics. Advances in neural information processing systems.

[5] Chang, M. B., Ullman, T., Torralba, A., & Tenenbaum, J. B. (2016). A compositional object-based approach to learning physical dynamics. arXiv preprint arXiv:1612.00341.

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

This paper studies whether Transformers can learn lifted STRIPS world models from symbolic action traces. It proposes two architectures: LS, which explicitly represents symbolic preconditions and effects, and LB, which uses stick-breaking attention and randomized object embeddings. Experiments across five planning domains evaluate model extraction, planning performance, and generalization to longer traces and larger object sets.

Critical strengths
  1. I find the connection between Transformers and symbolic planning interesting. The LS Transformer provides interpretable STRIPS rules, while LB explores a less structured approach. Extending previous work to object-independent representations is technically meaningful.

  2. I appreciate that the paper evaluates actual planning performance rather than relying only on prediction accuracy. The final-state supervision results are strong, and the authors have clarified several technical concerns in their rebuttal, particularly the injective-grounding assumption and action masking.

Critical weaknesses
  1. My main concern remains the gap between trace accuracy and reliable world-model recovery. The authors provide useful explanations involving sparse predicates and missing preconditions, but high prediction accuracy still does not consistently translate into successful planning.

  2. The architectural contributions of stick-breaking attention and positional encoding remain insufficiently isolated. The authors' explanation is reasonable, but controlled ablations would provide stronger evidence for the proposed design choices.

Overall, I appreciate the authors' detailed responses, which have addressed several of my original concerns. However, I still have reservations about model recovery and the architectural evidence. I therefore maintain my rating of 2 (Weak Rejection).

Questions for authors

None.

3 comments
AuthorsOpenReview ↗

Answer to Questions (1/3)

Thank you for your feedback. Below, we adress specific weaknesses (W) and questions (Q). We will also update our paper in the next few days.

Q1. Correctness under repeated object arguments. Does the LS Transformer explicitly exclude ground actions in which different action arguments refer to the same object? If not, how does Eq. (39) handle multiple effect patterns matching the same ground atom? Could the authors clarify the intended assumptions and address the counterexample described above?

We agree with the reviewer and thank them for identifying an assumption that was not explicitly stated. We assume injective groundings of domain actions, meaning that distinct action arguments must be instantiated with distinct objects. Thus, ground actions such as (\alpha(a,a)) are excluded. Since valid action patterns use distinct argument positions, injective groundings guarantee that two different effect patterns of the same action cannot match the same ground atom. Consequently, at most one term in Eq. (39) is nonzero, and the reviewer's counterexample falls outside the scope of our formulation. Avoiding repeated action arguments also prevents some ambiguous situations, e.g., if \alpha(x,y) had add effect p(x) and delete effect p(y), then \alpha(a,a) would have p(a) both as a delete and add effect, so it is unclear whether p(a) is added or removed.

We will make this assumption explicit in Section 3.1 and Appendix C. Our training and test traces already exclude non-injective ground actions, so the experimental results are unaffected.

Q2. Masking inapplicable actions. Section 4 states that preceding inapplicable actions are masked from subsequent attention. At inference time, are these masks obtained from ground-truth applicability labels or from the model's predictions? Since the test protocol contains only applicable prefixes, could the authors clarify how the architecture would handle sequences containing multiple intermediate inapplicable actions?

In our applicability setting, the task is to predict the applicability (0 or 1 label) of an action a_n given an initial state s_0 and a sequence \tau=(a_0, …, a_n-1) of applicable actions. Thus, an applicable (positive) trace contains only applicable actions, whereas an inapplicable (negative) trace consists of an applicable prefix followed by a single inapplicable action. This is the setting we evaluate at test time, where no additional masking is needed.

At training time, we found that using traces where only the last action may be inapplicable resulted in substantial class imbalance. To address this, we combine several applicable and inapplicable actions into a single mixed training trace. This fixes class imbalance. To train on these mixed traces, we mask preceding inapplicable actions using the ground truth labels, treating them as no-ops. Note that this is only part of the training implementation; our learning task assumes applicable prefixes.

We agree that this distinction is not sufficiently clear in Section 4. We will revise the section accordingly and describe mixed traces and masking in Section 7 as part of the experimental protocol.

AuthorsOpenReview ↗

Answer to Questions (2/3)

Q3. World-model fidelity. How do the extracted models compare with ground-truth STRIPS schemas in terms of precondition/effect precision and recall, or transition accuracy over independently sampled states? Can the authors explain why some configurations achieve perfect trace accuracy but substantially lower planning accuracy, particularly LS GT in Logistics and LS Shared in Blocksworld?

