acceptodds
Under review as a conference paper at ICLR 2027

The Geometry of Empowerment

Abstract

Empowerment captures the capacity for an agent to actively control its environment. While conceptually appealing as an information-theoretic quantity, the connection between empowerment and MDP centrality has remained an open question. In this work, we link empowerment maximization and skill-learning methods to provide new geometries for interpreting and analyzing empowerment. Our analyses answer longstanding open questions on the connections between empowerment and centrality. Our analyses also reveal distinctions between information and reward geometries, highlighting important theoretical implications to build scalable empowerment-maximization methods.

Primary area:
reinforcement learning
Keywords:
reinforcement learning, empowerment, theory, intrinsic, adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wH4zOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper studies the relationship between empowerment, state centrality, and downstream adaptation. It provides geometric and hitting-time interpretations of empowerment and establishes an adaptation lower bound in tabular settings under an isotropic Gaussian reward prior. Counterexamples show that higher empowerment does not necessarily imply better adaptation, highlighting limitations in continuous state spaces and under anisotropic rewards.

Critical strengths
  1. The paper addresses the fundamental question of whether empowerment benefits downstream tasks and has a clear research motivation.
  2. The paper is generally well written. The geometric interpretations and visualizations in small-scale environments help explain the main results.
Critical weaknesses
  1. The centrality result offers limited theoretical novelty. The authors describe Theorem 4.1 as a straightforward extension of Eysenbach et al. (2021), but do not sufficiently explain what new insights its centrality interpretation provides beyond existing information-geometric results.

  2. The adaptation lower bound offers limited practical guidance. The authors explicitly acknowledge that the bound does not guarantee consistent rankings between empowerment and adaptation returns. Thus, although the result establishes a conditional theoretical connection, it provides little direct guidance on selecting states or skills to improve downstream adaptation.

  3. The empirical evaluation is insufficient. Experiments are limited to small-scale tabular demonstrations, without systematic evaluation of practical skill learning and downstream adaptation. The mismatches in continuous state spaces and under anisotropic rewards also lack empirical validation with learning algorithms, leaving their impact on practical methods unclear.

  4. The claims about scalable methods are too strong. The authors argue that scalable empowerment methods must address the mismatch between information geometry and reward geometry. However, practical methods are not inherently required to align these geometries. Without showing that the mismatch limits downstream performance, the paper does not establish why resolving it is necessary.

I would be willing to raise my score if the authors adequately clarify the main concerns above in their response.

Questions for authors
  1. Beyond applying the existing information-geometric result to the reachable occupancy distributions conditioned on each initial state, what new theoretical insight does Theorem 4.1 provide?
  2. Beyond the ranking counterexample in Figure 5, are there explicit additional conditions under which comparisons of empowerment can guide the selection of initial states or skill sets? If not, what do the authors consider the main practical implication of Theorem 4.3?
  3. Could the authors provide empirical evidence beyond small-scale tabular demonstrations by evaluating concrete empowerment-based or mutual-information-based skill-learning methods, examining whether the mismatch occurs in continuous state spaces or under anisotropic rewards and how it affects downstream adaptation?

Other

  1. Appendix I contains only a heading and no text.
  2. The readability of Figure 3(b) could be improved.
3 comments
AuthorsOpenReview ↗

Official Author Response to Reviewer wH4z

Thank you for the thoughtful feedback. We strengthen the original adaptation bound (Theorem I.1) and show a generalized version of the bound for a general class of GP priors using a noise-regularized version of empowerment (Theorem I.9). We have added experiments on adaptation under matched and anisotropic reward priors. Do you have any further concerns after these revisions?

New theoretical insight of Theorem 4.1 beyond Eysenbach et al. (2021)

We have revised the text to specifically discuss those related works [1, 2, 3, 4]. There are substantial differences between our theoretical framing and that of [1]. Eysenbach et al. [1] optimize the skill prior and analyze minimax (worst-case) reward initialization against the induced state marginal. We instead index the reachable occupancy polytope by the start state and prove an average-case zero-shot adaptation bound under a Gaussian reward prior. As such, our formulation shows skill/latent-based empowerment can be useful for selecting empowered states, not just partitioning a marginal behavior space effectively.

