acceptodds
Under review as a conference paper at ICLR 2027

ATLAS: From World Laws to Skills in LLM-Guided Hierarchical Reinforcement Learning

11 views

Abstract

Long-horizon reinforcement learning (RL) often relies on structured skill specifications to overcome sparse rewards and long dependency chains. Existing LLM-Hierarchical Reinforcement Learning (LLM-HRL) methods typically ask language models to both infer environment dynamics and derive skill preconditions, effects, and dependencies, making specification errors difficult to avoid. We propose ATLAS (Acquiring and Testing Laws for Actionable Skills), a world-law-driven LLM-HRL framework in which explicit world laws connect language-model reasoning with skill construction and learning. ATLAS represents environment dynamics as a BC+ world model and deterministically derives executable skill specifications from accepted laws for hierarchical RL. Test-driven development provides validation constraints on law construction and revision through symbolic unit tests and environment-level tests. To acquire and revise laws from experience, ATLAS uses successful and failed trajectories together with counterfactual experiments to check candidate world laws. Accepted revisions propagate to skill specifications and training, translating evolving environment knowledge into continually updated execution capabilities. With detailed Craftax-Classic materials, the generated laws support operator grounding and achievement completion across resource and tool dependencies; in an unfamiliar Mars world, interaction evidence supports law acquisition without supplied environment-specific rules.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Large Language Model

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_97xqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces ATLAS, a framework for LLM-guided hierarchical reinforcement learning that introduces an explicit world-law layer between language-model reasoning and skill learning. An LLM constructs and revises environmental laws in BC+, while independently constructed symbolic and environment-level tests constrain their acceptance. A deterministic compiler extracts skill prerequisites, effects, resource costs, and dependencies, which are grounded through separately trained RL policies and a curriculum that revisits downstream skills as prerequisites improve. Execution trajectories and controlled interventions support further law acquisition and revision. Experiments demonstrate improved agreement with a fixed behavioral test suite, successful execution of long prerequisite chains in Craftax-Classic, and an illustrative acquisition of unfamiliar rules in Mars.

Critical strengths
  1. The separation between world modeling and skill specification is well motivated. ATLAS maintains a shared BC+ theory and deterministically derives skill prerequisites, effects, resource consumption, and dependencies. This provides a coherent way to reuse environmental knowledge across skills and identify which specifications need updating when a law changes. In particular, retaining the supporting laws for each specification gives the proposed revision process a useful structure.
Critical weaknesses
  1. The experiments do not isolate the benefit of the central architectural choice. The main argument is that learning world laws and deriving specifications deterministically is more reliable than having an LLM generate specifications directly. However, there is no comparison that holds the surrounding system fixed while changing the source of the specifications.
  2. The reported results have several evaluation concerns. In Craftax, ATLAS reports the best evaluation within a training task rather than a separate final evaluation, and its results come from a single training run, whereas CODE-SHARP reports averages across five runs. The equivalence of their survival settings also needs clarification. In Mars, ATLAS and the baseline are evaluated under different world configurations and procedures, making the reported performance improvement difficult to interpret. These inconsistencies weaken the empirical evidence for ATLAS's superiority.
Questions for authors
  1. Does the world-law representation improve over direct specification generation under matched conditions?

  2. Do the principal performance gains survive a matched evaluation protocol?

1 comment

Dear Reviewer 97xq,

Thank you again for your time and constructive feedback.

As approximately two days remain in the author–reviewer discussion period, we would be most grateful if you could take a moment to review our rebuttal. In it, we have addressed your concerns through additional analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and welcome your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_t3p5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ATLAS, an LLM-HRL framework that represents environment dynamics as a BC+ world model (which the paper refers to as world laws), uses it to derive skill specifications from it, trains RL policies for the skills, and uses a controller to select between skills to solve long horizon RL tasks. The BC+ world model is revised from unit tests and the experience trajectories. ATLAS uses various techniques during trajectory-based revisions including trajectory simplification, semantic purification, and counterfactual experiments. The updated BC+ world model is then used to refine the skill specifications and training. Their main contribution is the ATLAS LLM-HRL pipeline, and the shared world law representation for the skills derived from it. They validate their result on Craftax-Classic and Mars, notably showing a large improvement in collecting diamonds, one of the more difficult long-horizon tasks in the Craftax-classic benchmark.

