acceptodds
Under review as a conference paper at ICLR 2027

Learning Quality-Controllable Robot Policies with Iterative Hindsight Relabeling

Abstract

Robot policy evaluation often relies on a binary success signal for task completion. In real-world deployment, however, desired execution behaviors are also defined by quality attributes such as execution speed and motion smoothness, which need to be dynamically controlled according to task requirements. To address this, we introduce Quality-Controllable Policy (QCP), a robot policy that enables inference-time control over both individual quality attributes and their combinations. To extend the controllable quality range beyond initial demonstrations, we introduce an iterative hindsight relabeling loop: QCP collects online rollouts under target quality labels, adds successful trajectories to the dataset, relabels the accumulated trajectories in hindsight according to their measured quality, and fine-tunes the policy. Experiments on robomimic and RoboEval demonstrate that QCP consistently improves both task execution quality and success rates across iterations, producing successful rollouts that exceed the quality of the initial demonstrations. Finally, we validate QCP on real-world manipulation tasks, showing that controlling speed increases throughput in conveyor operations and controlling transport stability improves success when carrying wine glasses, showing that quality-controllable policies yield practical gains in physical deployment.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, Imitation Learning, Quality-Controllable Policy

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Cb5TOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The manuscript tackles a limitation of robot foundation models. Standard Behavior Cloning treats all successful demonstrations as equally good, so the resulting policies average over diverse execution qualities such as speed and smoothness and offer no inference-time steerability.

To address this, the authors propose Quality-Controllable Policy. It measures trajectory quality along the desired axes using rule-based scoring functions, discretizes the scores into quantile bins such as 4 bins, and conditions a flow-matching policy on these discrete quality tokens, either via learned embeddings or prompt injection for VLAs such as π0.5. To push execution quality beyond the limits of the initial human demonstrations, the authors further introduce an Iterative Hindsight Relabeling loop. Rollouts are sampled requesting the highest quality bin, successful rollouts are retroactively relabeled by their actual achieved physical metric, added to the dataset, and the policy is retrained on updated quantile boundaries.

The method is evaluated on standard simulation benchmarks including Robomimic and RoboEval, and on physical Franka arm tasks covering wine serving for stability and conveyor picking for speed. It shows consistent joint improvements in task success and execution quality, outperforming quality-filtered BC and offline and online RL baselines.

Critical strengths

Rather than resorting to complex and brittle online reinforcement learning, such as tuning multi-objective reward trade-offs with critics, QCP frames quality control as a supervised conditioning problem. The iterative hindsight relabeling loop elegantly expands the policy's quality boundary beyond human demonstration limits by leveraging policy stochasticity and adaptive quantiles. The paper provides exhaustive empirical evidence, including controlling up to six attributes simultaneously, clear comparisons against competitive RL baselines such as DSRL, EXPO and SCIQL, and crucial ablation studies in Fig. 8 and Fig. 9. The finding that discrete quantile bins outperform continuous targets because of density support in the training distribution is an important takeaway for the robot learning community. Demonstrating QCP on a physical Franka robot via $\pi$0.5 fine-tuning for two distinct physical requirements, namely reducing tilt in wine serving by 19.7 percent and boosting conveyor throughput by 23 percent beyond human teleoperator speeds, validates the practical utility in real-world deployment.

Critical weaknesses

The method's premise assumes access to an analytical scoring function mτ, such as jerk, execution steps or tilt angle. While suitable for kinematic and geometric attributes, it cannot trivially scale to perceptual, semantic or non-analytical qualities such as naturalness, safety near humans or delicateness of contact without introducing human-in-the-loop annotations or learned reward models. In Algorithm 1, Line 9, the policy πθk+1 is retrained from scratch, or reset to the pretrained checkpoint, at each iteration to avoid stale label-to-behavior mappings. As iterations K progress and accumulated data grows, the cumulative training cost becomes substantial, which may hinder rapid online deployment or continuous lifelong learning. In Footnote 3, the authors mention that incorporating failed rollouts as low-quality data degraded performance. In standard Hindsight Experience Replay and offline RL, failed trajectories often provide valuable negative signals or boundary information. Completely discarding them represents a missed sample-efficiency opportunity that warrants further theoretical and empirical clarification.