State/skill ranking conditions beyond Fig. 5 & practical implication of Theorem 4.3

Theorem 4.3 lower-bounds adaptation from a given state based on that state's empowerment. This does not necessarily imply a ranking over states. Fig. 5 shows why a general ranking claim would be too strong.

With correct priors and γ=0.95\gamma=0.95, the correlation between effective empowerment and return is 0.97 for center-doorway grids and 0.92 for corner-doorway grids.

Empirical validation beyond small tabular grids in continuous states & under anisotropic rewards

Our core contribution in this paper was the geometric and theoretical exploration of a concrete and practical empowerment objective based on skill learning. We have revised the text to clearly state the empirical limitations in §5. A key direction for future work is building and testing scalable methods that maximize this potential empowerment quantity.

We also test anisotropic rewards with equal-trace covariance (rotations) and fixed skills.

Necessity claim that scalable methods must align information and reward geometries

Revised manuscript (§5, p. 10): "More broadly, Theorems 4.4 and 4.5 refute universal MI-only guarantees rather than every practical MISL method: smooth encoders, Wasserstein constraints (Metric-aware Abstraction (METRA); Park et al., 2023), and contrastive successor features (Zheng et al., 2025) implicitly regularize against pathological state splitting."

References

Dear Reviewer wH4z,

Thank you for your thoughtful feedback. We have worked hard to incorporate your feedback with added experiments, revisions, and a stronger, generalized version of the 4.3 bound that applies to a general class of GP priors.

Have our revisions and response sufficiently resolved your questions and concerns? We appreciate your engagement as the end of the discussion phase nears. Thank you.

Thank you for the revision, which appropriately softens the claims on scalable methods and adds helpful diagnostics. However, my main concern regarding empirical depth remains, as the experiments are still limited to exact solvers and controlled toy settings without evaluating practical neural MISL algorithms. Given the improved clarity and rigor, I am willing to marginally raise my score from 2 to 3.

Reviewer_cmh1OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper provides an in-depth theoretical analysis of the concept of "empowerment" in reinforcement learning, adopting the dual perspectives of information geometry and temporal distance geometry. The authors first draw a rigorous distinction between "potential empowerment" —a measure of maximum latent control capacity—and "effective empowerment" (Eeff\mathcal{E}_{\text{eff}}), which is the quantity actually optimized by mutual information-based skill learning. Furthermore, the paper reformulates traditional open-loop action sequences into closed-loop, skill-conditioned policies based on discounted state occupancy measures. Building on this foundation, the paper achieves three key theoretical advancements.

Critical strengths
  1. This paper bridges information theory and convex geometry—specifically reachable polytopes and Gaussian widths—with unsupervised skill discovery in reinforcement learning (e.g., DIAYN, VIC, Metra). In particular, the negative results presented in Section 4.3—concerning the divergence between "information geometry" (based on KL divergence) and "reward geometry" (based on Gaussian widths or metric distances)—pinpoint exactly why intrinsic reward methods driven purely by mutual information are prone to "skill collapse" in continuous, high-dimensional spaces (as illustrated in Fig. 6).

  2. This work provides a formal resolution to the boundary conditions for the validity of the AI-empowerment hypothesis proposed by Salge et al. (2014). Furthermore, it demonstrates that employing a continuous uniform prior—as opposed to a discrete uniform prior—achieves full universality in reaching channel capacity (Lemma D.1 vs. Lemma E.1), thereby offering a robust theoretical foundation for the selection of priors in modern MISL algorithms.

Critical weaknesses
  1. In the abstract and introduction, the paper introduces a dense array of concepts, including structurally central states, skill learning, information and reward geometries, as well as potential and effective empowerment—without clearly defining the precise mappings or the logical derivations connecting them at the outset. Although the introduction employs an intuitive "mountaineer" analogy, the underlying causal and progressive relationships between "potential empowerment," "effective empowerment," and "MDP structural centrality" remain somewhat obscure, imposing a significant cognitive burden on the reader in grasping the core narrative thread of the work.

