acceptodds
Under review as a conference paper at ICLR 2027

GIRO: Robust Reinforcement Learning with Generative Model Induced Uncertainty Set

19 views

Abstract

Offline model-based reinforcement learning uses learned transition dynamics to generate training experience, but policies optimized against an estimated model may exploit its errors and fail under distribution shift. We introduce GIRO, a distributionally robust RL method that defines an uncertainty set of generative transition models through a training-loss budget. Because shared model parameters couple transitions across state–action pairs, the resulting set is generally non-rectangular, making robust policy evaluation challenging. GIRO alternates adversarial model updates with robust policy improvement. For robust policy evaluation, it uses a primal-only update that decreases policy value when the loss constraint is satisfied and reduces constraint violation otherwise. Under stated assumptions, we bound robust evaluation suboptimality and constraint violation and establish a stationarity guarantee for policy improvement. Across nine D4RL task–dataset combinations and eight test perturbations, GIRO attains the highest aggregate worst-case return and lowest aggregate across-perturbation variance among the compared methods.

Primary area:
reinforcement learning
Keywords:
Distributional Robust Reinforcement Learning, Generative World Models, Model-Based Reinforcement Learning

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_7Mf4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GIRO, a robust reinforcement learning method that defines an uncertainty set of generative world models through a training-loss constraint, using diffusion models or conditional VAEs. GIRO alternates between an inner loop that adversarially updates the world model to minimize the current policy’s return while learning a value function from generated trajectories, and an outer loop that improves the policy under the resulting model. The method is evaluated on D4RL tasks with observation and action perturbations, as well as changes to body mass and friction.

Critical strengths
  1. The method design is well aligned with its goal of learning robust policies. Updating the policy using the worst-case generative model within an uncertainty set, along with its corresponding value function, provides a convincing rationale for improved generalization. The method consistently follows this objective.

  2. The paper provides clear theoretical insights into how GIRO mitigates world-model overoptimization differently from KL-regularized optimization and characterizes the convergence bounds for both the value function and the policy.

Critical weaknesses
  1. Modest gains over a strong baseline. In Table 1, GIRO with VAE improves the aggregate worst-case score over MOBILE by only approximately 1% (714.2 vs. 707.1), while GIRO-Diffusion underperforms MOBILE in Table 7 (691.0 vs. 707.1). These marginal gains raise doubts about whether GIRO’s benefits justify the complexity of its adversarial updates.

  2. Lower average performance. Averaging the reported scores in Table 6 equally across the eight perturbations and nine task–dataset combinations yields 90.37 for MOBILE, 85.80 for GIRO-Diffusion, and 84.46 for GIRO-VAE. Thus, both GIRO variants underperform MOBILE on average, indicating weak robustness capability across diverse perturbations.

  3. Limited perturbation range. The mass and friction evaluations only use scaling factors of 0.95 and 1.05. A broader sweep, such as 0.8–1.2, would better characterize how performance degrades as the dynamics shift becomes more severe. The current narrow range provides limited evidence of robustness beyond small deviations from the nominal environment.

Questions for authors

Please see the weakness section.

4 comments

Response to Reviewer 7Mf4

We thank the reviewer for the careful assessment and for recognizing the alignment between GIRO’s design and its robustness objective, as well as the value of its theoretical analysis. We address the three empirical concerns below.

Response to Weakness 1: Modest Gains Over a Strong Baseline

We agree that the improvement in aggregate worst-case return over MOBILE is modest (714.2 versus 707.1). However, GIRO (VAE) also reduces the sum of disturbance variances from 496.09 to 137.78, a reduction of approximately 72%, with lower variance on seven of nine datasets. This stability is consistent with the intended benefit of our minimax formulation: optimizing against adverse dynamics encourages policies that remain effective across plausible model variations. Thus, the results show substantially greater consistency across disturbances alongside the modest improvement in aggregate worst-case return.

Response to Weakness 2: Lower Average Performance

We acknowledge that MOBILE and RAMBO achieve higher overall average returns. As discussed in our response to Weakness 1, GIRO achieves a higher aggregate worst-case return than MOBILE and substantially lower disturbance variance, consistent with its minimax objective.