Questions for authors

Could the authors elaborate on whether warm-starting or partial fine-tuning, for example only updating the quality embedding layer and final projection layers, was attempted instead of full reinitialization from scratch. What specifically causes policy collapse if the previous checkpoint is continually trained on the updated quantile boundaries. How does the policy behave if an out-of-distribution quality request is provided, for example extrapolating beyond q equals B minus 1, or requesting contradictory multi-attribute combinations such as maximum speed combined with near-zero jerk. Does the policy fail gracefully or exhibit erratic kinematics. Could you provide more intuition on why including failed rollouts as q equals 0 degraded performance. Was it due to the policy learning to replicate task-failure modes when lower quality bins were conditioned, or did it degrade the feature representation overall.

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer Cb5T (1/3)

 

We deeply appreciate your insightful comments and efforts in reviewing our manuscript. We respond to each of your comments one-by-one in what follows.

 

1. Extending QCP Beyond Analytical Metrics

We thank the reviewer for raising this important point. While our experiments use analytical metrics such as jerk and execution steps, we expect that QCP could also be used with non-analytical metrics. The framework only requires a scalar quality score for each trajectory, so any scoring function can be used, including learned ones. For qualities that are difficult to measure analytically, such as naturalness, safety near humans, or delicacy of contact, the score can come from a reward model learned from human preferences, as in PEBBLE [1], or from VLM feedback, as in RL-VLM-F [2].

We have not yet evaluated this extension, but we believe it makes QCP applicable to a much broader range of qualities. We added a brief discussion of using learned reward models to score trajectories in Section 6 (Discussion).

 


 

2. Warm-starting and training cost

Following your valuable suggestion, we compared short warm-start updates with longer runs from a fresh initialization. Warm starts may be a practical option when training cost is a concern, while fresh initialization may help push the performance frontier for specific quality metrics (e.g., execution speed).

On Robomimic Square (MH), we compared fresh initialization at each iteration (100K updates) with warm-starting from the previous checkpoint (5K, 10K, or 100K updates). For each setting, we trained separate policies for speed and smoothness over six iterations with three seeds. Appendix J of the revised draft now includes the results, checkpoint-wise learning curves, and diagnostics.

Policy trained forMetricOriginal QCP (Fresh 100K)Warm 100KWarm 10KWarm 5K
SpeedTask success rate (%) ↑82.2 ± 0.475.9 ± 4.364.2 ± 7.363.0 ± 6.9
SpeedSteps to success ↓120.3 ± 2.1163.9 ± 5.8163.7 ± 7.1164.1 ± 2.8
SmoothnessTask success rate (%) ↑79.7 ± 2.377.8 ± 4.466.8 ± 9.263.3 ± 9.6
SmoothnessAction variation ↓0.00849 ± 0.000510.01030 ± 0.000780.01143 ± 0.000520.01177 ± 0.00037

At a matched training budget, the warm-started policy approached the original QCP in smoothness but failed to match its execution speed. Future work could explore more efficient adaptation methods for QCP.

AuthorsOpenReview ↗

Author Response to Reviewer Cb5T (2/3)

 

3. Learning from failed rollouts

Thank you for this thoughtful question. We agree that failed trajectories can carry useful information. We would like to share our intuition on why simply including them as q = 0 degraded performance in QCP.

Conflating task outcome with execution quality. Assigning every failure to q = 0 mixes two different notions: whether the task succeeded, and how well the motion was executed. For example, a rollout may fail the task while moving very fast. In terms of speed, this trajectory should belong to a high-quality bin, but labeling it q = 0 tells the model that it represents the lowest quality. Such inconsistencies may make it harder for the model to learn what each quality level represents, which could weaken the conditioning signal. Excluding failures may therefore provide cleaner quality supervision.

Negative transfer from imitation. As the reviewer suggests, the policy may also learn to replicate failure modes. Because QCP is built on behavioral cloning, actions from failed rollouts remain imitation targets even under the q = 0 label. The label gives no explicit signal to avoid those actions; it may only increase their likelihood under a particular condition. This differs from HER or offline RL, where failed trajectories are used through rewards or value estimates that can explicitly assign them low credit.

