acceptodds
Under review as a conference paper at ICLR 2027

BRM: Better Reward Modeling via Critic Feature Transfer in Continuous Control

Abstract

Preference-based Reinforcement Learning (PbRL) replaces hand-crafted reward functions by learning a reward model from pairwise trajectory preferences, achieving good performance in several continuous control tasks. However, the generalization of the reward model quickly becomes a bottleneck of the pipeline: it has to predict rewards for the millions of visited state-action pairs from a few hundred preference trajectories. This data asymmetry results in the reward model overfitting to the labeled data, ultimately degrading final performance. In this work, we address this issue by turning this data asymmetry into an opportunity. We propose BRM (Better Reward Modeling), which copies the features of the critic, trained on the full replay buffer, and uses them as an input to the reward model, trained on only a few hundred labeled preferences. Across twelve tasks from the DeepMind Control Suite and MetaWorld and against seven PbRL baselines, BRM reaches a task-normalized Interquartile Mean (IQM) of 0.85 versus 0.55 for the strongest of them. Our empirical analysis indicates that BRM's gains are consistent with better reward generalization, most clearly on trajectories held out from the preference distribution. Our results suggest that the critic the agent already trains is an underused starting point for the reward model in existing PbRL methods.

Primary area:
reinforcement learning
Keywords:
Deep Reinforcement Learning, Preference-based reinforcement Learning, Continuous Control

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_mK2cOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a new approach to reward in the context of PbRL. The proposed method, BRM, uses the first two layers of the critic in a soft-actor critic (SAC) policy as input features to the reward model learned from preference labels over pairs of trajectories. In the experiments section, BRM is compared against several PbRL baselines and the various design decisions are evaluated to understand their value. There results suggest including features from the critic improve the performance of both the policy trained on the learned reward and the learned reward on the preference labeling task.

Critical strengths

The paper is well grounded in its writing and narrative arc (e.g. how the work is motivated and how readers are introduced to the problem). The experiments run dig deeper into what about including critic layers as input improves reward model performance.

Critical weaknesses
  • There are key experiments results missing from the experiments section to truly understand the impact of the proposed method and to appropriately contextualize the key learning:
  1. The most important one is, there is no comparison to REED, which the authors call out as the most related of prior work to what they propose. The lack of a comparison to REED leaves the ablation in Section 5.4 rather weak. The conclusion from Section 5.4 is that extra capacity in the form of extra reward model parameters, random features, or actor features lead to similar gains as the critic features. However, evaluating the benefits of a learned representation via something like a reconstruction loss is not explored. Therefore, it is not clear if the gain is from having a representation that is learned from a somewhat related objective (i.e. the critic) or having a representation learned with some objective. The temporal consistency objective from REED would have been a reasonable comparison point to help fill out this ablation and understand the benefits of critic features.
  2. There are no experiments to understand if the critic features are beneficial simply because they are exposed to more data than the reward model is, or because the something about the critic objective is beneficial for learning more generalizable features. A comparison to REED would have helped here or an experiment where the number of preference features or the frequency of soliciting preference feedback is increased to see how much preference data is needed to match the performance of BRM.
  3. SAC performance on each of the tasks is not provided, so it is not clear how far BRM moves towards recovering performance achieved from the ground truth reward function.
  4. The extent to which BRM improves the reward model on truly OOD samples is not clear as the reward model is only evaluated on samples from SAC's replay buffer, which is where the preference samples the reward model is trained on are also sampled from. It is not possible to tell how similar the samples used in the "Does the reward model generalize beyond the preference distribution?" experiments are to the samples the reward model is trained on. This makes it difficult to know how much the critic features help with generalization.
  5. The authors claim that the skip connection allows the reward model to selectively pick to use the critic features based on whether they are mature enough. However, there are no experiments to test if this is actually the case. A simple experiment to look at this would have been to look at how much influence the critic features have at different training phases, e.g. m's.
  • Not a critical weakness per se, but something that would be beneficial to address. There are awkward and unconventional turns of phrase, word usages, and sentence construction. The authors include using a LLM to help with paper refinement, so I am guessing these come from the LLM. However, it would help the reader to clean these up. Some examples, (1) over use of ":" that seem to have been swapped in where LLMs are prone to use m-dashes and (2) in "...the reward model is refit..." and "...the reward model is refit...", the word "refit" doesn't really have a meaning,
Questions for authors
  • Why does combining BRM with other PbRL methods drop performance relative to BRM alone?
  • What parts of PEBBLE are not used by BRM? You claim PEBBLE is the base algorithm, but then have an experiment that is PEBBLE+BRM, which is confusing.
  • In PEBBLE the reward model quickly achieved 100% accuracy on the preference data and the authors had to hardcode a 97% accuracy training cut off. Does the BRM reward model exhibit similar behaviors?
  • What is meant by "dual-critic ablation" on line 279? It would be helpful to have a little more detail so the reader can better understand the claim, "...no significant effect of decoupling the two."
  • It seems odd that later layer critic features correlate more with the reward, but are worse as reward features. What is your intuition behind this tension?
4 comments

We thank the reviewer for their detailed comments and concrete suggestions, which we address below.