We analyze the mismatch between trace and planning accuracy for the three configurations below.

  • LS Shared Blocksworld, applicability setting: the mismatch was due to a bug in the code. We have repeated the experiments. Now, the best seed for this configuration achieves perfect training, test and planning accuracy (previously, 2% planning accuracy). We will update Tables 1 and 2 in our revised paper. STRIPS models in Appendix G are not modified.
  • LS GT Logistics, applicability setting: we analyzed seed 1. Despite high training (99.93%) and test (99.9995%) accuracy, it obtains low planning accuracy (7%). The reason is an error in the extracted model, where action fly(x,y,z) lacks (is_airplane ?x) as precondition, which allows trucks to fly. This missing precondition causes many errors during planning, but has low impact on trace classification (train/test accuracy) because the remaining preconditions of fly(x,y,z) are enough to correctly predict applicability in most situations. In other words, the mismatch between train/test and planning accuracy is due to data imbalance/sparsity, where very few occurrences of fly(x,y,z) are inapplicable specifically due to that single precondition. If we measure predicate-wise accuracy (as requested in Q5) for fly(x,y,z) ocurrences where every precondition is true except (is_airplane ?x), then the model obtains 36.79% training accuracy and 50% test accuracy.
  • LB Logistics, final-state setting: the best seed obtains high trace accuracy (98% for training, 90% for test) but only 1% planning accuracy. This is due to errors in the STRIPS model (see Listings 3-4), where actions drive, fly and load have missing preconditions for predicate at. If we measure accuracy specifically for goal queries involving true at atoms, then the model obtains 94% training accuracy and 72% test accuracy. These prediction errors are responsible for the missing at preconditions in the extracted model.

Therefore, the mismatch between trace classification (train/test accuracy) and planning accuracy in some configurations is due to these metrics measuring different things. Trace accuracy measures prediction errors on random traces. Conversely, planning accuracy measures how errors impact planning. Errors affecting sparse predicates (see (is_airplane ?x) above) may have low impact on trace accuracy precisely due to sparsity, but may result in missing preconditions/effects in extracted STRIPS models that severely degrade planning performance.

Errors on sparse predicates could be reduced by training on predicate-balanced traces. However, this assumes additional prior knowledge, i.e., not only knowning if an action is inapplicable but, also, why it is inapplicable (e.g., due to (is_airplane ?x) being false in our previous example). As we assume this information is unknown, we do not use predicate-balanced traces. Regardless, most configurations learn successfully despite traces being random and unbalanced.

Thus, trace classification and planning accuracy are complementary, and they are both needed to evaluate model fidelity. To further reconcile train/test and planning accuracy, we will add a new appendix to the revised version providing per-class and per-predicate accuracy for all domains, as requested in Q5.

Finally, we explain why we do not evaluate precision and recall for particular preconditions and effects, as requested. For a given ground-truth model M, there may exist many models M’, M’’, … that are equivalent in terms of applicability but contain different action preconditions and effects. Take Listing 1(a) in Appendix G, for instance. This domain achieves perfect planning accuracy, as it is equivalent to the ground-truth Blocksworld domain. And yet, its (pick-up ?x) action contains (ontable ?x) as add effect, whereas it appears as delete effect in the ground-truth model. As such, evaluating precondition/effect precision and recall may be misleading. Similarly, equivalent models M’, M’’ with different effects may result in different state transitions (e.g., s→s’ for M’ and s→s’’ for M’’), so measuring “transition accuracy over independently sampled states” can also be misleading. In summary, model fidelity should be evaluated in terms of applicability, by measuring trace prediction accuracy and planning accuracy, including per-class and per-predicate accuracy as we will now include in our paper.

AuthorsOpenReview ↗

Answer to Questions (3/3)

Q4. Controlled architectural ablations. Could the authors isolate the effects of stick-breaking attention and positional encoding, for example by evaluating Softmax without positional encoding and Stick-Breaking with positional encoding? This would help substantiate the explanation given for LB's generalization advantage.

The choice of attention normalization (stick-breaking vs softmax) and positional encoding (NoPE vs Sinusoidal vs RoPE) are interdependent, so they must be considered together instead of in isolation.

Ablation 1 (Softmax without positional encoding): softmax aggregation is order-insensitive. This means that, when predicting the applicability of an action a_i, prefixes with the same actions but in different order would receive the same attention scores, e.g., (pick-up(a), put-down(a)) and (put-down(a), pick-up(a)). To solve this, introducing explicit positional information via positional encodings is the most common choice.