  2. The example of climbing up or down a mountain used in the introduction to distinguish between "being in control" and "actively exerting control" is ambiguity; readers are left to puzzle over whether the former implies a state of stasis without a specific objective, while the latter denotes goal-oriented execution. More importantly, the paper fails to clearly articulate how these two modes of control transition into one another and how they support the core problem the study aims to address: the identification and characterization of "central states." Although the text draws extensively upon and integrates existing literature from fields such as information theory, graph centrality, temporal distance, and unsupervised skill discovery, it lacks a cohesive logical thread to unify these disparate theoretical fragments into a coherent whole.

  3. The paper provides a profound analysis of the fundamental divergence between KL information geometry and reward-adaptive geometry (characterized by Gaussian width or metric distance)—both in continuous state spaces (Theorem 4.4 / Proposition G.2) and under anisotropic priors (Theorem 4.5 / Proposition H.2)—identifying this as a core obstacle to developing scalable empowerment-maximization methods. However, the work stops short at a purely theoretical "diagnosis" of the problem; it fails to propose any constructive objective functions or algorithmic refinements that would align these two geometric structures (for instance, by explicitly incorporating metric constraints or prior covariances into latent empowerment maximization).

  4. Both the abstract and the main text emphasize that the theoretical analysis provides a significant foundation and valuable insights for developing "scalable empowerment-maximization methods." However, the experimental validation is extremely limited, as all experiments are confined to very small-scale, tabular environments.

Questions for authors
  1. The authors should further clarify the conceptual framework in the introduction and the preliminaries: how are the concepts of "potential empowerment," "effective empowerment" (or skill learning), "information geometry," and "reward geometry" interconnected through the central link of "structurally central states"? It is recommended to include a clear conceptual relationship diagram or a comparison table to enhance the readability of the paper.

  2. In the mountain-climbing example shown in Figure 1, does "being in control" essentially correspond to a state of exploration or preparation—lacking a specific downstream objective—while "actively exerting control" (descending via various routes) corresponds to goal-directed execution tailored toward a particular target or skill branch? Please clarify: how does this two-tiered distinction in control behavior logically lead to a resolution of the tension between "MDP structural centrality" and "adaptation to downstream tasks"?

  3. Given that the paper claims its geometric analysis is intended to guide the development of "scalable empowerment-maximization methods," Try to add some large scale experiments.

2 comments
AuthorsOpenReview ↗

Official Author Response to Reviewer cmh1

Thank you for your thoughtful review. We strengthen the original adaptation bound (Theorem I.1) and show a generalized version of the bound for a general class of GP priors using a noise-regularized version of empowerment (Theorem I.9). We have added experiments on adaptation under matched and anisotropic reward priors. We also address some of your detailed concerns and add clarifications in the responses below. Do these revisions fully address your concerns? We look forward to continuing the discussion.

Conceptual mapping & diagram/table linking potential/effective empowerment, centrality, and geometries

Each initial state leads to a set of possible discounted state occupancies. Potential empowerment is the KL radius of this set, while effective empowerment measures the spread of the distributions for a particular skill family. We clarify this in §4.1.

These have intuitive interpretations. Low hitting times connect potential empowerment to environment centrality. The adaptation result relates effective empowerment to expected post-adaptation return given a centered Gaussian process reward. In the latter case, expected best-skill return is the Gaussian width of the skill-occupancy offsets. The form of the empowerment in this bound depends on this reward prior, and the counterexamples show why it does not directly extend to arbitrary reward geometry (§4.3).

Fig. 1 mountain analogy ("being in control" vs. "exerting control") & link from centrality to adaptation

Potential empowerment measures the control available from a state. Navigating to the mountaintop increases the options available before a task is specified. Effective empowerment measures the control exercised by a particular skill family, such as descending the mountain in different directions. Once the reward is revealed, the agent selects the skill with the highest return. We clarify this distinction in Fig. 1.

The connection to adaptation requires an additional assumption about rewards. Under the matched Gaussian prior, Theorem 4.3 lower-bounds expected best-skill return using effective empowerment.

Constructive covariance- or metric-aware objective aligning information and reward geometries

Our primary contribution here is not a new MISL algorithm. What we instead show is that the notion of empowerment derived from skill learning methods can be directly connected to downstream average-case adaptation. In the revision, we strengthen this result for a general class of GP priors that have anisotropic variance and nontrivial covariance structure.

