acceptodds
Under review as a conference paper at ICLR 2027

SWAN: Scaling Long-horizon World Models via Next-Scale Autoregression

12 views

Abstract

Robotic world models must faithfully predict the consequences of prescribed actions, including failures, while preserving physical states over extended horizons. However, existing generative models frequently hallucinate success due to strong visual priors, and struggle with error accumulation or prohibitive memory growth during recursive rollouts. To overcome these challenges, we introduce SWAN, which Scales World modeling through Autoregressive Next-scale prediction by integrating unified action-conditioned prediction with a coarse-to-fine hierarchical streaming memory. First, scale-aligned action representations guide generation across all resolutions to strengthen action-future coupling. Second, a time-scale memory preserves fine recent and coarse distant causal states under a bounded resident-memory budget, coupled with scale-wise dream forcing to bridge the exposure gap. Finally, we align visual dynamics on fresh rollouts via action-following and consistency rewards under fixed controls, without task-success rewards. Experiments demonstrate stable streaming rollouts up to 60 seconds. Across 600 matched sim-to-real rollout pairs with frozen external policies, SWAN achieves a success-rate MAE of 0.062 and a within-task rank correlation of 0.88, while an auxiliary action expert further enables effective real-robot control. Despite residual optimism in contact-rich stages, these results demonstrate the promise of next-scale autoregression for scalable robotic simulation and execution.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, Robotic Manipulation, Next-Scale Autoregression, Policy Evaluation, Long-Horizon Simulation

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

This paper proposes SWAN, a long-horizon robotic world model built upon the pretrained InfinityStar framework. The method introduces scale-aligned action conditioning to improve action-guided visual prediction and a time-scale memory mechanism to preserve historical states under bounded memory constraints. It further incorporates scale-wise dream forcing and on-policy visual alignment to reduce error accumulation and improve the consistency of generated dynamics. Experiments on long-horizon video prediction, simulated policy evaluation, and real-world robotic manipulation demonstrate the effectiveness of the proposed framework.

Critical strengths
  • The paper addresses an important problem in robotic world modeling, particularly the challenges of action-following accuracy, long-horizon consistency, and efficient memory management.

  • The proposed framework integrates action-conditioned generation, hierarchical temporal memory, and rollout-based refinement into a unified world model, providing a practical approach to long-horizon robotic simulation.

  • The experimental evaluation covers multiple aspects, including generation quality, action following, long-horizon consistency, and real-world robotic manipulation. The matched simulation-to-real rollout evaluation further provides useful evidence of the model's potential for robotic policy assessment.

Critical weaknesses
  • The ablation studies do not sufficiently disentangle the contributions of the proposed architectural components from the additional domain-specific training. Since SWAN is initialized from InfinityStar and further trained on extensive robotic data, it is unclear how much of the performance improvement comes from the proposed modules rather than domain adaptation. A more controlled comparison using the same pretrained backbone, training data, and training budget, with individual components progressively introduced, would help establish their respective contributions.

  • The ablation study only compares the full model with a variant without action conditioning, which demonstrates the benefit of action information but does not establish the effectiveness of the proposed scale-aligned design. A comparison with standard action-conditioning strategies, such as feature concatenation or cross-attention without scale alignment, is needed to justify this architectural contribution.

  • The ablation study does not sufficiently validate the necessity of scale-aligned action conditioning. The authors should compare the proposed multi-scale action conditioning with a standard single-scale conditioning strategy under identical settings to demonstrate the specific benefits of scale alignment.

  • The related work provides only a high-level discussion of existing action-conditioned world models, without adequately comparing their action-conditioning mechanisms with the proposed scale-aligned design. The authors should clarify how their approach differs from prior action-conditioning strategies and what specific limitations of these methods motivate scale alignment.

Questions for authors

Refer to Weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer fKFE

We thank Reviewer fKFE for the thoughtful review and constructive feedback. Below, we address W1–W4 in detail.