Critical strengths

S1. The idea of world laws, maintaining a shared environment dynamics representation is interesting and may be useful for maintaining consistency across skills.

S2. The improvement over baselines in the Craftax-classic diamond collection task seems substantial. This demonstrates promising empirical performance on a challenging long-horizon task.

Critical weaknesses

W1. The paper presents several techniques for trajectory based refinement of the world laws. However, there is a lack of systematic ablations that show the motivation for these choices (trajectory simplification, counterfactual validation). This seems important considering one of the main contributions listed in the introduction is the method for revising world laws using trajectory evidence and counterfactual validation.

W2. The paper's presentation could benefit from giving a small example or illustration of what the synthesized BC+ laws and the skill specifications look like.

Questions for authors
  1. Could the authors address the weaknesses raised above, particularly I am wondering about the lack of systematic ablations for the trajectory-based world law refinement?
4 comments

Dear Reviewer t3p5,

Thank you again for your time and constructive feedback.

As approximately two days remain in the author–reviewer discussion period, we would be most grateful if you could take a moment to review our rebuttal. In it, we have addressed your concerns through additional analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and welcome your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Dear reviewer t3p5,

We fully acknowledge that the paper still has room for improvement. Throughout the rebuttal, we have made every effort to address the two concerns you raised, including providing additional analyses and revising the manuscript.

In particular:

  • W1 (Trajectory-based refinement): We explained the motivation behind trajectory simplification and semantic purification, and added a quantitative analysis of the refinement process. Specifically, the candidate state fields are reduced from 35 to 5 and 7 for two representative tasks, requiring 32 and 128 counterfactual queries, respectively.
  • W2 (Concrete examples): We added a detailed iron-pickaxe crafting example illustrating how BC+ world laws are represented and how the compiler derives the corresponding skill specifications, including prerequisites, effects, resource consumption, and dependencies.

At the same time, we hope you might reconsider the overall assessment in light of the positive aspects you highlighted in your review, particularly:

  • Reusable world laws and consistency across skills:

    • Your assessment: "The idea of world laws, maintaining a shared environment dynamics representation is interesting and may be useful for maintaining consistency across skills."
    • Reviewer 97xq similarly considered the separation between world modeling and skill specification "well motivated," noting that it provides "a coherent way to reuse environmental knowledge across skills." Reviewer q62x also recognized the "clean decomposition" and the use of a shared BC+ layer whose revisions propagate to dependent skills.
  • End-to-end performance improvement:

    • Your assessment: "The improvement over baselines in the Craftax-classic diamond collection task seems substantial. This demonstrates promising empirical performance on a challenging long-horizon task."
    • Reviewer q62x similarly recognized that ATLAS achieved a 45.7% success rate for mining diamonds, "significantly outperforming" the 16.0% reported by CODE-SHARP. Our updated evaluation across five independent training seeds further demonstrates a 42.65% mean success rate (SD: 3.28 percentage points), retaining a substantial improvement over the baseline.

These are central aspects of ATLAS, and we sincerely appreciate your recognition of the potential value of our approach.

Taking a step back and considering the paper as a whole, particularly its technical contributions and the additional analyses, experiments, and clarifications provided during the rebuttal, we sincerely hope you might reconsider whether a Weak Accept would better reflect the overall merits of the paper, even with some remaining issues and concerns.

Thank you again for your thoughtful feedback and the time and effort you have devoted to our paper.

Dear Authors,

Thank you for your detailed rebuttal and efforts to improve the paper. They address my concerns. I have raised my score to 3.

Dear Reviewer,

Thanks for your positive response and for raising your score to 3. We greatly appreciate the time and effort you invested in reviewing our paper, and we are glad that our rebuttal and revisions have addressed your concerns.

We will carefully incorporate all the suggestions from your review into the final version of the paper to further improve its quality. Thank you again for your constructive comments and for your support of our work.