If average performance across disturbed transition models were the primary goal, the corresponding objective would be max⁡π1m∑j=1mVπPj(ρ)\max_{\pi}\frac{1}{m}\sum_{j=1}^{m}V_{\pi}^{\mathsf{P}_j}(\rho). GIRO instead optimizes against the worst admissible dynamics. The average-return comparison identifies a limitation of our method, but alone does not establish weaker worst-case robustness.

Response to Weakness 3: Limited Perturbation Range

We thank the reviewer for this suggestion. We evaluated all methods under body-mass and friction scaling factors of 0.8, 0.9, 0.95, 1.05, 1.1, and 1.2, using 20 episodes per condition and three seeds. Our original hyperparameters were tuned for small perturbations. Under this wider range, the original GIRO (VAE) checkpoints achieve a lower sum of worst-case returns than RAMBO (476.4 versus 530.5).

To adapt GIRO to the broader perturbation range, we fine-tuned the existing GIRO (VAE) checkpoints for an additional 200,000 training steps. We used the same settings across all datasets: a larger uncertainty budget (budget scale 1.02, compared with the original 0.94–1.00) and a larger adversarial update step size (3×10−63\times10^{-6}, compared with 10−710^{-7}).

For every dataset, we report the checkpoint after these additional 200,000 steps. The fine-tuned GIRO checkpoints use 1.2 million training steps in total, whereas the evaluated CQL, MOPO, IQL, and RAMBO checkpoints use 2 million steps. Experiments with GIRO (Diffusion) are still running, and we will include those results once available.

MethodWorst-case sum ↑Variance sum ↓
Model-based SAC453.51366.7
CQL408.92078.1
MOPO465.61753.2
IQL395.42029.6
RAMBO530.5829.0
MOBILE483.62703.2
GIRO (VAE)476.41445.4
GIRO (VAE, fine-tuned)603.4711.5

Fine-tuned GIRO (VAE) achieves the highest sum of worst-case returns among the evaluated methods (603.4, compared with 530.5 for RAMBO). Relative to the original GIRO checkpoints, fine-tuning improves worst-case returns on seven of nine datasets. The largest gains occur on Hopper-ME (37.5 → 102.6), Walker2d-ME (63.8 → 86.1), Walker2d-M (65.8 → 84.4), and Walker2d-MR (49.2 → 61.1). Performance is essentially unchanged on HalfCheetah-M (62.9 → 62.8) and decreases slightly on HalfCheetah-ME (59.3 → 57.1).

These results demonstrate improved robustness after adapting GIRO’s budget and update settings to the broader perturbation range. We include the detailed per-task, per-disturbance results and the fine-tuning configuration in Appendix C.3.

Thank you for the detailed rebuttal and additional experiments. I understand the worst-case objective of GIRO, but remain unconvinced that its robustness benefits justify the lower average returns relative to RAMBO and MOBILE. Lower variance alone does not establish this benefit. For example, the RAMBO variance of approximately 85 on HC-ME corresponds to a standard deviation of 9.2, which alone does not indicate severe degradation given the original score of RAMBO (93.7).

Under the wider perturbation range, the original GIRO-VAE underperforms RAMBO on both metrics. The fine-tuned variant achieves a improvement, but its settings were specifically adjusted for the expanded range. The rebuttal does not establish whether the baselines received comparable tuning, leaving uncertainty about the comparability of these results.

Given these remaining concerns, I maintain my original rating.

Thank you for the follow-up. We agree that comparing a fine-tuned GIRO with baselines that were not tuned for the wider range is not a fair comparison, so we fine-tuned RAMBO and MOBILE with the same protocol.

Protocol. Starting from the released OfflineRL-Kit checkpoints used in the paper (seeds 1–3), each baseline is trained for another 200K steps, as for GIRO. Only the robustness hyperparameters are changed; all other settings follow the released per-dataset configurations.

  • RAMBO: adversarial weight λ∈{3×10−4,10−3,3×10−3}\lambda \in \lbrace 3\times10^{-4}, 10^{-3}, 3\times10^{-3} \rbrace (released: 00 or 3×10−43\times10^{-4}, depending on the dataset), and a larger adversary learning rate (10−310^{-3} vs. 3×10−43\times10^{-4}).
  • MOBILE: uncertainty penalty β\beta at ×1\times 1, ×2\times 2 and ×4\times 4 the released per-dataset values.