Q1 (W1):

Comparing pretrained InfinityStar against progressive SWAN variants trained on the exact same robotic dataset directly separates domain adaptation from our proposed modules:

Model Variant (Same Backbone)Robotic DataImage Quality ↑\uparrowAction Following ↑\uparrowTrajectory Acc. ↑\uparrowPolicy Spearman ↑\uparrowSR MAE ↓\downarrow
Pretrained InfinityStarNo0.602—0.251——
SWAN w/o Action (Domain-Adapted)Yes0.6830.0190.4180.6600.112
Supervised SWAN (+ Action, Memory, Dream Forcing)Yes—0.2640.5120.8310.076
Full SWAN (+ On-Policy Visual Alignment)Yes0.6970.2900.5390.8760.062
  1. Domain adaptation alone (Pretrained InfinityStar →\rightarrow SWAN w/o Action) improves Image Quality (0.602→0.6830.602 \rightarrow 0.683) and unconditioned Trajectory Accuracy (0.251→0.4180.251 \rightarrow 0.418), but leaves Action Following near zero (0.0190.019) and policy ranking weak (0.6600.660 Spearman, 0.1120.112 MAE).

  2. SWAN's architectural and alignment modules on the exact same robotic dataset drive the gains in controllable simulation and policy evaluation—lifting Action Following from 0.019→0.264→0.2900.019 \rightarrow 0.264 \rightarrow 0.290, Trajectory Accuracy from 0.418→0.512→0.5390.418 \rightarrow 0.512 \rightarrow 0.539, and policy-ranking Spearman from 0.660→0.831→0.8760.660 \rightarrow 0.831 \rightarrow 0.876 (with component-wise ablations in our response to Reviewer 8FsE Q2 further isolating time-scale memory and dream forcing under the same robotic training set).


Q2 (W2 & W3):

We clarify what SWAN changes relative to standard concatenation or vanilla cross-attention and why multi-scale conditioning matches next-scale generation:

  1. Mechanism vs. standard concatenation and cross-attention. Standard concatenation or vanilla cross-attention exposes a uniform, single-resolution sequence of raw action tokens to all visual queries, requiring attention layers to implicitly infer 3D-to-2D camera projection, clip boundaries, and spatial scale alignment at once. SWAN instead projects 3D actions into each camera's 2D coordinate frame across 8 spatial scales (s0s_0–s7s_7). For visual queries at clip cc, camera vv, and visual scale ℓ∈{0,…,11}\ell \in \{0, \dots, 11\}, SWAN restricts attention to the matched (c,v)(c, v) window and routes queries to the finest action scale κ(ℓ)≤ℓ\kappa(\ell) \le \ell (using s0s_0–s7s_7 for ℓ≤7\ell \le 7 and reusing s7s_7 for s8s_8–s11s_{11}).

  2. Why multi-scale over single-scale representations. In next-scale autoregression, coarse scales determine global arm motion direction and scene layout, while fine scales refine local gripper-object contact geometry. Matching the spatial resolution of the action prefix κ(ℓ)\kappa(\ell) to each visual scale ℓ\ell avoids granularity mismatch across coarse-to-fine prediction steps and directly motivates our choice of a next-scale autoregressive backbone.


Q3 (W4):

Prior action-conditioned world models align actions and video predictions at two primary granularities: (1) Sequence-level conditioning (e.g., UniSim), which conditions generation on global language or low-level control vectors; and (2) Frame-aligned single-scale conditioning (e.g., IRASim, WorldGym, Ctrl-World), which aligns per-frame action tokens with latent video frames inside diffusion or autoregressive transformers operating on a single fixed spatial latent grid. SWAN addresses an orthogonal dimension unique to next-scale autoregression: as generation progresses from coarse layout to fine detail within the same camera view and physical time window, combining 3D-to-camera projection with explicit scale routing κ(ℓ)\kappa(\ell) ensures that each visual scale reads action features at a matched spatial resolution.

Reviewer_8FsEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SWAN, a next-scale autoregressive world model for long-horizon robotic simulation. Building upon InfinityStar, the authors first introduce scale-aligned action conditioning to improve the consistency between prescribed actions and generated future videos. They then propose a time-scale memory mechanism and scale-wise dream forcing to reduce memory consumption and error accumulation during long-horizon generation. Additionally, the authors introduce an on-policy visual alignment strategy to improve action following and temporal consistency through reinforcement learning, together with an auxiliary action expert for robotic control. Experiments demonstrate the effectiveness of SWAN in video generation, 60-second long-horizon rollouts, sim-to-real policy evaluation, and real-world robotic manipulation tasks.