Sincerely, The Authors

Reviewer_q62xOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ATLAS, a framework for LLM-guided hierarchical reinforcement learning that decouples the task of "understanding environmental dynamics" from "deriving skill specifications." The process begins with an LLM using an action language, BC+, to articulate the rules governing the environment. These rules must pass a suite of pre-constructed symbolic unit tests and environmental tests; if a test fails, the feedback is provided to the LLM for refinement. Subsequently, a symbolic compiler deterministically derives the preconditions, effects, resource consumption, and interdependencies for each skill based on the validated rules. Each skill is trained using an independent PPO policy, with training proceeding hierarchically according to the established dependencies. For skills with low success rates, the system collects trajectories, proposes candidate rules, and verifies them through counterfactual experiments to update the environmental model. The experimental evaluation consists of three parts. On Craftax, test-guided refinement increased the average test pass rate for rules generated by four different models from 52.5% to 99.6%. On Craftax-Classic, the system achieved a 45.7% success rate for mining diamonds, significantly outperforming the 16.0% reported by CODE-SHARP. In the "Mars" environment—characterized by counterintuitive rules—the system learns the environmental dynamics through interaction (without being provided with the rules beforehand), achieving a score of 19.69% compared to a baseline of 2.45%.

Critical strengths
  1. Clean decomposition with a working symbolic toolchain. The paper separates LLM law authoring from deterministic specification derivation, using a shared BC+ layer whose revisions propagate to dependent skills. The test-gated acceptance loop is practical: candidate pass rates rise from a 52.5% mean to 99.1–100% within five rounds, at $0.05–0.21 per run for lower-cost models (Tables 1, 5–7). Appendix B.1 is also candid about the compiler’s scope.

  2. Inspectable knowledge and unusually transparent reporting. Learned laws are explicit and can be checked by intervention (Table 3), which makes them easier to audit than opaque reward programs. The appendix reports the full climb trace, per-operator results, token costs, and wall-clock time (Tables 6–9). Releasing the code made the concerns below possible to identify.

Critical weaknesses
  1. The experiments could isolate the proposed pipeline more directly. The paper does not compare deterministic compilation with LLM-derived specifications from the same laws, although this is the most direct test of its central claim. It is also unclear how much of the Craftax-Classic run is end-to-end. For example, Table 2’s diamond result comes from scan flag collect diamond rather than the compiled mine diamond (Appendix D). Clarifying which skills were compiled from LLM-authored laws would help.

  2. The method relies on stronger assumptions than the framing suggests. In Craftax, the tests are built from the source code and verified by setting up states in the simulator. The law author also appears to work from the source, which makes RQ1 closer to source-to-BC+ translation than to learning world laws. Compilation also assumes a given mapping from fluents to environment predicates. It is unclear how these components would work in a truly unfamiliar environment, and RQ3 does not use the test-driven loop. A discussion of these assumptions would clarify the scope.

  3. The choice of BC+ could be better motivated. The laws in the experiments look mostly like preconditions with additive effects, which PDDL with numeric fluents and derived predicates could also express. Without an example that requires BC+’s extra expressiveness, its benefit over PDDL is unclear, while ASP solving adds costs (bounded quantities and horizons).

Questions for authors
  1. Could the authors clarify which BC+ theory was used for RQ2: the LLM-generated theory accepted in RQ1, or the hand-written knowledge/diamond_chain.py in the supplementary code? I ask because that file notes that it is meant to match a hand-written contract and that it intentionally leaves out resource consumption, which seems slightly different from the description of consumption C in Appendix B.1. A short clarification would be very helpful.

  2. If I read the code correctly, the law-authoring LLM in RQ1 has access to the Craftax C++ source (classic.cc), and the tests are also derived from this source. If so, it would help readers to mention this in the main text. It would also be useful to hear the authors’ view on whether RQ1 is best understood as translating source code into BC+, or as understanding world laws more generally.

  3. Could the authors share how many training seeds were used for Tables 2 and 10, and if possible include variance or confidence intervals? I also had a small question about Table 9: in round 1, several skills restarted from random weights but reached exactly their earlier success rates (e.g., 98.83% → 98.83%). Were these skills re-evaluated after retraining, or are the earlier values carried over?