The same information content can support very different adaptation returns under different priors. The tightened bound models this effect by noise-regularizing the mutual information based on the covariance structure of the reward prior (Theorem I.9).

Small-scale tabular scope vs. scalable empowerment claims & larger-scale validation

We have revised the manuscript to more clearly state that the experiments do not establish scaling to larger, more complex continuous environments. With the exact simplex geometry, we are able to cleanly study the behavior of empowerment and adaptation without the complexity of neural estimators.

To strengthen the evaluations, we ablate horizon, doorway placement, grid size, and transition noise. We also evaluate adaptation under non-diagonal Gaussian rewards and rotated reward covariances. These experiments test the theoretical predictions and their limitations, rather than the performance of a new skill-learning algorithm (Appendix I).

Reviewer_cmh1OpenReview ↗

Response

I appreciate the authors' response and revision, and I have read the revision, especially for the abstract and introduction. It is more clear for me. I keep my positive score.

Reviewer_5o1wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies empowerment, an information-theoretic measure of an agent’s control over its future. It distinguishes two forms. Potential empowerment (Klyubin et al.) is recast as the channel capacity from latent skills to a future state reached at a geometric stopping time. Effective empowerment is the mutual information achieved by a particular skill-conditioned policy, which is the quantity optimized by mutual-information skill-learning (MISL) methods. The main result (Theorem 4.1), which extends Eysenbach et al. (2021), shows that the potential empowerment of a state equals the maximum KL divergence between the occupancies reachable from it and their center. The authors interpret this as a formal version of a conjecture that highly empowered states are central, where centrality is measured by KL divergence. The paper also gives a hitting-time interpretation of empowerment in discrete MDPs (Lemma 4.2). It further shows that in tabular MDPs, the mutual information between skills and the future states they reach lower-bounds the expected return of the best pre-trained skill when each state’s reward is drawn independently from a Gaussian with the same variance (Theorem 4.3). The paper then presents negative results showing that this guarantee does not extend beyond those assumptions: a state can have arbitrarily large empowerment yet allow arbitrarily little adaptation, both with Lipschitz rewards in continuous-state MDPs (Theorem 4.4) and in discrete MDPs when Gaussian rewards are correlated across states (Theorem 4.5).

My recommendation is weak reject. Overall, the contribution is modest relative to the claims.

Critical strengths

The paper’s appeal is a formal treatment linking empowerment and skill learning.

Its strongest result is Theorem 4.1, which connects the mutual information between skills and future states to the maximum KL divergence between reachable occupancies and their center.

Critical weaknesses

There are 3 main weaknesses: Theorem 4.1 is a modest extension of known work, the remaining results claim more than they show, and the presentation makes the contribution hard to assess.

Theorem 4.1 is a modest extension of known work. It combines Lemmas 6.1–6.2 of Eysenbach et al. (2021) with the standard form of channel capacity. Equation 6 is essentially their Lemma 6.2. The difference is that Eq. 6 takes the maximum KL divergence over the entire reachable polytope P(s₀) rather than over a given set of skills. Equation 5 applies a result that a channel’s capacity equals the maximum KL divergence between its output distributions and their center (Theorem A.2, citing Polyanskiy and Wu, 2022) to the channel from skills to occupancies.

Lemma 4.2’s connection to bottleneck states is unclear. Lemma 4.2 restates potential empowerment in terms of expected hitting times. The repeated-rollout hitting time is geometric with mean one over the termination probability, so the lemma follows by substituting this into the definition of mutual information. To be fair, the reading "committing to a skill cuts the number of tries needed to land at s₊" is reasonable intuition, but the biggest problem is presenting it as a result about bottleneck states, which are never defined in this paper, and hitting times are not usually how bottleneck states are defined (e.g., see interconnectedness (Simsek, 2005)). I also think the example is misleading. The choice of γ\gamma will affect πpot∗\pi_{\rm pot}^* and change the very provided example. For small γ\gamma, a hallway-state connecting two rooms (an obvious bottleneck state with high interconnectedness) would not be chosen as its potential empowerment would be bounded by the number of possible next states. In summary, the Lemma is not about bottleneck states as claimed.