Ablation 2 (Stick-Breaking with positional encoding): stick-breaking (SB) attention was proposed as a replacement for positional encodings [2]. Whereas positional encodings such as Sinusoidal y RoPE encode relative distance between actions, SB permits encoding action order without encoding action distance. Therefore, combining SB (action order) with positional encodings (action distance, including order) defies the purpose of SB in the first place.

Q5. Final-state supervision and class imbalance. The abstract and Figure 1 describe observing final states, whereas Section 7 uses ten sampled ground-atom queries per trace. Could the authors clarify the observation protocol and report class-balanced or predicate-wise prediction metrics, particularly for sparse predicates in Logistics?

In final-state supervision, the task is to learn from traces containing an initial state, a sequence of applicable actions, and a final state. Unlike for the initial state, for the final state we must encode both true and false atoms. This may result in a large number of goal atoms to encode (i.e., goal queries QFQ_F), greatly increasing the length of the traces. To avoid this, we sample 10 goal atoms (queries) from the final state of the training and test traces, at random and without class-balanced weights. This suffices for learning, since our transformers generalize over objects and they do not need to see every possible grounding for each predicate. We will clarify this in the revised version.

As mentioned above, we will also add a new appendix containing class-wise (true vs false atoms) and predicate-wise accuracy results. We will also include results for the applicability supervision setting. This will provide insight into specific errors (e.g., sparse predicates in Logistics) and will make it possible to understand the mismatch between train/test and planning accuracy.

References

[1] Núñez-Molina, C., Gómez, V., & Geffner, H. (2026). From Next Token Prediction to (STRIPS) World Models. In Proceedings of the 23rd International Conference on Principles of Knowledge Representation and Reasoning.

[2] Tan, S., Yang, S., Courville, A., Panda, R., & Shen, Y. (2025, May). Scaling stick-breaking attention: An efficient implementation and in-depth study. In International Conference on Learning Representations.

Reviewer_BEeuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper extends the STRIPS transformer paradigm to learning lifted domain models. The problem addressed is significant as it addresses a large body of work on theory and practice of solving large planning problems. It is particularly impressive that this work is able to discover models with very little data (just applicable actions with observations about the final state). Although the framework is impressive, much of it carries over from prior work at KR 2026 and it is not clear how this work compares with recent literature on action model learning.

Critical strengths
  • Learning STRIPS models is a foundational representational learning technique with many potential applications and extensions towards learning world models in more general settings

  • This work extends the prior approach to learning lifted models, which further extends the scope and applicability of this line of work.

  • Empirical results show strong generalization to test problems with larger number of objects than those seen during training.

Critical weaknesses
  • As presented, the paper appears to be an incremental extension of recent work by Nunez-Molina et al. (2026)

  • It is not clear how these methods would apply beyond domains where symbolic state representations are available

  • In its current writing the paper makes several references to the aforementioned paper, to the point where it is difficult to understand much of the methodology without first reviewing the prior work. A better summarization of that work would help address this concern.

  • One of the main strengths of learning lifted models is their generalizability. However, this aspect is not evaluated empirically in comparison with existing (non-transformer) lifted action model learners.

Questions for authors
  1. Can you provide quantitative details for the planning experiments?

  2. What are the main technical challenges in extending Nunez-Molina's prior work to the lifted setting?

  3. How does the method perform if predicate and/or action names are replaced by randomly generated strings?

  4. How would this work compare with work by Gosgens et al., 2025 and Gosgens et al., 2026? Can you present a summary of empirical performance in terms of number of training samples needed and performance of the resulting models?

6 comments
AuthorsOpenReview ↗

Answer to Questions (1/2)

Thank you for your feedback. Below, we adress specific weaknesses (W) and questions (Q). We will also update our paper in the next few days.

Q1. Can you provide quantitative details for the planning experiments?

Quantitative planning results are already reported in Tables 1–2, including average planning success rates across five seeds and results for the best-training seed. Our full experimental setting (including the planning experiments) is detailed in Section 7 and Appendices D-E, and Appendix G contains the learned STRIPS models. Additionally, we will add to the revised version a new Appendix containing extended planning results. If the reviewer has specific additional metrics or analyses in mind, we would be happy to address them.

Q2. What are the main technical challenges in extending Nunez-Molina's prior work [1] to the lifted setting?

In the propositional setting of [1], tokens represent ground actions from a fixed vocabulary shared between training and testing. In our lifted setting, action schemas remain fixed, but test instances contain new objects and previously unseen ground actions, with potentially many more objects that during training.

The main technical challenge is thus variable binding: the transformer must identify when preconditions and effects of different ground actions refer to the same atom. For example, if move(x,y,z) requires at(x,y) and adds at(x,z), the effect of move(a,b,c) makes true the precondition of move(a,c,d), despite involving different argument positions.