Experiments. Consistent with this intuition, on Square-MH, including failed rollouts as q = 0 reduced task success from 81.7% to 75.0% in the speed experiment and from 63.3% to 61.3% in the smoothness experiment, compared with training on successful rollouts alone. To address the first issue, we also tested conditioning separately on success and quality, which separates task outcome from execution quality. This did not consistently improve task success over training on successful rollouts alone. This suggests that the second issue, imitating failed actions under BC, remains even when the labels are disentangled.

We therefore view the effective use of failed rollouts as a promising direction for future work, for example through outcome-based weighting or value-based credit assignment that can explicitly down-weight failed actions instead of imitating them. We discuss these results in Section 6 (Discussion) of the revised draft.

AuthorsOpenReview ↗

Author Response to Reviewer Cb5T (3/3)

 

4. Out-of-distribution quality requests

Thank you for raising this interesting question. In QCP’s discrete formulation, any quality score above the highest threshold maps to the top bin (q=B−1q = B - 1), so higher scores all correspond to the same quality level. To explore how the policy responds to requests beyond the training range, we ran an experiment with continuous quality conditioning.

On Robomimic Square (MH), we evaluated the continuous variant using three training seeds and 300 episodes per request.

Metricq=1.2q=1.1q=1q=0q=-0.1q=-0.2
Speed SR (%)0.00.011.3 ± 1.418.3 ± 2.02.7 ± 0.00.0
Speed execution steps ↓––208.6 ± 8.1298.5 ± 10.2307.0 ± 8.7–
Smoothness SR (%)0.00.049.0 ± 3.520.7 ± 1.40.00.0
Smoothness action jerk ↓––0.02754 ± 0.003920.47062 ± 0.02103––

 

Extrapolation on either side substantially reduced SR and did not reliably produce the requested behavior. These results support QCP’s iterative approach, which can reliably target progressively higher-quality behavior.

For multiple attributes, individually valid requests can form an unsupported or infeasible combination. We have not explicitly addressed such requests. Future work could use attribute weights to select target combinations supported by the data, as in multi-objective Decision Transformers [3].

 

References

[1] Lee et al. PEBBLE: Feedback-Efficient Interactive Reinforcement Learning via Relabeling Experience and Unsupervised Pre-training. ICML 2021.

[2] Wang et al. RL-VLM-F: Reinforcement Learning from Vision Language Foundation Model Feedback. ICML 2024.

[3] Zhu et al. Scaling Pareto-Efficient Decision Making via Offline Multi-Objective RL. ICLR 2023.

Reviewer_a9QyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper aims to make robot execution quality controllable and improve it beyond the range covered by demonstrations. The proposed method conditions a robot policy on discrete quality labels, allowing users to adjust attributes such as speed and smoothness. It then iteratively collects successful rollouts, relabels all accumulated trajectories by their measured quality, and retrains the policy. Experiments on Robomimic and RoboEval show improvements in both execution quality and task success, with quality exceeding the initial demonstrations and support for controlling multiple attributes together. Real robot experiments show increased throughput in conveyor operations and higher success when transporting wine glasses.

Critical strengths
  • The paper is well motivated. Learning quality-controllable policies is a practically relevant problem, and the authors propose a conceptually simple and effective method.
  • The simulation experiments cover a range of settings, and the results on combinations of multiple attributes are interesting.
Critical weaknesses
  • Several experiments show a tradeoff between success rate and execution quality (e.g., in figure 9, more bins improve final execution speed but limit gains in success rate over iterations), but it is unclear how to balance these objectives. In settings where scaling up demonstrations can already yield high task success, a key remaining challenge is to improve execution quality while preserving that success rate. Therefore, I think another informative evaluation can be to start from a policy with a high success rate (e.g., >90%) and compare QCP across hyperparameter settings and against other baselines on quality improvement while maintaining this success rate.
  • The real-world experiments use only BC_filter as an iterative baseline. In simulation, this baseline often achieves high quality but has among the lowest success rates, which makes the comparison insufficient.
Questions for authors
  • How to exploit high-quality segments within otherwise low-quality trajectories is an important question, especially for longer-horizon tasks. Although Appendix G shows that the naive segment-level variant underperform trajectory-level labeling, do the authors see promising directions for incorporating local quality information to improve QCP?
  • How are the interaction budgets matched between QCP and the RL baselines in Table 1?
6 comments
AuthorsOpenReview ↗

Author Response to Reviewer a9Qy (1/3)

 

We deeply appreciate your insightful comments and efforts in reviewing our manuscript. We respond to each of your comments one-by-one in what follows.

