acceptodds
Under review as a conference paper at ICLR 2027

Minimal Complete Specification Languages via Specification Equivalence Backups

Abstract

As AI systems become increasingly able to solve any task we can specify, the bottleneck becomes being able to more efficiently and robustly specify our tasks. In this work, we move towards the theoretical limit of optimal specification: finding the minimum specification language which allows for precise specification of any task in a domain. We formalize the Specification Language Design Problem, and define the minimum amount of information necessary to determine an optimal policy for the specification, which we call the Specification Equivalence Set. This set has the recursive substructure necessary for a Bellman backup to compute it exactly, allowing us to find optimal specification language efficiently and exactly for a set of specification problems. We then extend these insights to a deep RL algorithm, using a VQVAE-inspired method to train an agent that is controllable via a latent specification language. We find that this deep RL algorithm matches our exact algorithm at small scales, and allows the system to scale to larger domains. In introducing this problem and our methods of addressing it, we aim to provide a primary look into the benefits of more precise specifications for AI systems.

Primary area:
reinforcement learning
Keywords:
Specification, Planning, AI Safety, State Representations, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_e8yDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to minimize the information needed to specify a task while still obtaining an optimal policy. It defines Minimal Complete Specification Languages (MCSLs) using sets of tasks that share an optimal policy. The paper proposes SEB, an exact algorithm based on Bellman-style backups, and ALSES, a deep RL approach using discrete codes to approximate the specification language. Both methods are evaluated in gridworld environments.

Critical strengths

The problem of identifying which task details are necessary for optimal behavior is interesting and potentially useful. The idea of grouping tasks that share an optimal policy provides a simple and intuitive way to formalize this problem. It also offers an interesting perspective on task specification by focusing on the information needed for optimal behavior rather than fully describing the task.

Critical weaknesses
  • The overall contribution is not sufficiently clear. SEB seems closely related to the related works, while ALSES takes a rather different approach without clearly building on SEB. The connection between the theoretical and empirical parts feels weak, making it difficult to identify the main technical contribution. Also, minimizing the number of specifications does not necessarily mean reducing human effort, which is the main motivation of the paper.

  • I have concerns about the theoretical assumptions and the empirical evidence. Some statements and examples seem inconsistent, and the exact scope of SEB's correctness is unclear. ALSES also shows unstable performance and often fails to recover the optimal partition. The experiments are limited to gridworlds and do not provide sufficient comparisons to establish the benefits of the proposed approach.

Questions for authors
  1. Could the authors clarify the main technical differences between SEB and the related works? In particular, what is new in the backup procedure beyond applying existing equivalence-based compression ideas to task specification? Also, ALSES seems to use a different approach based on discrete latent codes. Is there a specific property of SEB that ALSES is designed to approximate or preserve?

  2. The UMDP definition allows transitions to depend on the level, but Algorithm 1 uses a single successor state (T(s,a)). Does SEB assume shared transitions across levels? For example, if the same action leads to different states in different levels, a single state-dependent policy could still be optimal in both. How would the backup handle this case, and does the correctness proof still apply?

  3. Theorem 1 seems to treat partitions and covers as equivalent, although maximal SESs can overlap. Would it be more precise to state that their minimum cardinalities are equal, rather than that the objects themselves are equivalent? Also, in Section 5.1, the final minimum cover seems inconsistent with the previously listed maximal SESs and their policy labels. Could the authors clarify these points?

  4. ALSES minimizes a combination of expected return and code-use entropy, which seems different from minimizing the number of specifications while preserving optimality for every task. What justifies this surrogate objective? In Figure 7, even the best run achieves optimal behavior in only 1,512 out of 2,048 worlds. Could the authors report per-world optimality gaps alongside code counts to better show this trade-off?

  5. ALSES recovers the correct partition in only 4/8 seeds in the noisy maze and 2/8 seeds in the colored key maze. What is the main source of this instability? It would also be useful to compare against simpler policy-based clustering or discrete representation learning baselines to better understand whether the proposed training method offers a clear advantage.