  4. Since the training schedule builds on SCALAR’s frontier checkpointing (Appendix B.2), I think readers would appreciate a brief qualitative comparison with SCALAR, or, if feasible, a reimplementation of its operator-learning component.

  5. Appendix D mentions that the diamond result in Table 2 comes from scan flag collect diamond (45.70%), while the BC+-compiled mine diamond reaches a very similar 44.92%. Could the authors clarify how scan flag collect diamond was specified, and whether it was compiled from the accepted BC+ theory or taken from a predefined operator table? Because the two numbers are so close, reporting the compiled operator in Table 2 (or both, with their provenance) would make the paper’s main claim even clearer at little cost.

4 comments

Dear Reviewer q62x,

Thank you again for your time and constructive feedback.

As approximately two days remain in the author–reviewer discussion period, we would be most grateful if you could take a moment to review our rebuttal. In it, we have addressed your concerns through additional analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and welcome your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Thank you for the detailed rebuttal, which resolves several of my questions. My concerns has been clarified, but the assumptions in the actual implementation are much stronger than the paper describes, which limits the scope of the contribution. The central claim still lacks a direct comparison. I will maintain my score.

Dear Reviewer,

Thank you again for your careful consideration and constructive feedback. We sincerely appreciate your acknowledgment that our rebuttal has addressed several of your concerns.

We would be grateful if you could further clarify your remaining concerns.

First, regarding the stronger assumptions in our implementation, could you specify which assumptions you find stronger than those described in the paper, and how they limit the scope of our contribution?

We suspect that two aspects of our implementation may underlie your concerns about stronger assumptions:

  • Full source-code access.
    • We acknowledge that this is an undesirable assumption. However, RQ1 and RQ2 follow the settings of prior LLM-HRL works to ensure fair comparisons, making this assumption difficult to avoid.
    • We have also addressed this limitation in RQ3 (Mars), where ATLAS acquires world laws purely through interaction, without any prior source code, rules, or documentation.
  • Predefined fluent-to-predicate mapping.
    • We agree that automatic predicate discovery is an interesting research direction. However, scientific progress is incremental, and our focus is on decoupling world-law modeling from skill-specification derivation.
    • Previous LLM-HRL works also rely on predefined symbolic predicates. We therefore believe requiring automatic predicate discovery in this work would be overly demanding, as it introduces an additional research problem beyond our central contribution.

We would appreciate knowing whether these are indeed the assumptions underlying your remaining concerns, or whether you have other specific concerns in mind.

Second, regarding the missing direct comparison, we would like to respectfully highlight three points:

  1. The underlying problem is well established in prior literature. Text2World reports F1 scores below 0.60 for preconditions and effects, while LLM+AL finds that only approximately 20% of one-shot generations pass syntax checks.
  2. It is a well-stated common challange in this area. CODE-SHARP and SCALAR also explicitly identify LLM-generated misspecifications and missing prerequisites as important challenges.
  3. We have added direct comparisons in our rebuttal. Using CODE-SHARP's published prompts, we found that 12 out of 20 generated specifications omit necessary conditions, even for a simple skill. More importantly, we also conducted an end-to-end ablation that holds the surrounding system fixed while changing the source of specifications, directly following your suggestion.
  4. Our method demonstrates substantial performance improvements over prior approaches, providing further empirical evidence of the practical value of our world-law-driven design.

Given these additional experiments and the existing evidence from prior work, we would sincerely appreciate understanding why you still consider the direct comparison insufficient, which specific limitations remain, or which aspects of our presentation require further clarification.

We sincerely hope that you will take these additional evidence into account in your overall assessment. Any further suggestions would be greatly appreciated and would help us improve the manuscript.

Thank you for your time and consideration.

Dear reviewer q62x,

We fully acknowledge that the paper still has room for improvement. Throughout the rebuttal, we have made every effort to address the questions and suggestions you raised, including conducting the requested additional experiments and providing further clarifications.