 

1. Improving Quality While Preserving High Success

We thank the reviewer for this constructive suggestion. We would first like to clarify that the trade-off in Figure 9 is specific to the ablation on the number of quality bins. With our default setting, execution quality and task success improve together across our experiments (Figures 2 and 7), so this trade-off is not a general property of QCP.

Nevertheless, we agree that the high-success regime is an important and practical setting. Following your suggestion, we evaluated QCP on Robomimic Lift (MH), starting from a policy with a success rate of 92.3 ± 2.0%. We compared QCP at various bin counts with quality-filtered BC in separate experiments targeting speed and smoothness.

 

Speed (after 6 iter., 3 seed avg.)

MetricInitial policyBC_filter baseline2 bins4 bins8 bins16 bins
SR (%) ↑92.3 ± 2.095.0 ± 1.9 (+2.7)98.1 ± 0.7 (+5.8)99.6 ± 0.2 (+7.2)97.8 ± 0.8 (+5.4)95.6 ± 3.7 (+3.2)
Steps ↓107.1 ± 4.045.5 ± 3.0 (−61.6)61.2 ± 1.9 (−45.8)45.0 ± 1.0 (−62.0)42.1 ± 1.3 (−65.0)41.6 ± 4.4 (−65.4)

Smoothness (after 6 iter., 3 seed avg.)

MetricInitial policyBC_filter baseline2 bins4 bins8 bins16 bins
SR (%) ↑92.3 ± 2.093.9 ± 3.2 (+1.6)97.2 ± 0.8 (+4.9)98.3 ± 0.7 (+6.0)97.4 ± 1.3 (+5.1)88.4 ± 6.2 (−3.9)
Normalized jerk (×10⁻³) ↓132.9 ± 10.717.4 ± 1.1 (−115.5)18.4 ± 0.6 (−114.5)14.8 ± 0.2 (−118.0)13.3 ± 0.3 (−119.6)12.3 ± 0.8 (−120.6)

 

Even starting from a high-success policy, QCP substantially improves execution quality without sacrificing success. Only at the extreme setting of 16 bins does the success gain diminish or reverse. Consistent with the discussion of Figure 9, too many bins make conditioning closer to continuous labels and leave fewer training samples per bin.

AuthorsOpenReview ↗

Author Response to Reviewer a9Qy (2/3)

 

2. Strengthening Real-World Baselines

We thank the reviewer for this valuable suggestion. To strengthen the real-world comparison, we added EXPO as an RL baseline on the conveyor-belt task. We chose EXPO because it was the strongest baseline in our simulation experiments targeting speed, improving both success rate and the quality metric (Table 1).

MetricMethodIteration 0Iteration 1Iteration 2
Throughput (cans/min) ↑EXPO9.0 ± 0.39.3 ± 0.510.1 ± 0.1
QCP (ours)9.3 ± 0.310.6 ± 0.613.0 ± 0.5
EEF speed (mm/s) ↑EXPO180.7 ± 6.8202.5 ± 8.2195.3 ± 3.5
QCP (ours)208.0 ± 7.0241.2 ± 13.3257.6 ± 8.8

Even against this stronger baseline, QCP achieved substantially larger gains over two iterations, improving throughput by +3.7 cans/min (vs. +1.1 for EXPO) and end-effector speed by +49.6 mm/s (vs. +14.6 for EXPO).

AuthorsOpenReview ↗

Author Response to Reviewer a9Qy (3/3)

 

3. Incorporating Local Quality through Semantic Trajectory Segmentation

We thank the reviewer for this insightful question. We agree that exploiting high-quality segments within otherwise low-quality trajectories is a promising direction, especially for longer-horizon tasks. To explore this, we ran a small pilot study that assigns quality labels to semantically meaningful segments rather than arbitrary ones, and we are happy to share the encouraging results.