2 comments

We firstly wanted to say thank you for the thorough review and helping to strengthen this paper through your feedback. To address your review:

  • “The overall contribution is not sufficiently clear”. We have added a new paragraph in the paper clarifying the contribution (with more details in the general response). We hope to make this clear for the reader and to be more distinct about what this work aims to introduce. We also address the argument we hold for minimization of specifications and its relation to human effort and the HCI angle we come from in making such claims. We hope in clearing up these weaknesses we can better demonstrate the strength of our paper and the foundations of the problem it is trying to resolve.
  • “ALSES also shows unstable performance and often fails to recover the optimal partition.” We hope that the new results and performance of ALSES help to solidify it as a strong method for solving such problems, and agree that while the experiments are limited to toy examples, the aim of this work is to introduce the problem and its foundation while supplying the first methods for solving such problems.

Answering your questions about the work:

  • The general response above alongside the new paper section aims to make the distinction and contribution of this work clear. The ALSES objective gives a differentiable lower bound to the SEB objective, and as shown in practice, can often resolve to the same codebooks. This should also answer part of your question 4.
  • Yes, all the proofs hold and this is a notational error, this has been fixed in the new version of the pdf, we appreciate the comment as it allowed us to correct this error.
  • And here again, in section 5.1, this was another notation error. This has been fixed in the latest pdf version.
  • The new version of the draft goes into more details in the second paragraph of 7.2, 7.4, and the second paragraph of 7.5 about where the performance gaps are from and which levels this applies to. We thank you for the feedback and have incorporated it into the main text. Additionally, the overall performance however has significantly shrunk in this version with the new results.
  • The instability has been greatly resolved with the new results, now stemming from the necessary random starts as this introduces a need for information to survive even for information outside the policies which start from the starting states. Our response to reviewer TWEs (the third bullet point) goes into further detail about the cases in which SEB itself would be far too expensive to run and only a sampling based method such as ALSES can efficiently answer.

We hope that with these new results and clarifications we have addressed your main concerns. We would be grateful if you would consider updating your scores, and if anything remains unclear or any concerns remain, please let us know before the discussion period closes.

Thank you to the authors for the detailed response, which addresses some of my concerns. However, I still have reservations about the technical novelty and theoretical soundness of the proposed approach.

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

This paper introduces a way to help a human form the smallest specification to execute an optimal policy in a world. The authors compute common specifications for multiple worlds, by constructing the smallest required specification such that the same policy is optimal for multiple worlds. A full specification might require a human to define all labels, but using the specification equivalence backup, the authors identify the minimal set of labels required to define which optimal policy is to be executed given scenario. This reduces the possible choices required to be made by the designer, while ensuring the language is still complete. They introduce SEB, which provides exact solutions in a small discrete environment with limited labels and ALSES, which converts environments/worlds into discrete codes using VQ-VAE's and training a policy to identify optimal actions given worlds. This allows the codebook to combine multiple worlds in a single code, hoping to recover the same minimal specification as SEB.

Critical strengths

This is an extremely interesting paper to read. The problem focuses not on creating better specifications, or reward functions but rather identifying what is the minimum required specification to identify optimal policies to execute given multiple worlds sharing similar configurations. To the best of my knowledge, this is the first approach to such a problem and opens up interesting avenues for future research.

  1. The SEB algorithm is sound and can identify exact solutions for smaller worlds. It is an NP-Hard problem that requires approximate solvers like ALSES to solve.
  2. The ALSES algorithm is quite intuitive and a solid solver for the SEB problem. Along with finding the minimum set cover, it also learns the policy. This opens up very nice avenues to use this to solve problems that have near similar MDPs but solve different tasks. Even if it does not share an MCSL, it will allow optimal reuse of world knowledge. This is like multi-task RL with discrete task embedding and a compression objective.
Critical weaknesses

As with any foundational paper, there are some weaknesses, but they motivate future research than harm this paper. However, it is necessary to mention this in the paper.

  1. The SEB algorithm assumes that optimal policies are available and the transition kernel is independent of the worlds. While it is critical to the exactness of the solution, SEB is limited to worlds having the same transition kernel. However, it is not impossible to imagine that this could be the scenario for a lot of environments, like crafter, mario etc.

  2. The ALSES algorithm suffers from codebook collapse. Additionally, the code-usage entropy penalty could prefer collapsing every world into a discrete code. This would minimise complexity, but make learning an optimal policy impossible. The phased training somewhat mitigates this, but it is helpful to mention. I hope future research could solve this problem/mitigate it.