At the same time, we hope you might reconsider the overall assessment in light of the positive aspects you highlighted in your review, particularly:

  • Motivation for decoupling:

    • Your assessment: "Clean decomposition with a working symbolic toolchain."
    • Reviewer 97xq similarly considered the separation between world modeling and skill specification "well motivated," while reviewer t3p5 found the idea of maintaining a shared environmental dynamics representation "interesting."
  • Reusable world laws:

    • Your assessment: "The paper separates LLM law authoring from deterministic specification derivation, using a shared BC+ layer whose revisions propagate to dependent skills." You also noted that learned laws are explicit and can be checked by intervention, making them easier to audit.
    • Reviewer t3p5 noted that this representation could be "useful for maintaining consistency across skills." Reviewer 97xq also recognized that retaining supporting laws gives the revision process "a useful structure."
  • Test-driven refinement:

    • Your assessment: "The test-gated acceptance loop is practical."
    • Reviewer 97xq acknowledged that our experiments demonstrate "improved agreement with a fixed behavioral test suite." Our additional results further show that the mean test pass rate increases from 52.52% to 99.62% across four models.
  • End-to-end skill composition and performance improvement:

    • Your assessment: "The system achieved a 45.7% success rate for mining diamonds, significantly outperforming the 16.0% reported by CODE-SHARP."
    • Reviewer t3p5 similarly considered the improvement "substantial" and recognized its "promising empirical performance on a challenging long-horizon task."

These are the central technical contributions of ATLAS, and we sincerely appreciate your recognition of them.

We are very glad to hear that our rebuttal has helped clarify your concerns and resolve several of your questions.

Taking a step back and considering the paper as a whole, particularly its technical contributions and the additional evidence provided during the rebuttal, we sincerely hope you might reconsider whether a Weak Accept would better reflect your overall assessment of the paper, even with some remaining issues and concerns.

Thank you again for your thoughtful feedback and the time and effort you have devoted to our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

Dear Reviewers, ACs:

We thank Reviewers q62x, t3p5, and 97xq for their careful reading and constructive feedback on ATLAS.

Key contributions and results

  1. Motivation of decoupling. Motivated by prior findings that LLMs often produce incomplete skill specifications when generating them directly, we identify the coupling between environment-dynamics modeling and skill-specification construction as a key source of this difficulty.

    • 97xq: "The separation between world modeling and skill specification is well motivated... This provides a coherent way to reuse environmental knowledge across skills and identify which specifications need updating when a law changes."
    • q62x: "Clean decomposition with a working symbolic toolchain."
    • t3p5: "... The idea of world laws, maintaining a shared environment dynamics representation is interesting ..."
  2. Reusable world laws. ATLAS addresses this by introducing an explicit representation of reusable world laws, which can be efficiently and deterministically propagated to multiple skill specifications, avoiding repeated and potentially inconsistent encoding of the same environmental regularities across individual skills.

    • t3p5: "The idea of world laws, maintaining a shared environment dynamics representation is interesting and may be useful for maintaining consistency across skills."
    • 97xq: "In particular, retaining the supporting laws for each specification gives the proposed revision process a useful structure."
    • q62x: "Critical Strengths1: The paper separates LLM law authoring from deterministic specification derivation, using a shared BC+ layer whose revisions propagate to dependent skills."
  3. Test-driven refinement. Our test-driven refinement raises the mean pass rate of LLM-generated world laws from 52.52% to 99.62% across four models after five revision rounds. For the three lower-cost models, mean API costs range from US0.05 to US0.21 per run.

    • q62x: "The test-gated acceptance loop is practical" and "Learned laws are explicit and can be checked by intervention (Table 3), which makes them easier to audit than opaque reward programs."
    • 97xq: "Experiments demonstrate improved agreement with a fixed behavioral test suite"
  4. End-to-end skill composition and Substantial Performance gain. This design achieves a 42.65% mean end-to-end success rate on the challenging diamond-collection task, compared with the 16.0% reported by CODE-SHARP (NeurIPS 2026). The updated ATLAS result is based on five independent training seeds (sample SD: 3.28 percentage points; 800 episodes per seed).