Addressing this challenge requires substantial modifications. First, the B-RASP construction (Appendix B) is considerably more complex than for the propositional case, as it needs to consider patterns when matching action preconditions and effects (see Eqs. 6-11). Second, the lifted STRIPS (LS) Transformer (see Section 5.1, Appendix C) learns lifted rules (using patterns) which are grounded on the particular objects in the traces. Doing so requires encoding objects with one-hot vectors (Eq. 16) which are permutted according to patterns (Eqs. 25-33) when computing attention. Both operations allow learned parameters to be independent of object identities and number of objects, and are absent in the propositional STRIPS Transformer. Third, the lifted Stick-Breaking (LB) Transformer leverages randomized object embeddings, consistent within each trace but resampled across presentations, to encourage object-independent reasoning without explicitly encoding symbolic rules, which is also missing in the propositional SB Transformer.

We agree that the technical challenges introduced by lifting were not sufficiently emphasized. We will revise Sections 4--6 to clarify this.

Q3. How does the method perform if predicate and/or action names are replaced by randomly generated strings?

Unlike pre-trained LLMs, our transformers do not exploit the semantic meaning of action or predicate names, but only their discrete IDs. For example, renaming the action move to a7 and the predicate at to p3 would have no effect on the model.. Therefore, replacing names with arbitrary strings leaves predictions unchanged, provided that the renaming is one-to-one and consistent across training and testing.

AuthorsOpenReview ↗

Answer to Questions (2/2)

Q4. How would this work compare with work by Gosgens et al., 2025 and Gosgens et al., 2026? Can you present a summary of empirical performance in terms of number of training samples needed and performance of the resulting models?

SIFT [2] is a symbolic, constraint-based learning algorithm that constructs lifted models by identifying features consistent with observed traces. In contrast, our work formulates lifted world-model learning as a statistical learning problem, using differentiable transformer architectures trained by gradient descent. Our goal is not to outperform SIFT in sample efficiency, but to investigate the capabilities of transformers to learn symbolic world models, focusing on their generalization to longer traces and different, larger sets of objects. Thus, our work aligns with [3], which studies whether transformers can verify plans. Notice that plan verification can be trivially performed by symbolic algorithms and, yet, it is interesting to study whether transformers are expressive enough to perform plan verification. Analogously, studying whether transformers can learn lifted world models provides insight into their expressivity and generalization capabilities, regardless of the existence of symbolic learning algorithms like SIFT.

Regardless, our approach offers several advantages over SIFT:

  • Differentiability: our approach could potentially be integrated with perceptual learning components as in ROSAME [4,5], e.g., to learn from images.
  • Model alignment: SIFT lacks model alignment, meaning that the learned STRIPS model is encoded using a set of predicates P’ different to those P of the ground-truth domain. Since problems are also encoded in terms of P, we cannot use the model learned by SIFT to perform planning, unless we first manually translate each problem to the encoding learned by SIFT (in terms of P’). Unlike SIFT, our transformers exploit initial-state observations to align the learned model, so it can be directly used for planning (see paragraph Direct extraction via symbolic alignment in Section 6).
  • Scalability: SIFT runs in time that is exponential in the number of action patterns, which are linear in the number of action names and exponential in the arity of the actions. This computational bottleneck does not arise in our approach, where learning is done via gradient descent (our LS Transformer scales quadratically in the number of action patterns, and the LB Transformer does not depend on it).

W2. It is not clear how these methods would apply beyond domains where symbolic state representations are available.

Our current experiments assume symbolic action traces and initial states. However, the LS Transformer could be adapted to learn from action traces alone, without observing any states, following the action-only formulation of SIFT [1]. In this case, the learned STRIPS models would not be aligned, meaning they could not be directly used for planning (see Model alignment above).

More generally, our transformer architectures are differentiable and, thus, could be integrated with perceptual components, allowing models to learn from non-symbolic observations such as images, as in ROSAME [4,5]. This is an important direction for future work, but beyond the scope of the present paper.

References

[1] Núñez-Molina, C., Gómez, V., & Geffner, H. (2026). From Next Token Prediction to (STRIPS) World Models. In Proceedings of the 23rd International Conference on Principles of Knowledge Representation and Reasoning.

[2] Gösgens, J., Jansen, N., & Geffner, H. (2025, September). Learning lifted strips models from action traces alone: A simple, general, and scalable solution. In Proceedings of the International Conference on Automated Planning and Scheduling.

[3] Sarrof, Y., Du, Y., Stein, K., Koller, A., Thiébaux, S., & Hahn, M. (2026). On the Ability of Transformers to Verify Plans. arXiv preprint arXiv:2603.19954.