As for GIRO, each method uses one configuration for all nine datasets, selected by total worst-case return on screening runs, and we report the final checkpoint. This selects λ=3×10−3\lambda = 3\times10^{-3} for RAMBO (10×10\times the largest released value) and the original β\beta for MOBILE.

MethodNominalMean (perturbed)Worst-case
RAMBO (released)734.3708.4530.5
RAMBO (fine-tuned, λ=3×10−3\lambda = 3\times10^{-3})746.1681.5515.2
MOBILE (released)860.1766.3483.6
MOBILE (fine-tuned, original β\beta)815.2763.6499.7
GIRO-VAE (original)777.2731.4476.4
GIRO-VAE (fine-tuned)815.8792.0603.4

Normalized scores summed over the nine datasets (3 seeds, 20 episodes). Body mass and friction are scaled by {0.8,0.9,0.95,1.05,1.1,1.2}\lbrace 0.8, 0.9, 0.95, 1.05, 1.1, 1.2 \rbrace; worst-case is the minimum over these 12 settings of the 3-seed mean.

Larger robustness hyperparameters do not help the baselines. Fine-tuned RAMBO stays below its released checkpoint under perturbation, and doubling or quadrupling MOBILE's penalty lowers its worst-case return on the screening runs (502→425→304502 \rightarrow 425 \rightarrow 304). Training also becomes less stable: RAMBO diverges on hopper-medium-expert with the larger adversary learning rate, and MOBILE collapses on halfcheetah- and walker2d-medium-expert with the ×4\times 4 penalty. Under the same protocol, GIRO-VAE has the highest worst-case and mean perturbed return, and its nominal return is on par with fine-tuned MOBILE.

Why GIRO benefits where RAMBO does not. RAMBO adds the adversarial term to the model's supervised loss with a fixed weight λ\lambda, so how far the model may move from the data is never set directly. It depends on λ\lambda, on the adversary's step size and number of updates, and on the scale of the supervised loss, which differs between datasets. This is why the released λ\lambda varies by dataset, and why a larger λ\lambda or step size gives inconsistent gains or divergence. GIRO instead keeps the model's fit loss on the offline data within a budget relative to the nominal model's loss, so its uncertainty set is explicitly the set of models that explain the data almost as well as the nominal one, and one relative budget works for all nine datasets.

Why fine-tuning. In the paper, GIRO's budget had to be kept small: with a large budget, training from scratch diverges, because the critic is still inaccurate early on and the adversary can push transitions far from the data. That set does not cover $\pm$20% perturbations, which explains the original results in the wider range. Starting from a converged checkpoint, the critic is stable enough for a larger budget, so fine-tuning works as a two-stage budget schedule. Nothing in it is specific to GIRO, which is why we applied it to the baselines as well.

On variance and large perturbations. We agree that variance alone is not sufficient, which is why worst-case return is our main metric. With fine-tuning, GIRO-VAE is now the strongest method at the largest perturbations. At body mass and friction ×0.8\times 0.8 and ×1.2\times 1.2, its worst-case sum is 603.4, compared with at most 530.5 for the baselines (released or fine-tuned), and its mean return over these four settings is 754.5, compared with at most 670.8. It has the highest worst-case return at these settings on 5 of the 9 datasets, with the largest margins on hopper-medium-expert (102.6 vs. 61.0 for RAMBO and 40.9 for MOBILE) and walker2d-medium-expert (86.1 vs. 76.1). HalfCheetah, including HC-ME, remains the exception: the baselines lose little there under these perturbations, and GIRO does not improve on them.

We have added these results and the corresponding implementation details to Appendix C.3 of the revised manuscript.

Dear Reviewer 7Mf4, as the discussion period is ending soon, we would appreciate it if you could review our follow-up experiments with fine-tuned baselines and let us know whether they address your concerns. Thank you for your time.

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

This paper introduces GIRO, a diffusion model-based method for generating uncertainty sets in robust offline reinforcement learning. A generative model is trained to predict an underlying distribution that could have produced the offline RL dataset, such that the proposed distribution induces worst-case performance in a policy trained on the offline dataset. Then, a robust policy is trained to improve performance on this distribution, and this iteration continues until convergence.