  3. Merging codes can delete optimal policies, which could happen in phase 2 compression. The experiments show this exact failure. Additionally, these codes are non-interpretable, making it difficult for any human designer to define specifications over the recovered language. I would hope that future extensions recover symbolic codes rather than learned codes, making it more interpretable.

Questions for authors

What real world scenarios do you expect this research to help?

2 comments

We want to thank you for the kind words and the interest in the foundation of this problem. Following your suggestions, we aim to carefully present our results so as to lay a strong foundation for the field:

  • “The SEB algorithm assumes that optimal policies are available and the transition kernel is independent of the worlds. While it is critical to the exactness of the solution, SEB is limited to worlds having the same transition kernel.” In principle this is always factorable to the parameters. When they significantly differ however, areas of environment space become "which game are you playing". With a wide enough distribution one could expect that these clusters would describe game mechanics, and could be found in the latent space of a Genie model, but this stays far beyond our current scale, and has a higher likelihood to be uninterpretable.
  • “The ALSES algorithm suffers from codebook collapse.” Our new results have overall reduced this problem almost entirely, given the objective of the model this problem is still not strongly resolved but we now demonstrate that often the model will now overstate the number of specifications as it focuses on distinctions unnecessary for policies beginning from the starting state. As you mention, this is a strong avenue for future work, and lay it out so in the paper, hoping to spark advancements on this problem.
  • “Additionally, these codes are non-interpretable, making it difficult for any human designer to define specifications over the recovered language” We fully agree that interpretable codes is a natural future work and that the current methods don't guarantee interpretability. However, we find experimentally that these codes do often end up being somewhat understandable, as shown with the “is lava there or not?” question from the four mazes or the “is lava blocking the only path?” from the lava maze.

Finally, to your question, while this work is extremely nascent, we believe it has the possibility to be impactful as we transition into a world with systems which can solve problems behind current human understanding, by identifying the aspects of the problem which are important to specify. For example, perhaps this could be used to help determine the parameters combinations that are necessary to know for a fusion reactor simulation to allow us to know which policy to deploy.

We hope that with these new results and clarifications we have addressed your main concerns. We would be grateful if you would consider updating your scores, and if anything remains unclear or any concerns remain, please let us know before the discussion period closes.

I have no further comments

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

The paper asks how little information a designer must give an agent so that the agent acts optimally. It models tasks as levels θ of an underspecified MDP. A "specification language" is a finite set Σ and a map Λ: Σ → Π. Theorem 1 shows that a minimal complete language is a minimum cover of Θ by maximal Specification Equivalence Sets, which are sets of levels that share an optimal policy. SEB computes these sets by backward induction. ALSES learns an approximation with a VQ-style encoder and a code-conditioned policy. All experiments use small gridworlds.