    • t3p5: "The improvement over baselines in the Craftax-classic diamond collection task seems substantial. This demonstrates promising empirical performance on a challenging long-horizon task.”
    • q62x: "the system achieved a 45.7% success rate for mining diamonds, significantly outperforming the 16.0% reported by CODE-SHARP"

Responses and revisions

All three reviewers highlighted strengths of ATLAS's design, and q62x and t3p5 noted the substantial improvement on diamond collection. The reviews also raised questions about the experimental evidence and how the method was presented. During rebuttal, we ran additional experiments, revised the manuscript, and responded point by point. Three concerns were central:

  1. Direct comparison with prior specification generation. We agree that the contribution of deterministic compilation is best assessed by comparing it directly with LLM-generated specifications. In the added comparison, 12 of 20 specifications generated under CODE-SHARP-style prompts omitted at least one required condition. For example, direct generation omitted the nearby crafting-table requirement for a wooden pickaxe; ATLAS's compiler preserved that prerequisite from the accepted world laws. This illustrates the advantage of compilation in preserving requirements encoded in the laws. (Responses to 97xq W1 and q62x W1(a); revised manuscript §6.3 and Appendix F.2.)
  2. Evaluation protocol and variability across training runs. We now report five independent training runs and a separate final end-to-end evaluation with 800 episodes per seed. The mean diamond-collection success rate was 42.65%, with a sample standard deviation of 3.28 percentage points. The mean is close to the originally reported 45.7% and retains a 26.65-percentage-point advantage over the 16.0% reported by CODE-SHARP. (Responses to 97xq W2(a) and q62x Q3(a); revised manuscript §§5.3 and 6.2, Table 2, and Appendix F.3.)
  3. A concrete example of a world law and its compiled specification. We added an iron-pickaxe crafting example that shows how a BC+ law represents a furnace requirement and how the compiler turns it into a skill prerequisite. The appendix gives the corresponding prerequisites, effects, and resource consumption. (Response to t3p5 W2; revised manuscript §4.3 and Appendix F.4.)

We believe the added experiments and the crafting example address the three main concerns above.

The table below maps each reviewer point to our response and the corresponding manuscript revision. In the marked PDF, additions are blue, deletions are shown in red strikethrough, and purple superscripts identify reviewer comments.

Reviewer pointResponse and manuscript changeManuscript location
97xq W1; q62x W1(a)Added a direct-generation comparison under CODE-SHARP-style prompts: 12 of 20 specifications omitted a required condition.§6.3; App. F.2
97xq W2(a); q62x Q3(a)Reported results across five independent training seeds and added a separate final end-to-end evaluation for diamond collection.§§5.3, 6.2; Table 2; App. F.3 (Table 12)
97xq W2(b)Clarified the Craftax death-prevention setting and the qualitative scope of the Mars experiment.§§5.3, 6.4
t3p5 W1Quantified the reduction from 35 candidate fields to 5 and 7, yielding 32 and 128 contrast queries, respectively.§4.4
t3p5 W2Added the crafting-law example and the compiled skill specifications.§4.3; App. F.4 (Table 14)
q62x W1(b), Q5Clarified operator provenance and reported results for three diamond channels across five training seeds.§6.2; App. F.3 (Table 13)
q62x W2, Q2Specified what source materials the law and test authors use and documented the interface-provided predicate mapping.§5.1; App. F.6
q62x W3Reported symbolic query timings for testing and planning.App. F.5 (Table 15)
q62x Q1Clarified that the RQ2 operator specifications are compiled from the theory generated and accepted in RQ1.§6.2
q62x Q3(b)Documented checkpoint selection in the climb trace and added per-skill cross-run means and standard deviations.App. D (Table 9)
q62x Q4Clarified how SCALAR differs in observation inputs and policy architecture.§5.2; App. B.3

We thank all three reviewers again for the time and care they devoted to the paper and for helping us improve it. Their requests for a direct comparison with LLM-generated specifications and a fuller account of our evaluation protocol helped us substantially strengthen the empirical evidence for ATLAS. The suggestions to quantify the reduction in counterfactual search and include a concrete crafting example also made the method easier to assess.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.