Critical strengths
  • The paper addresses an important topic in robust reinforcement learning, i.e. constructing plausible uncertainty sets for the purpose of determining potential worst-case performance.
  • The paper sets up the intuition well such that the later formal sections are easy to anticipate and follow.
Critical weaknesses
  • There is some confusion around the diffusion indexing that may just be notational. In Eq. 10, how is Pζ(x0:K∣s,a)P_\zeta (x_{0:K}|s,a) computed precisely? Is this the joint probability for x0,x1,⋯ ,xKx_0, x_1, \cdots, x_K?
  • The presentation of the results via a summed performance is an unusual choice. It may be worth considering highlighting the best score in each row, with some analysis of when GIRO or other methods are the best performing and why.
Questions for authors
  • Are the results in the main body of the paper with a diffusion model or a VAE? This should be stated for clarity in the table captions.
  • Could the authors clarify on the notation in the weaknesses section?
  • What is the distinction between GIRO and each of the prior works used as baseline methods?

Clarifying the above points will be sufficient for an accepting score.

1 comment

Response to Reviewer WRzq

We thank the reviewer for the constructive feedback and positive assessment of the problem’s importance and the paper’s presentation. We address each point below.

Response to Weakness 1: Diffusion Indexing and Notation

Thank you for pointing out this ambiguity. We have clarified the definition of PξP_{\xi} around Eq. 10. For a discrete-time diffusion process, Pξ(x0:K∣s,a)P_{\xi}(x_{0:K}\mid s,a) is the conditional joint density of (x0,x1,…,xK)(x_0,x_1,\ldots,x_K). By the Markov structure of the reverse denoising process,

Pξ(x0:K∣s,a)=ν(xK)∏k=1KPξ,k−1∣k(xk−1∣xk,s,a).P_{\xi}(x_{0:K}\mid s,a) = \nu(x_K)\prod_{k=1}^{K} P_{\xi,k-1\mid k}(x_{k-1}\mid x_k,s,a).

where ν=N(0,I)\nu=\mathcal{N}(0,I) is the noise prior and Pξ,k−1∣kP_{\xi,k-1\mid k} denotes the reverse transition density. Its log-density is therefore calculated by summing the log-density of the noise prior and those of the reverse transitions along the trajectory. Here, x0=s′x_0=s' is the generated next state, and kk indexes diffusion steps rather than environment time.

Our theoretical analysis uses continuous-time diffusion, with the path-measure KL characterization derived through Girsanov’s theorem in Appendix A.1. Our experiments use discrete-time diffusion and do not require evaluating a continuous-time path density.

Response to Weakness 2: Presentation and Analysis of Performance Results

We have added a task-level discussion in Appendix C.2. Across the nine datasets, GIRO (VAE) achieves the highest worst-case return on Hopper-M and Hopper-MR, while GIRO (Diffusion) leads on HC-ME and Walker2d-MR. RAMBO leads on four datasets and MOBILE on one.

The discussion also identifies shared vulnerabilities. State-IID is the worst disturbance for every policy on HC-ME, and all worst cases on HC-MR occur under State-IID or State-Static. On Hopper-M, Action-IID is worst for four baselines, whereas GIRO (VAE) maintains a return of 100.5 under this condition.

GIRO (VAE)’s largest gains over the strongest baselines occur on Hopper-M and Hopper-MR, where several baselines experience large performance drops under perturbations. Its worst-case returns are 97.9 and 95.9, compared with 82.4 and 83.5 for the strongest baselines, respectively. These gains contribute to its higher aggregate worst-case score and are consistent with its adversarial training objective.

Response to Question 1: World Model Used in the Main Results

Thank you for pointing out this omission. The GIRO results in Tables 1–2 use a conditional VAE (CVAE) world model. Model-based SAC also uses a CVAE. We have clarified this in the main text and both table captions.

Response to Question 2

Please see our response to Weakness 1 above.

Response to Question 3: Distinctions Between GIRO and the Baseline Methods

We have added a baseline comparison in Appendix B.2. CQL and IQL are model-free: CQL regularizes Q-values to mitigate overestimation, while IQL uses expectile value regression and advantage-weighted behavioral cloning. MOPO and MOBILE introduce uncertainty-based penalties: MOPO penalizes synthetic rewards using model uncertainty, while MOBILE penalizes Bellman estimates using disagreement across dynamics models. GIRO instead searches for an adverse world model subject to a training-loss budget.

RAMBO is the closest baseline because it also alternates policy improvement with adversarial model updates. Although RAMBO is motivated by a constrained robust RL formulation, its practical algorithm combines adversarial return minimization with a fixed-weight negative log-likelihood penalty.