Critical strengths
  • The question is interesting. It is useful to ask how much information an agent needs to act optimally when the designer cannot solve the task.
  • The proofs of Theorem 1 and Proposition 2 appear correct.
  • The paper reports mixed ALSES results openly.
Critical weaknesses
  1. The specification framing hides the contribution. In formal methods, a specification is a property of behaviors with a satisfaction relation, as in LTL or PCTL. I first read the paper with this meaning, and it does not apply. Here σ is an opaque codeword, and Λ is a lookup table. σ restricts the environment parameter θ, not the behavior of the agent. The paper does not define what an element of Σ is.
  2. Under a communication framing, the novelty is limited. The problem becomes one-shot, zero-error communication. An informed sender knows θ, and the receiver knows the model. This is a special case of the minimal-knowledge setting of Lauffer et al. (2023), with θ in place of the co-policy. Theorem 1 is close to a restatement of the definitions. Please state the reduction to Lauffer et al. and explain what is new beyond it.
  3. The efficiency claims have no complexity support. SEB enumerates Θ, and its final step is minimum set cover. Any set system is an SES family: use a one-step UMDP with one action per set. So the problem is NP-hard, and no polynomial algorithm approximates it better than ln|Θ| (Dinur and Steurer, 2014). The paper calls SEB "efficient" but gives no complexity analysis.
  4. The experiments do not measure communication. The theory minimizes |Σ|, which is a worst-case measure. ALSES penalizes code-usage entropy, which is an average-case measure. These objectives can disagree, and the paper does not discuss this. No experiment tests generalization to unseen levels, which is the test that would show real compression. The paper reports no rate and no capacity measure. It also gives no baselines, such as clustering the optimal trajectories of sampled levels or greedy set cover. Exact recovery succeeds in 2 to 4 of 8 seeds. The lava maze shows only the best run.
  5. The notation conflicts. T denotes the transition function, the horizon, and the true task. Algorithm 1 uses H for the horizon. The text defines U(π*) as an argmax. T depends on θ, but SEB and the proof write T(s, a). The method is correct only if the dynamics do not depend on θ.
Questions for authors
  • What is an element of Σ? How does a human sender choose σ without knowing which set contains θ?
  • Can you frame the work as minimal communication for optimal control? Please relate it to contextual and multi-model MDPs, parameter-space partitioning for parametric MDPs, and UED (Dennis et al., 2020).
  • Can you add a prior over Θ and measure the rate in expected bits? A rate-regret trade-off with ε-optimal policies would make the problem less arbitrary.
  • Can you evaluate ALSES on held-out levels and compare it with simple baselines?
2 comments

We would first like to thank you for the incredibly thorough and insightful review especially as it pertains to pushing its theoretical boundaries. We agree with the critiques though we believe we still have a strong paper regardless, importantly however, from your review we understand the importance of conveying to the reader these points to make sure the results are not read over.

  • “The specification framing hides the contribution.” We thank the reviewer for pointing out the collision in terms between research communities that would have caused confusion in many readers. Our work follows how “specification” has been used in “reward specification” or “task specification” in the RL community in works such as IRD (Hadfield-Menell et al., 2020), our “environment specification” is a generalization of these since reward and task parameters can be included as part of the environment parameters. While this is definitely not as direct of a specification of behaviour as in formal methods, we would hope that the subsets defined Λ would tend to have some semantic interpretation about the environment. To the extent they do not have natural semantics (as noticed by reviewers qspY, e8yD, and WNDM) that are easy to learn, we hope future work studies how we can improve this and will direct readers towards work pushing in that direction. We will ensure in the intro that our use of this term is clarified.
  • “This is a special case of the minimal-knowledge setting of Lauffer et al. (2023), with θ in place of the co-policy.” We hope that with the changes to the paper (more detail about this are in the general comment above) the distinction is now clear where this work stands with its contribution. In addition to the fundamental algorithmic differences, we also found that from communication with experts in the area, this connection is often missed or unintuitive and hope this work helps bridge that gap.
  • “No experiment tests generalization to unseen levels”, This paper focused on the case of making a gradient based method to approximate the SES exact calculation where the whole problem is already known and generalization isn’t as essentially, leaving it to future work to study when we can get the full ML model to generalise. Our exact algorithm is the baseline, but is computationally infeasible and simply doesn’t finish in our time budget in the large-scale noise test since the exponential aspect makes the problem blow up for the exact algorithm. Extrapolating the runtime and memory cost, given the 32bit noise, the exact SEB algorithm would have a runtime of around 3.5hours (compared to 4 mins with ALSES), and a memory cost of 100TB to hold all 2^32 * 4 worlds in our implementation. Lastly, we hope new performance of the ALSES experiment in this version of the pdf, including experimenting on 8 seeds rather than just the best run, helps demonstrate the promise of such an algorithm in necessarily scaleable problems.

Finally, we thank you for pointing out the notation conflicts and have tried to fix all such issues.