Critical strengths
  1. Clear motivation and practical relevance. The paper addresses important challenges in robotic world modeling, particularly action following, long-horizon consistency, and the accurate simulation of unsuccessful actions. These problems are highly relevant to using generative world models for robot policy evaluation.

  2. Comprehensive framework for long-horizon world modeling. The authors integrate scale-aligned action conditioning, time-scale memory, dream forcing, and reinforcement learning into a next-scale autoregressive framework. The experiments demonstrate improvements in action following, trajectory accuracy, and long-horizon consistency.

  3. Extensive evaluation with real-world validation. Beyond video generation benchmarks, the authors conduct 600 matched sim-to-real rollouts to evaluate policy success-rate calibration and ranking, achieving a success-rate MAE of 0.062 and a mean within-task Spearman correlation of 0.88. Additional real-robot experiments further demonstrate the practical potential of the proposed framework.

Critical weaknesses
  1. Insufficient comparisons with existing policy evaluators. The experiments mainly compare SWAN with general-purpose video generation models, such as Wan, CogVideoX, and InfinityStar, while lacking comparisons with specialized world-model-based policy evaluators, such as WorldGym, Ctrl-World, and dWorldEval. This makes it difficult to assess SWAN's relative effectiveness and contributions to robot policy evaluation.

  2. Limited ablations on policy evaluation. Although the authors evaluate individual components using generation quality and long-horizon consistency metrics, their contributions to the final policy evaluation performance remain unclear. Additional ablations on success-rate calibration, policy ranking, and failure detection would strengthen the experimental validation.

  3. Insufficient validation of scale-aligned action conditioning. The current ablation only compares models with and without action conditioning, which does not demonstrate the advantage of the proposed scale-aligned design over conventional action conditioning methods. A controlled comparison with simpler action injection strategies would better support its effectiveness.

Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8FsE

We thank Reviewer 8FsE for the constructive comments and valuable suggestions. Below, we address W1–W3 in detail.


Q1 (W1):

We clarify the advantages of our matched real-robot evaluation benchmark over simulated benchmarks and provide direct comparisons with specialized world-model evaluators below.

  1. Advantage of matched real-robot evaluation over simulated benchmarks (e.g., LIBERO). Prior evaluators often validate rankings inside deterministic simulators such as LIBERO, which omit physical sensor noise, control latency, and complex contact failures, and frequently pool correlations across tasks of varying difficulty. Because the core goal of a world-model evaluator is to replace costly physical testing, our benchmark evaluates 600 matched sim-to-real rollout pairs (6 real-world bimanual tasks ×\times 5 checkpoints ×\times 20 shared initial states) directly against physical robot outcomes (including 332 real failures), measuring strict within-task ranking and selection regret.

  2. Direct comparison on the 600 matched real-robot rollouts. We compare SWAN directly against WorldGym and Ctrl-World under the exact same protocol below (dWorldEval has not publicly released code or model weights):

MethodSR MAE ↓Within-Task Spearman ↑Pairwise Score ↑Selection Regret ↓Balanced Acc. ↑False-Success Rate ↓
WorldGym0.0780.8470.8690.0540.8680.157
Ctrl-World0.0830.8350.8600.0420.8600.169
dWorldEval——————
SWAN (Ours)0.0620.8760.8900.0250.8800.136

SWAN consistently outperforms both WorldGym and Ctrl-World in success-rate calibration (0.0620.062 vs. 0.0790.079 and 0.0830.083 MAE), within-task Spearman correlation (0.8760.876 vs. 0.8470.847 and 0.8350.835), and checkpoint selection regret (0.0250.025 vs. 0.0540.054 and 0.0420.042).