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

This paper proposes a robust offline model-based RL algorithm GIRO. Its uncertainty set contains all generative models of which the loss can stay within a budget. This makes the set non-rectangular. The proposed GIRO can alternate between the updating of adversarial world model and that of robust policy. The convergence proof of these two loops is provided. GIRO also shows great performance on selected benchmarks in the paper.

Critical strengths
  1. Since the uncertainty set is defined on the budget of training loss, the specific tuning for each radius is avoided and meanwhile the coupling of different transitions is reserved.
  2. The performance is very robust across tasks.
Critical weaknesses
  1. There is a gap between the theory and implementation. The theory uses a lot of natural gradients and many inner loops but the plain gradients are used in the implementation.
  2. The novelty of GIRO over two algorithms: RAMBO (Rigter et al., 2022) and Wen & Yang (2026) is not clear.
Questions for authors
  1. Please explain the novelty of GIRO compared with the mentioned two methods in the Weakness. These two papers are cited in your draft.
  2. The diffusion version does worse than the version of CVAE in 1) total worst-case return and 2) variance. But diffusion is more expressive, why didn't this method help? Will this weaken the motivation for the usage of generative world models?
4 comments

Response to Reviewer pAqb

We thank the reviewer for the careful evaluation and for recognizing the benefits of our loss-budget uncertainty set and the robustness of GIRO across tasks. We address the main concerns below.

Response to Weakness 1: Gap Between Theory and Implementation

We use natural-gradient updates in the theoretical algorithm because their Fisher-based geometry supports global convergence analysis under function approximation, subject to the stated assumptions and approximation errors. This structure enables us to analyze adversarial model optimization and its interaction with policy improvement. The theoretical inner-loop schedule also controls the model-optimization error before each outer update.

In practice, we use plain-gradient updates to reduce computational cost, avoiding the Fisher–vector products and linear-system solves required by natural gradients. The adversarial objective, uncertainty budget, and projection step remain unchanged. These implementation choices are practical approximations; the theoretical guarantees do not directly apply to the implemented update scheme.

To assess the gradient choice, we compared plain-gradient and approximate natural-gradient adversarial updates for GIRO (Diffusion). We approximate the direction d=(F+λI)−1gd=(F+\lambda I)^{-1}g using implicit Fisher–vector products and 10 conjugate-gradient iterations, with damping λ=10−2\lambda=10^{-2} and a KL trust-region target of δ=3×10−6\delta=3\times10^{-6}. The number of inner updates, rollout length, uncertainty budget, and projection step are identical between the two versions.

On HalfCheetah-medium-replay, using 10610^6 training steps, three seeds, and 20 evaluation episodes per condition, both versions achieve a reported worst-case return of 55.255.2 across the 16 perturbations. Their average returns are also close: 63.763.7 for plain gradients and 64.464.4 for approximate natural gradients.

Evaluation conditionPlain gradientApproximate natural gradient
Nominal65.6±0.365.6 \pm 0.365.9±0.865.9 \pm 0.8
State/action noise (mean of 4)59.6±0.659.6 \pm 0.660.8±0.760.8 \pm 0.7
Mass/friction scaled by 0.95,1.050.95, 1.05 (mean of 4)65.5±0.365.5 \pm 0.365.9±0.765.9 \pm 0.7
Mass/friction scaled by 0.8,0.9,1.1,1.20.8, 0.9, 1.1, 1.2 (mean of 8)64.9±0.364.9 \pm 0.365.5±0.765.5 \pm 0.7
Mass/friction scaled by 0.8,0.9,1.1,1.20.8, 0.9, 1.1, 1.2 (worst of 8)61.8±0.661.8 \pm 0.662.6±1.362.6 \pm 1.3
Mean over all 16 perturbations63.7±0.463.7 \pm 0.464.4±0.764.4 \pm 0.7
Worst over all 16 perturbations55.2±0.355.2 \pm 0.355.2±0.255.2 \pm 0.2

This comparison supports the use of the less expensive plain-gradient update on this benchmark, although it does not establish equivalence between the two update rules more generally.