On the Robomimic Transport task, we divided each trajectory into four semantic phases: reaching and grasping, transport by the first arm, handoff, and placement by the second arm. Phase boundaries were reviewed using Qwen3-VL-4B-Instruct. We then assigned smoothness labels at the phase level instead of the trajectory level.

In this pilot, phase-level labeling achieved a success rate of 35.3 ± 2.0%, compared with 12.3 ± 3.3% for trajectory-level labeling in our main results (+23.0%). Smoothness also improved, with the metric decreasing from 0.02225 ± 0.00804 to 0.01068 ± 0.00493 (a 52.0% reduction).

These preliminary results suggest that local quality information can benefit QCP when segments reflect the semantic structure of the task. A more systematic study of how to identify effective segments, and of how to combine local and trajectory-level labels, is an exciting direction for future work. We have added this discussion to Section 6 (Discussion) of the revised draft.

 

4. Matching Interaction Budgets

For each task, we matched the RL baselines’ interaction budget to QCP’s maximum budget using six rounds of 300 rollouts and the task-specific episode limit. This gave 0.72M environment steps per task for Robomimic Square and RoboEval (400 steps per rollout) and 1.44M for Robomimic Transport (800 steps per rollout).

Reviewer_L5ULOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses improving robot behavior beyond task success rates by considering additional quality metrics, such as speed and smoothness. The proposed approach consists of two steps. First, demonstration trajectories are scored by pre-defined rules and divided into quality bins to train a quality-conditioned policy, allowing users to specify the desired execution quality at inference time. Second, an iterative hindsight relabeling procedure collects rollouts under the highest-quality label. This enables the policy to explore more high-quality behaviors beyond initial demonstrations. Experiments on Robomimic, RoboEval, and real-world tasks show improvements in both execution quality and task success.

Critical strengths
  • The paper addresses execution quality beyond task success, including speed and smoothness. Quality labels are derived from automatically computed metrics, avoiding additional human preference annotations. The method is straightforward to implement and can be easily incorporated into existing policies.

  • Experiments on Robomimic, RoboEval, and physical manipulation tasks demonstrate improvements in both execution quality and task success. The results also show successful behaviors extending beyond the quality range of the initial demonstrations.

  • Comparisons with accumulated-data BC, quality-filtered BC, and RL methods help contextualize the results. Analyses of quality distributions, quantization levels, and multi-attribute conditioning provide further insight into the method’s behavior. Reporting task success alongside execution quality is particularly valuable.

Critical weaknesses
  • The core procedure iteratively relabeling and retraining closely follows established approaches such as Quark. Although demonstrating its effectiveness in robotic manipulation has practical value, the methodological contribution currently appears to lie primarily in transferring this recipe to a new domain. The paper should more clearly establish what robot-specific challenges require new algorithmic designs or yield insights beyond those already available from prior work.

  • The method assumes that relevant quality metrics are specified in advance, but does not sufficiently examine how these metrics should be selected or how their importance varies across tasks and execution stages. Faster or smoother behavior is not uniformly preferable, particularly when it conflicts with precision, safety, or task completion. Updating quantile boundaries adapts the requested quality range, but does not resolve whether the underlying metrics capture the behaviors that actually matter. Unlike reward-driven optimization, which can assign credit according to downstream task outcomes, conditioning on preselected trajectory-level attributes does not by itself establish their task relevance. This may constrain the method’s achievable benefits when the chosen metrics are incomplete or misaligned. Analysis of attribute selection, conflicting objectives, and task-dependent relevance would therefore strengthen the paper.

Questions for authors

Please see weakness.

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer L5UL

 

We deeply appreciate your insightful comments and efforts in reviewing our manuscript. We respond to each of your comments one-by-one in what follows.