Theorem 4.3’s guarantee is narrow. The adaptation lower bound is established for a task distribution with no structure across states. Rewards are drawn independently at each state, so neighbouring states are no more alike in reward than distant ones. The paper’s later results show the lower bound can fail once such structure is present (Theorems 4.4 and 4.5). The claimed "partial answer to Salge’s longstanding hypothesis" (L58) therefore holds only in a very narrow setting.

Presentation. The term "skill" refers to at least four different objects: a latent vector z \in \mathbb{R}^d (L110–112; Fig. 1 caption), the single policy obtained by fixing z (L152–153), the entire skill-conditioned policy family, which the paper also calls the effective empowerment policy (L177–181), and the state-occupancy distribution a skill induces (Fig. 3a; L709–710). Several terms are used without definition, including "long horizon generalizations" (L54), "arbitrarily empowered" (L63, which does not say whether potential or effective empowerment is meant), and "extremized state distributions" (L237). The introduction ends (L67) without stating the paper’s contributions, and "providing a partial answer to Salge’s longstanding hypothesis" (L58) does not say what is actually being answered. Figures 2 and 4 are not referenced in the text.

Post response comments:

Concerns largely unaddressed, and I don't see how this paper is suitable for publication. The paper made claims that it settled long-standing open questions but then it turns out that nearly ever one of the paper's claims is far weaker than claimed. The authors admit those weaknesses and soften language, completely rewriting their contributions from the introduction. But that's not what author response is for. This is not a collaborative exercise where we collectively discover what your paper is about and what its contributions are.

I'll just repeat the same remarks from the review. Theorem 4.1 is still the most interesting result and it's only a modest extension of a known result. Lemma 4.2 is not about bottleneck states (which the original paper emphasized) and making it clear that it's about hitting times means it's not a formalization of a sought after result. Theorem 4.3's guarantee was exceptionally narrow as to be nearly vacuous. My glance at the new 4.3 (but seriously! we're writing new central theorems in the author response and expecting that to change the outcome) is that it doesn't really change the criticism that it is only dealing with unstructured rewards.

This is in addition to the presentation being extremely hard to follow (review scores on presentation of 1, 1, 2). On that alone I would say it's not ready for publication. But we also don't have a set of significantly interesting theoretical advances, which were initially overclaimed, and are now being rewritten in the response.

Minor comments and typos:

  • L106: "over an Markov" → "over a Markov".
  • L125: p(z|s_0) is used before it is introduced as the source (L184).
  • L189: "relabeling skills" is not explained.
  • L235: "Effective empowerment is a mutual information" repeats Definition 3.3.
  • L289: "at the the bottom"
  • L331–332: "a Gaussian distribution centered at a fixed baseline distribution" does not match Appendix F, where rewards are mean zero and the baseline sets the variance.
  • L689: "simplications"
  • L848: "Rather selecting"
  • L875: "intial"
  • L959: "the the RHS"
  • L1281: "adapation"
Questions for authors

None.

3 comments
AuthorsOpenReview ↗

Official Author Response to Reviewer 5o1w (1/2)

Thank you for the detailed review. We clarify the novelty of our contribution: connecting a modified empowerment construction to a general expected adaptation bound. We also strengthen the original adaptation bound (Theorem I.1) and show a generalized version of the bound for a general class of GP priors using a noise-regularized version of empowerment (Theorem I.9). We have added experiments on adaptation under matched and anisotropic reward priors. Do these additions and edits address your concerns fully? We hope to continue the discussion.

Theorem 4.1 novelty vs. Eysenbach et al. (2021) & Polyanskiy–Wu (2022)

We clarify the novelty of our approach relative to established results on channel capacity and information radius duality. Eysenbach et al. [1; Lemmas 6.1 and 6.2] work in a discrete setting, whereas Theorem 4.1 applies the duality from [3, 4, 5].

Lemma 4.2 geometric hitting-time identity & bottleneck overclaim (Simsek, 2005)

With independent resets to start state s0s_0, the number of rollouts to reach a target state is geometrically distributed with mean 1/pπ(s+∣s0)1/p^\pi(s_+\mid s_0), or 1/pπ(s+∣s0,z)1/p^\pi(s_+\mid s_0,z) for a fixed skill. This differs from the per-step stopping probability 1−γ1-\gamma. Substituting this into the MI definition gives the expected log ratio of rollout counts in Eq. (10). This is distinct from contiguous rollouts without any resets, with continuation discount β\beta.