Inner-loop updates. We use one inner update per outer update to keep training computationally practical. Running the inner loop to convergence before every policy update would exceed our computational budget. We start from a behavior-cloning-initialized policy and observe smaller critic changes as training progresses. We therefore interleave inner and outer updates as a practical approximation, similar to actor–critic training. This choice does not ensure the inner-loop accuracy required by the theoretical schedule.

Response to Weakness 2: Novelty Relative to RAMBO and Wen & Yang (2026)

We acknowledge that GIRO shares adversarial world-model updates with RAMBO and a training-loss-based uncertainty-set construction with GAS-DRO. However, GIRO goes beyond combining these components in three respects.

First, although RAMBO formulates a robust maximin problem, its practical algorithm replaces the uncertainty constraint with a fixed-weight likelihood penalty. This regularized surrogate does not, by itself, guarantee satisfaction of a prescribed uncertainty budget. Our analyzed primal algorithm directly addresses the constrained robust RL problem and provides convergence and constraint feasibility guarantees under the stated assumptions.

Second, extending the uncertainty-set construction of GAS-DRO to multi-step RL is nontrivial. The adversary optimizes conditional transition distributions Pξ(s′∣s,a)P_{\xi}(s'\mid s,a) rather than a single data distribution, and changing these transitions also changes state visitation and long-term returns. Moreover, shared world-model parameters and the global training-loss budget couple transitions across state–action pairs, yielding a non-rectangular uncertainty set that prevents independent adversarial optimization at each pair. To the best of our knowledge, this specific setting—constrained robust RL with a training-loss-defined, non-rectangular uncertainty set over conditional generative world models—has not previously been analyzed with both convergence and constraint feasibility guarantees. GIRO addresses this setting through a primal algorithm and an analysis of its iterative inner updates and outer policy improvement.

Third, GAS-DRO’s convergence proof assumes that the generative model globally maximizes the Lagrangian for each dual variable (Appendix C.2.2, Eq. (42)); it does not establish convergence of the model updates used to obtain that maximizer. Our analysis explicitly studies the iterative inner updates and outer policy improvement, providing convergence and feasibility guarantees under the stated assumptions rather than assuming an exact model-optimization oracle.

Response to Question 1

Please see our response to Weakness 2 above.

Response to Question 2: Why Does Diffusion Underperform CVAE, and What Are the Implications for Generative World Models?

We investigated this difference on Hopper-ME, one of the tasks where the diffusion version is less robust than the CVAE version. For the same 3,0003{,}000 dataset state–action pairs, we replaced the recorded next states with simulator-generated next states under body-mass or friction scaling of 0.950.95 and 1.051.05. We then evaluated each fixed nominal world model’s training loss on these disturbed transitions.

The body-mass perturbations increase the CVAE loss by 4.8% of its nominal value, compared with 162% for diffusion. The corresponding increases under friction perturbations are 70% and 459%. Thus, small physical changes produce substantially larger relative changes in the diffusion objective.

For the body-mass diagnostic, the disturbed-data losses are approximately 1.051.05 times the nominal loss for CVAE and 2.622.62 times the nominal loss for diffusion. These ratios measure loss sensitivity; they do not directly establish the uncertainty budget required to include the disturbed transition kernel, because the diagnostic evaluates a fixed nominal model on disturbed data.

Increasing the diffusion budget also introduces a practical tradeoff: it permits greater model variation but weakens the constraint that keeps the world model consistent with the offline data. On this task, raising the budget to 1.281.28 times the nominal loss already lowers the nominal return from 105.9105.9 to 82.782.7 without improving the worst-case return.

Together, these observations suggest that the uncertainty set induced by our chosen diffusion model, objective, and budget may be less well aligned with the tested physical disturbances than the CVAE-based construction. We regard this mismatch as a plausible contributor to diffusion’s lower robust return, although the diagnostic does not directly establish uncertainty-set coverage. The broader implication is that greater generative-model expressiveness alone does not guarantee better robustness: model choice, training objective, and uncertainty-budget calibration must also account for the disturbances encountered at deployment.

Thank the authors for providing additional information and respond to my questions. I raised my review score to 2.

Thank you for your updated assessment and constructive feedback. We would be happy to address any further questions or concerns you may have.

Comments on the paper

1 comment

We have uploaded a revised manuscript incorporating the clarifications, additional experiments, and discussions detailed in our responses to the reviewers. Changes are highlighted in blue for ease of reference.

We thank the reviewers for their constructive feedback and careful evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.