As for the responses to your questions:

  1. “What is an element of Σ? How does a human sender choose σ without knowing which set contains θ?“ This is an arbitrary set. We assume that the human has to learn to use the language first or that it has to be explained to the human. In the spirit of considering the theoretical limit we chose this to be as flexible as possible 2.”Can you frame the work as minimal communication for optimal control? Please relate it to contextual and multi-model MDPs, parameter-space partitioning for parametric MDPs, and UED (Dennis et al., 2020).“ Yes the framing fits, however our focus on this paper was the designer’s choice set rather than the communication channel itself. This work sits alongside parameter-space partitioning and contextual/multi-model MDPs and complementary to UED, in that UED works to select which level to train on while this work specifies such levels.
  2. “Can you add a prior over Θ and measure the rate in expected bits? A rate-regret trade-off with ε-optimal policies would make the problem less arbitrary.” While this is a reasonable experiment to run, this introduces the question of a prior over the tasks which itself needs correct specification. In this work we focused on minimizing the codes in order to give a prior-free answer to the problem, though future work focusing on this would be highly interesting and beneficial.
  3. “Can you evaluate ALSES on held-out levels and compare it with simple baselines?” Using the updated ALSES results, drawing 10_000 noise examples for our noise experiments to create 40_000 levels (4 settings x 10_000 noise) for each setting, noise affecting the reward and not, the trained ALSES networks from the paper match the correct policy 100% of the time across all seeds. Although some seeds do have extra unnecessary codes (this is expanded on in the paper) those extra codes are merely redundant and produce the same policies from the starting state.

We hope that with these new results and clarifications we have addressed your main concerns. We would be grateful if you would consider updating your scores, and if anything remains unclear or any concerns remain, please let us know before the discussion period closes.

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

The authors asks a simple question: how much information a person need to give AI to make it optimally perform a task? The key idea is: different tasks can share the same instruction if there is a single policy that is universally optimal. The paper bring forward an algorithm that identifies tasks that can share a policy, looks for the fewest instructions needed to cover all tasks. They also develop a nn method that learns to compress task descriptions into discrete labels for guidance. Maze experiments show that this approach can remove information, which doesnt affect optimal behavior. Yet the neural method is quite unstable and fails to preserve optimal behavior for all tasks in the larger experiment.

Critical strengths
  1. The paper asks a interesting and quite relevant question: what is the smallest set of instructions that lets an agent act optimally across all tasks in a given domain? I like that the authors separates completeness from minimality. This produces a useful framework for studying which information user needs to provide.

  2. I learned something useful. The maze exps show how various environmental details can be omitted and not changing policy optimality, and how the required information changes when the optimization objective change. The exact solutions also provide fairly useful reference for checking the learned task labels achieve compression and optimal behavior.

Critical weaknesses

This paper is difficult for me to evlauate, as I do not have substantial direct research exp on the RL methods that this paper engage with. Nevertheless, here are some weaknesses I think are relevant:

  1. The ALSES objective does not enforce completeness or minimality. The objective likely trades optimality for compression rather than guaranteeing both. The experiments show the failure mode. In Section 7.2, 25% of seeds collapse to a single “always take the bottom path” code, and in Figure 6, 3/8 seeds end with one code.

  2. In a intuitive sense, Fewer labels doesnt mean less effort for users. The experiments give the encoder full task info and assume users know which label to select. Yet the paper doestn explore whether people can understand the learned labels or choose the correct one.

  3. (minor) It is encouraged that the authors remove the reproducibility and impact statement templates are write their own. Currently, I find the paper's reproducibility details unsatisfying.

Questions for authors
  1. Could the authors release the SEB and ALSES implementations together with the maze definition, report the full ALSES training config for each experimnt?

  2. For the 2048-world lava maze, how many maximal sets did F(s_0, 0) contain, how was the minimum cover computed, and what was the runtime?

2 comments

We aim to respond to the critiques brought up, and answer the questions posed to the authors.

“The ALSES objective does not enforce completeness or minimality. The objective likely trades optimality for compression rather than guaranteeing both.” We agree the ML approximation does not enforce these conditions, and agree this should be made clear for the reader. We state these properties to follow the tradition in ML of separating our formal objective from our deep learning approximation of it. We believe this is good practice as it will make it clear when a future approach is better than ALSES at achieving this objective. While we do not intend to claim that ALSES guarantees this objective, we believe the results discussed in our top-level response show it is a relatively effective approximation in our domain. “In a intuitive sense, Fewer labels doesnt mean less effort for users”. Thank you for this important point. We agree this needed further clarification for the reader and have made changes to the paper to help illustrate this point which we detailed in the general review above. With this paper our primary aim was to tackle one aspect of this problem at a time and had a relatively narrow scope, but agree that future work in this area which addresses usability would be highly beneficial. This is a multi-faceted problem with many interacting components, and ultimately all of these facets will be relevant. Finally, we appreciate the reproducibility section weakness and have updated it to reflect the access to the open source code (which we aim to release as soon as possible) and the ability to reproduce every experiment deterministically.