[4] Xi, K.; Gould, S.; and Thiébaux, S. (2024). “Neuro-Symbolic Learning of Lifted Action Models from Visual Traces”. In Proceedings of the International Conference on Automated Planning and Scheduling.

[5] Xi, K., Gould, S., & Thiébaux, S. (2026, June). Learning Lifted Action Models from Unsupervised Visual Traces. In Proceedings of the International Conference on Automated Planning and Scheduling.

AuthorsOpenReview ↗

Comparison with SIFT

Dear Reviewer, we provide below the results of our comparison with SIFT (Gosgens et al., 2025), as requested in Question 4. We empirically show that SIFT is intractable for certain domains due to its exponential complexity in the number of action patterns, unlike our transformers.

We run SIFT and our LS Transformer on a domain containing 50 zero-arity actions. SIFT results in a timeout after 12h. For this domain, it needs to evaluate 2^50-1 features, requiring around 8500 years to finish. Our LS Transformers are trained in under an hour. Accuracy results are below:

ModelTrain acc.Test acc.Plan acc.
LS final-state (ours)0.99 (1.0)0.99 (1.0)0.93 (1.0)
LS applicability (ours)0.41 (0.54)0.30 (0.67)0.04 (0.05)

Unlike our transformers, SIFT learns models that are unaligned, meaning they use different predicates to the ground-truth domain and cannot be used for planning. To illustrate this, we show below the blocksworld action pick-up learned by SIFT and our LS Transformer:

SIFT

(:action pick-up
:parameters (?Arg0)
:precondition (and
(Feature_3_v0_s1 ?Arg0) (Feature_2_v0_s1 ?Arg0) (Feature_1_v0_s1 ?Arg0) (Feature_0_v0_s1)
(Static_pick-up ?Arg0)
)
:effect (and
(Feature_3_v0_s0 ?Arg0) (Feature_1_v0_s0 ?Arg0) (Feature_2_v0_s0 ?Arg0) (Feature_0_v0_s0)
(not (Feature_3_v0_s1 ?Arg0)) (not (Feature_2_v0_s1 ?Arg0)) (not (Feature_1_v0_s1 ?Arg0)) (not (Feature_0_v0_s1))

LS Transformer (final-state supervision, see Listing 1b)

(:action pick-up
:parameters (?x)
:precondition (and
(clear ?x) (handempty) (ontable ?x))
:effect (and
(holding ?x) (not (handempty)) (not (clear ?x)) (not (ontable ?x))))

Comments on the paper

1 comment
AuthorsOpenReview ↗

Paper Modifications

Dear Reviewers,

Thank you for your constructive feedback. We have revised the manuscript to address the main concerns raised in the reviews. Modifications are highlighted in blue. The principal changes are summarized below.

  • Additional planning results (BEeu Q1). We have added Appendix H, providing additional planning metrics such as planning time and types of planning errors.

  • Class-wise and Predicate-wise accuracy (EdmJ Q5). We have added Appendix I, breaking down accuracy by class and predicate. Based on these results, we explain the mismatch between trace classification accuracy and planning accuracy observed in some configurations.

  • Motivation, novelty, and technical challenges (BEeu Q2, qnW6 Q1). We revised the introduction to clarify the distinction between propositional and lifted STRIPS learning and emphasize variable binding as the main technical challenge. We expanded Appendices B and C to explain the corresponding modifications to the B-RASP construction and LS Transformer. We also expanded Discussion and Conclusions to distinguish the ability of LS to represent object-independent rules from the challenge of learning accurate rules through gradient descent, and to clarify the complementary role of LB, which learns relational computations without explicitly encoding STRIPS rules.

  • Comparison with symbolic approaches (BEeu Q4). We expanded Discussion and Conclusions to distinguish our statistical, gradient-based formulation from symbolic approaches such as SIFT and SIFT+. We discuss differences in sample efficiency and supervision, as well as model alignment, computational scaling with action patterns, and the potential integration of perceptual components.

  • Repeated object arguments (EdmJ Q1). We made the injective-grounding assumption explicit in Section 3.1 and clarified in Appendix C why this assumption guarantees that at most one effect pattern matches a given ground atom in Eq. (39).

  • Masking inapplicable actions (EdmJ Q2). We clarified in Section 4 that the applicability-supervision task assumes applicable prefixes, with only the final action potentially inapplicable. We moved the description of mixed training traces and ground-truth masking to Section 7 and clarified the corresponding implementation in Appendix C.

We hope these revisions satisfactorily address the reviewers’ concerns, and we would be happy to provide further clarification on any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.