 

1. Differences from Quark and insights for robot learning

Like Quark [1], our method conditions the policy on quantized quality levels, but applying this idea to robot learning required several changes. First, on the algorithm side, we re-initialize the policy across iterations. Without this, the policy tends to keep fitting to trajectories from earlier rounds and collapses toward a narrow range of behaviors (see Appendix H, added in the revised draft). Second, Quark optimizes a single scalar reward. Robot execution quality has several dimensions, so we study conditioning on robot-specific metrics such as speed and smoothness, and analyze how each relates to task success. Third, Quark adds new special tokens for quality levels. We instead write quality levels as natural-language text in the VLA’s instruction, so the pretrained VLM backbone’s language prior (e.g., what “fast” or “smooth” means) carries over to the conditioning signal. Together, these changes make quality conditioning effective for robot policies rather than a direct application of Quark.

 


 

2. Quality Metric Selection and Task Relevance

We thank the reviewer for this insightful question. We agree that the relevance of a quality metric depends on the task, and that choosing a metric is an important practical consideration.

A simple selection criterion. To select quality metrics in a task-relevant way, we propose a simple screening procedure that uses data available before iterative retraining begins. Concretely, we collect a set of initial rollouts, then measure the correlation between each candidate metric and task success across those rollouts. We observe that metrics with a stronger correlation with task success are better candidates for quality conditioning. The initial trials can be rollouts of the initial BC policy, as in our experiments below.

Empirical support. In Robomimic Square, we tested whether stronger initial correlations predicted larger success-rate gains after iterative retraining.

Quality metricSmoothnessEEF jerkPath lengthRotation pathPauseGrasp timingSpeed (Duration)
Absolute Spearman ρ with initial rollout success0.118 ± 0.1500.290 ± 0.0650.101 ± 0.0390.182 ± 0.0620.141 ± 0.1250.037 ± 0.051– (termination bias)
Success-rate gain (pp, iterations 1→10)+31.6 ± 3.9+41.3 ± 3.8+32.6 ± 4.0+33.7 ± 8.5+30.2 ± 3.2+18.0 ± 2.7+34.6 ± 7.9

Across the six metrics, a stronger initial correlation went with a larger final gain (Pearson r = 0.90). EEF jerk had the highest correlation and produced the largest gain (+41.3 pp). Grasp timing had the lowest correlation and produced the smallest gain (+18.0 pp). This suggests that a small set of initial trials is enough to find useful metrics for a new task, with no need for costly trial and error over full retraining runs. We have added this analysis and the selection procedure to Appendix I of the revised draft.

 

References

[1] Ximing Lu et al. Quark: Controllable Text Generation with Reinforced Unlearning. NeurIPS, 2022.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer L5UL,

Thank you again for your thoughtful review and discussion.

Your questions and feedback have greatly helped us clarify our arguments and strengthen the paper.

As the discussion period draws to a close, we wanted to follow up and see whether there are any remaining concerns we could help address.

We would be happy to continue the discussion!

Many thanks,
The Authors

Thank you for your response. My concern has been addressed. I will keep my score.

AuthorsOpenReview ↗

Response to Reviewer L5UL

Dear Reviewer L5UL,

Thank you again for your thoughtful feedback. We are glad that our rebuttal addressed your concerns, and we appreciate your continued support for our work. Please feel free to let us know if you have any further questions or suggestions.

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response

We sincerely thank the reviewers for their time and thoughtful feedback. Below, we address both common and reviewer-specific concerns.

 

1. Expanded analysis and discussion.

We added Section 6 (Discussion) to address four topics raised by the reviewers:

  • Selecting Quality Metrics to Improve Task Success Rate (L5UL), with supporting analysis in Appendix I.
  • Learning from Failed Rollouts (Cb5T).
  • Incorporating Segment-Level Quality (a9Qy).
  • Extending QCP Beyond Analytical Metrics (Cb5T).

We also added an analysis of policy reinitialization and warm-starting in Appendix H (Cb5T).

 

2. Strengthened real-world evaluation.

We repeated the real-world experiments across three independent training seeds, each covering the full iterative procedure. Figure 7 has been updated to report these results.

 

3. Attached a supplementary video.

We have attached the Supplementary Video to show the differences in execution quality.

 

The revised text is highlighted in blue\textcolor{blue}{\text{blue}} in the updated manuscript for ease of review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.