Q2 (W2):

evaluate each proposed component directly on the 600 matched sim-to-real policy evaluation rollouts below:

ConfigurationSR MAE ↓\downarrowWithin-Task Spearman ↑\uparrowPairwise Score ↑\uparrowSelection Regret ↓\downarrowBalanced Acc. ↑\uparrowFalse-Success Rate ↓\downarrow
Full SWAN0.0620.8760.8900.0250.8800.136
w/o Action Conditioning0.1120.6600.7740.0970.7420.259
Recent-Window Memory (vs. Time-Scale)0.0880.7750.8320.0650.8130.196
w/o Scale-Wise Dream Forcing0.0820.8030.8470.0590.8300.180
w/o Visual Alignment (Supervised)0.0760.8310.8620.0510.8480.164

Each component contributes clearly to calibration, ranking, and failure detection: (1) Action conditioning raises within-task Spearman by +0.216+0.216 and lowers the false-success rate on real failures by 12.3712.37 percentage points; (2) Time-scale memory improves Spearman by +0.101+0.101 and reduces false-success rate by 6.006.00 percentage points over a recent-window baseline by preserving distant causal states; (3) Scale-wise dream forcing adds +0.073+0.073 Spearman and reduces false-success rate by 4.444.44 percentage points by bridging autoregressive exposure bias; and (4) On-policy visual alignment further improves Spearman from 0.8310.831 to 0.8760.876, halves selection regret (0.051→0.0250.051 \rightarrow 0.025), and lowers false-success rate by 2.872.87 percentage points without using task-success rewards.

Q3 (W3):

We clarify both the control-specific gains of our action conditioning and the architectural reason for scale alignment in next-scale autoregression:

  1. Control fidelity vs. visual quality. Under identical robotic training data, adding scale-aligned action conditioning changes Image Quality minimally (0.683→0.6970.683 \rightarrow 0.697, +0.014+0.014), whereas Action Following rises from 0.0190.019 to 0.2900.290 (+0.271+0.271) and Trajectory Accuracy rises from 0.4180.418 to 0.5390.539 (+0.121+0.121), confirming that the module specifically drives fine-grained action-future alignment rather than generic frame quality.

  2. Necessity of scale alignment. Standard single-scale conditioning injects uniform-resolution action tokens into diffusion models operating on a fixed spatial grid. Because SWAN generates each video clip coarse-to-fine across 12 spatial scales (s0s_0–s11s_{11}), projecting 3D actions into camera space across matched spatial scales (s0s_0–s7s_7, routed via κ(ℓ)≤ℓ\kappa(\ell) \le \ell) supplies global motion direction at coarse layout scales and local trajectory geometry at fine detail scales.

Thank you to the authors for the detailed rebuttal and additional experiments. I appreciate the efforts to address my concerns, particularly the comparisons with existing policy evaluators and the additional component ablations.

Although SWAN demonstrates consistent improvements over existing world-model-based policy evaluators, the gains in policy ranking and failure detection remain relatively modest. Combined with the limited diversity of evaluated policies and the lack of a direct comparison with conventional single-scale action conditioning, I remain unconvinced that the results sufficiently establish the methodological advantages of the proposed approach.

While I acknowledge the value of the real-robot experiments and the additional evidence, my concerns regarding the methodological novelty and significance of the improvements remain. I will therefore maintain my original rating.

Reviewer_TZ3POpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

SWAN adapts an InfinityStar-initialized next-scale generator for action-conditioned robotic simulation. It couples camera- and scale-aligned action prefixes with age-dependent retention of causal visual states, trains on imperfect histories, and aligns generated dynamics with fidelity rewards under fixed controls. An auxiliary action expert supports physical control. The evaluation covers 60-second streaming, matched executions of frozen external policies in the model and on a robot, and a separate control benchmark.

Critical strengths

The frozen-policy evaluation is the most persuasive contribution. The same five checkpoints run independently in the model and on the robot across six tasks, producing 600 matched episode pairs without real-future feedback in the generated branch. Within-task checkpoint ranking and selection regret address simulator utility more directly than visual quality or a correlation pooled across tasks. The reported mean Spearman correlation of 0.88 and mean selection regret of 0.025 are encouraging. The episode-level counts independently support the reported balanced accuracy of 0.88 (Sec. 4.5; Appendix G).