We have modified the wording: doorway preference depends on both the horizon and the layout, as our additional experiments confirm (Appendix I.2). We also rename Lemma 4.2 “Discounted rollout hitting-time reduction” and qualify the discussion of Fig. 4 in the revised text.

Small-γ\gamma sensitivity & short-horizon hallway counterexample

We agree that a doorway need not maximize empowerment at short horizons.

On 8×88\times8 center-doorway grids, the doorway maximizes empowerment in most runs from γ≈0.5\gamma\approx0.5 upward. The transition shifts to longer horizons on 12×1212\times12 grids. Corner doorways can exceed the grid mean at high discounts without being global maximizers (Table 3).

Theorem 4.3 narrow unstructured prior vs. structured rewards (Thms. 4.4–4.5)

We have added a stronger bound on a regularized empowerment objective which allows for a general class of GP priors (Theorem I.9). The key is to regularize the future state (representation) s+s^+ with some form of noise that matches the kernel of the reward prior. For a fixed repertoire and skill prior, let MM be the noisy observation marginal and ℓs=dQϵ(⋅∣s)/dM\ell_s=\mathrm{d}Q_\epsilon(\cdot\mid s)/\mathrm{d}M. The centered kernel is K(s,t)=∫(ℓs−1)(ℓt−1) dMK(s,t)=\int(\ell_s-1)(\ell_t-1)\,\mathrm{d}M. At covariance K/AK/A, we prove J≥(eIϵ−1)/(πA)J\ge\sqrt{(e^{I_\epsilon}-1)/(\pi A)} under weak measurability and regularity assumptions formally stated in Appendix I.10.

Our new bound also works for non-isotropic GP priors. Details are presented in Theorem I.18.

L110–181, Fig. 3a: Overloaded use of "skill" across four mathematical objects

We use “skill” generally to refer to the behaviors and characteristics induced by conditioning on a specific latent zz. We clarify in Table 1 the distinct ways we use this term: depending on context, it refers to a conditioned policy, a policy family, a latent, or a family of discounted state occupancy distributions.

L54, L63, L237: Undefined "long horizon generalizations", "arbitrarily empowered", and "extremized state distributions"

We now define “long-horizon generalization” as successive rollouts with continuation discount β\beta in Appendix C.2.

We clarify that “arbitrarily empowered” refers to mutual information growing arbitrarily with increasing skill count and state that the extremized state distributions are the vertices of the reachable occupancy polytope.

L58, L67: Missing contribution roadmap & scope of "partial answer to Salge’s hypothesis"

We clarify that our core contribution in this paper was the geometric and theoretical exploration of a concrete and practical empowerment objective based on skill learning.

Regarding Salge et al. [2], we clarify that Theorem 4.3 lower-bounds expected post-adaptation return rather than establishing a full ranking over states. We note that our new, strengthened bound (Theorem I.9) does, however, enable analysis of continuous environments and a general class of GP priors.

Figs. 2 & 4, L125, L189, L331–332: Figure callouts, skill source p(z∣s0)p(z\mid s_0), relabeling, and Gaussian prior variance

Thank you for noticing these. We now define the skill prior p(z∣s0)p(z\mid s_0) in §3.1 and explain how we fix an atomless continuous uniform prior by relabeling in §3.2. We have also clarified the relabeling scheme and prior variance notation.

AuthorsOpenReview ↗

Official Author Response to Reviewer 5o1w (2/2)

References

Dear Reviewer 501w,

Thank you for your thoughtful feedback. We have worked hard to incorporate your feedback with added experiments, revisions, and a stronger, generalized version of the 4.3 bound that applies to a general class of GP priors.

Have our revisions and response sufficiently resolved your questions and concerns? We appreciate your engagement as the end of the discussion phase nears. Thank you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Response Summary

We thank the reviewers for their thoughtful feedback. We have revised the manuscript with changes highlighted in blue and added experiments based on reviewer feedback and a stronger, generalized version of the bound that applies to a general class of GP priors. We address specific reviewer concerns in more detail below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.