Responses to your questions follow:

  1. We are planning on releasing the code for all experiments, and have since added an appendix section outlining every hyperparameter used for every experiment in the paper.
  2. In the lava maze experiment, and all of our experiments for this paper, the number of maximal SESs is the same as the resulting MCSL, is not a general case, and the minimum set cover operation is still required. However, given the scale of our experiments, the minimum set cover is an exhaustive search rather than the greedy approximation. In total, for SEB we have a cpu version which ran the lava maze experiment in 2s, while a GPU implementation runs in 0.07s.

We hope that with these new results and clarifications we have addressed your main concerns. We would be grateful if you would consider updating your scores, and if anything remains unclear or any concerns remain, please let us know before the discussion period closes.

Thanks for the response. My concerns are addressed. Thus, I maintain my recommendation of weak acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We first would like to thank the reviewers for their thoughtful analyses of this work, and are grateful for helping strengthen the paper.

In this general response we focus on the 3 main points found across reviews:

A common critique was the robustness of the ALSES algorithm, noted by reviewers qspY, WNDM, and e8yD.To address the these concerns we have revisited the experimental setup and found significant improvements in stability and algorithmic performance through careful application of common deep RL tricks, and hyper-parameter adjustment:

  • In the four maze experiment with and without noise, ALSES now recovers the exact codes 6/8 seeds in both set ups, up from 4 and 3 respectively, with the other two worlds having perfect performance but an extra redundant code.
  • In the Lava Maze experiment, ALSES now recovers a performance of 76,507.4 out of a possible 76,851.5, a fallback of only 344.1 (0.45%) up from 69,189. This is now the only experiment where ALSES doesn’t recover optimal performance. Given only the behavior found ALSES did manage to optimally partition those policies optimally in 6/8 seeds.
  • In the Mixed Hazard Maze experiment, the network recovers perfect performance of 567 in all 8 seeds, up from a mean of around 240, while having more seeds recovering the same codes as SEB (2 -> 3). In brief, these changes include random starts with no history from states in the UMDP with a slow anneal off of random starts, annealing embeddings to codes, and further hyperparameter tuning. More details are included in the paper, along with every hyperparameter used for each experiment.

Reviewers qspY, e8yD, and WNDM point out that a minimum number of specifications doesn’t necessarily translate to minimum human effort, this is true: nothing guarantees the codes will be interpretable or easy to use. While sometimes codes are interpretable like the “is there lava?” question in the simple four mazes case, or “is there certainly lava blocking the path?” as the complexity grew in our experiments, we do not want readers to assume this is inevitably the case. In order to make this limitation clear to the readers, we added the following line in the definition of the specification language design problem section: “Importantly, the semantics of as implied by need not be intuitive nor interpretable and would need to be learned by human users. In repeated tasks where the costs of teaching the language can be amortized, |\Sigma| characterizes the amount of human input necessary. This also sets aside the cognitive effort the human must exert to use the language.” We will also point the reader towards related work in this direction such as the HCI literature, where works like (Reddy et al., 2021) and (Reddy et al., 2022) design interfaces to be more human usable. We furthered this point in the future work section by adding interpretable codes as a future direction.

Reviewers TWEs, WNDM, and e8yD asked that we clarify the novelty of the algorithm relative to VSER. Conceptually, while this work builds on the VSER algorithm from Lauffer et al. (2023), VSER is limited to cooperative games due to its hidden determinism assumption. It also results in a much larger object than necessary if we only require a single best policy per world. In this regard, the algorithm we introduce in this section aims to reduce the amount of information in the backup by intersecting policies and keeping only the minimum number of policies needed. This then makes the resulting object of the complete languages a cover rather than a partition in VSER. We have added this passage in section 5 while expanding further on our contribution by introducing the distributional task case which necessitates further algorithmic development in section 5.3.

We thank the reviewers for strengthening this paper and highlighting important necessary clarifications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.