The paper also distinguishes several easily conflated claims. It separates simulator evaluation from action-expert deployment, control sensitivity from task success, and bounded historical allocation from measured hardware savings. Figure 4 supports the benefit of retaining distant context and training on generated histories, while acknowledging the remaining quality advantage of full-history KV. The residual contact optimism is quantified rather than hidden: 45 of 332 real failures are predicted as successes.

Critical weaknesses

The memory formulation is clearer than its experimental instantiation. Equation 3 and Appendix C define the retention mechanism, but do not give the actual age–scale schedule, retained boundaries, clip dimensions, retention horizon, or resident-byte budget used for Figure 4. These parameters determine how much history the experiment retains and are needed to reproduce its main long-horizon comparison. I accept the explicitly limited storage claim; the absence of hardware measurements mainly leaves the practical tradeoff unresolved.

The policy-ranking evidence is promising but its statistical interpretation needs more detail. There are only five checkpoints per task, evaluated from shared initial configurations. Appendix G reports stratified bootstrap intervals without specifying how this dependence is preserved, or how predicted ties and aggregation weights enter the pairwise score. This matters for interpreting small checkpoint differences and the precision of the reported utility estimates. It does not negate the substantial episode-level agreement.

Questions for authors
  1. What exact memory configuration produced Figure 4, including the schedule and total resident allocation? Providing this would make the streaming result reproducible; a matched-budget comparison would further clarify the design's practical value.
  2. How are shared initial configurations handled in the bootstrap, and how are predicted ties scored? Please provide the task–checkpoint success counts or rates so that the ranking and selection results can be checked.
1 comment
AuthorsOpenReview ↗

Response to Reviewer TZ3P

We thank Reviewer TZ3P for the thorough review and constructive questions. Below, we address Q1 and Q2 in detail.


Q1

We clarify the exact streaming memory configuration used in Figure 4 and its bounded-budget design as follows.

  1. Exact schedule and window parameters. Each camera view operates at 288×512288 \times 512 resolution across 12 spatial scales (s0s_0–s11s_{11}). Each training window spans 161 frames, consisting of 1 initial anchor frame followed by 2 motion clips of 80 frames each. After finishing each motion clip, SWAN caches selected multi-scale visual KV states together with the corresponding action representations under a maximum history length of 2 clips (including the anchor). At steady state during 60-second streaming rollouts (12 consecutive motion clips), each new clip attends to full-scale features up to s11s_{11} from the fixed initial anchor frame and coarse-to-mid features up to s9s_9 from the immediately preceding 80-frame motion clip, while older motion clips are evicted.

  2. Bounded resident allocation and practical tradeoff. Because the per-clip dimensions, retained scale boundaries (s11s_{11} for the anchor and s9s_9 for the preceding clip), and active history length (2 historical entries) are strictly fixed, the resident historical KV memory stays constant across rollout horizons rather than growing linearly as in full-history KV. Under the same two-clip history constraint, allocating part of the resident budget to the anchor state while compressing the recent motion clip to s9s_9 preserves long-term workspace and object states over 60 seconds, avoiding the cumulative drift of a sliding recent-only window (replacing time-scale memory with a recent window reduces downstream policy-ranking Spearman correlation from 0.8760.876 to 0.7750.775).


Q2 Below we provide the complete per-task, per-checkpoint success counts (formatted as Real successes / SWAN predicted successes, each out of 20 matched trials) and detail our ranking, tie-scoring, and bootstrap procedures.

TaskC1C2C3C4C5SWAN PickRegret
T1: Place glasses on shelf10 / 1014 / 1416 / 1512 / 1317 / 14C30.05
T2: Place fruits in basket4 / 47 / 1111 / 1315 / 1413 / 13C40.00
T3: Pour water into cup4 / 87 / 79 / 914 / 1610 / 10C40.00
T4: Thread ring onto rod0 / 02 / 28 / 810 / 88 / 12C50.10
T5: Put items into drawer0 / 05 / 87 / 713 / 1314 / 15C50.00
T6: Put clothes into hamper2 / 47 / 75 / 89 / 1015 / 12C50.00
  1. Pairwise ranking and tie scoring. Comparing 5 checkpoints within each of the 6 tasks yields 6×(52)=606 \times \binom{5}{2} = 60 pairs. One real-side tie on T4 (C3 and C5 both at 8/20) is excluded. Across the 59 valid pairs, 51 are concordant, 5 are discordant, and 3 are ties in SWAN predictions. Weighting all valid pairs equally and assigning 0.50.5 credit to predicted ties yields a pairwise concordance score of (51+3×0.5)/59=0.8898(51 + 3 \times 0.5) / 59 = 0.8898 (0.890.89). Within-task Spearman correlation assigns average ranks to ties within each task and averages equally across the 6 tasks, giving 0.87640.8764 (reported as 0.880.88), with a cell-wise success-rate MAE of 0.06170.0617 (0.0620.062).

  2. Checkpoint selection regret. SWAN selects the true best real-robot checkpoint on 4 of 6 tasks. It selects C3 on T1 (16/20 vs. optimal C5 at 17/20; regret 0.050.05) and C5 on T4 (8/20 vs. optimal C4 at 10/20; regret 0.100.10), giving a mean selection regret of (0.05+0.10)/6=0.025(0.05 + 0.10) / 6 = 0.025. Ties in SWAN's maximum score break toward the earlier checkpoint index, though the predicted maximum is unique on all 6 tasks.

  3. Bootstrap protocol. Appendix G runs 2,000 bootstrap iterations that first resample tasks and then resample 20 episodes with replacement within each (task, checkpoint) cell. Real and SWAN outcomes share identical resampled episode indices within each cell—preserving episode-level sim-to-real pairing—while resampling indices independently across checkpoints.

Comments on the paper

2 comments
Area_Chair_xyeuOpenReview ↗

Discussion

Dear reviewers,

We have mixed signals from the review ratings

  • TZ3P: 3: Weak acceptance
  • 8FsE: 2: Weak rejection
  • fKFE: 3: Weak acceptance

We don't have the response from the authors yet, however, could you take a look at each other's reviews and initiate a discussion to reach consensus? How severe are your concerns and unanswered questions? How do you weigh the strengths and weaknesses?

Please note that ICLR 2027 has Policy on LLM use.

Dear authors,

Please provide your response to the reviews to address the concerns from the reviewers.

Best, Your AC

AuthorsOpenReview ↗

Overall Response

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. We summarize the new results and clarifications from our rebuttal below.

Reviewer 8FsE

  1. Specialized baselines. On 600 matched sim-to-real rollout pairs, SWAN achieved higher within-task Spearman correlation than WorldGym and Ctrl-World (0.876 vs. 0.847 and 0.835) and lower checkpoint-selection regret (0.025 vs. 0.054 and 0.042).
  2. Policy-evaluation ablations. Replacing time-scale memory with a recent window reduced Spearman correlation from 0.876 to 0.775 and raised the false-success rate from 13.6% to 19.6%. Removing dream forcing or visual alignment also worsened both metrics.
  3. Scale alignment and evaluation scope. The ablation supports the value of action conditioning. The specific benefit of scale alignment and generalization across broader policy families remain open.

Reviewer TZ3P

  1. Memory configuration. We specified the Figure 4 retention schedule. SWAN keeps the initial anchor at full scale and the immediately preceding clip through scale s9s_9, keeping the historical KV allocation constant as rollout length increases.
  2. Ranking and uncertainty. We provided success counts for all 30 task–checkpoint combinations. Real-side ties are excluded, and predicted ties receive half credit. The bootstrap preserves real–simulated pairs within each checkpoint and resamples episode indices independently across checkpoints.

Reviewer fKFE

  1. Domain adaptation. With the same backbone and robotic dataset, policy-ranking Spearman correlation rose from 0.660 for the domain-adapted variant to 0.831 for supervised SWAN and 0.876 after visual alignment.
  2. Conditioning mechanism. SWAN projects 3D actions into each camera view and routes them to visual queries by clip, view, and spatial scale.
  3. Single-scale comparison. A matched single-scale baseline is needed to isolate the gain from scale routing.
  4. Relation to prior work. We clarified how SWAN extends frame-aligned conditioning by matching action resolution to each stage of coarse-to-fine generation.

Thank you for considering the additional evidence.

Sincerely,
Authors of Submission #15997

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.