W1 (REED comparison): We thank the reviewer for this important point. We agree that REED is an important comparison for this work. Per-task results on the 11 tasks used for the IQMs are in our answer to Reviewer quTf.

MethodIQM [95% CI]vs. QPA
BRM0.861 [0.819, 0.893]+49%
REED0.677 [0.616, 0.733]+17%
QPA0.579 [0.510, 0.642]

In aggregate, BRM reaches an IQM of 0.86 against 0.68 for REED, and the two 95% confidence intervals do not overlap. REED improves over QPA, so we believe a learned representation already helps the reward model. BRM remains clearly ahead of REED, which suggests that the critic's objective brings more than the self-predictive representation objective used by REED. We will add REED to Table 1.

W2 (more data vs. critic objective): We thank the reviewer for this constructive suggestion. REED gives a direct way to separate the two, since its representation is trained on the same replay buffer as the critic but with a different objective. As it remains clearly below BRM (W1), we believe the amount of data alone does not explain the gain.

W3 (SAC with ground-truth reward): We thank the reviewer for this suggestion. We will add SAC trained on the true reward as a reference line to the learning curves.

W4 (truly OOD evaluation): We thank the reviewer for this point. The SAC set does not come from the agent being trained, but from a separate SAC agent trained on the true reward, whose transitions the reward model never sees. Both methods use the same SAC buffer for each seed. The shift is visible in the accuracies in Table 3.

W5 (Skip connection): The skip connection is meant to let the reward model rely on the raw input when the features are less useful, and the design ablation suggests that it helps (IQM 0.79 without it, 0.87 with it, Table 9), though we agree this experiment does not isolate when each input is used. We thank the reviewer for this suggestion.

Q1 (combining BRM with other PbRL methods): We thank the reviewer for this question. BRM is not a standalone algorithm, as it changes the reward model's input and runs on top of a PbRL method. In Table 1, BRM runs on QPA, so "BRM alone" is QPA+BRM in Table 2. PEBBLE+BRM and RUNE+BRM are the same change on the PEBBLE and RUNE backbones, and each improves over its own base (six of seven and seven of seven tasks). The differences between the three combined versions follow the backbone rather than BRM, as QPA+BRM is higher on most DMC tasks, while RUNE+BRM is higher on the three MetaWorld tasks and walker-run.

Q2 (what parts of PEBBLE BRM uses): We thank the reviewer for pointing out this confusion. As explained in Q1, BRM's default backbone is QPA (Table 1). QPA and RUNE are themselves built on PEBBLE, so PEBBLE+BRM applies the same change directly to PEBBLE. We will clarify this in Section 5.1.

Q3 (reward model reaching 100% training accuracy): BRM keeps the same reward-training rule as its backbone, including PEBBLE's accuracy cutoff, so its reward model stops training at the same point as the baselines.

Q4 (dual-critic ablation): We thank the reviewer for asking for more detail, and we agree that our phrasing could be improved for clarity. Sharing features with the reward can also improve the Q-function, as in DeepMDP (Gelada et al., 2019), so the gain could come from a better critic rather than a better reward model. BRM-DC (Appendix D.2) separates the two by adding a critic whose only goal is to provide features for the reward model, but does not feed SAC's own critic and policy. This does not change performance (IQM 0.86 vs. 0.87 for BRM), which suggests the gain comes mainly from the reward model. The feature critic is still trained on the reward model's output, so "decoupling the two" refers only to SAC's critic. We will rephrase this in Section 4.2 and Appendix B.2.

Q5 (later layers more correlated, worse as features): We thank the reviewer for this question, and first clarify that BRM uses only the first layer of the critic. In Figure 7, the second layer often appears more correlated with the true reward, but its correlation rises and dips during feedback on door-unlock and drawer-open. In PbRL, such a dip can be critical, since a worse reward model can break the policy, and new queries then come from worse behavior, which makes it hard to recover. The critic's targets also change with each new reward model, and the second layer, closer to the output, follows these changes more than the first. The first layer then gives a more stable input across rounds, in line with Yosinski et al. (2014), where earlier features are more general and transfer better.


We look forward to hearing from you and providing any other clarifications during the discussion period. Thank you!

Thank you for your timely feedback. Please update your PDF to incorporate the information you have provided in this rebuttal (highlighted in a color of your choice). Additionally, please review the overall writing quality and clean up awkward and unclear constructions.

Let me know when you have made these changes, and then I will reconsider my score.

We thank the reviewer for their reply and for their constructive feedback throughout the discussion period.

We have updated the PDF, with the changes discussed in the rebuttal highlighted in red, including the REED comparison, SAC trained with the true reward as a reference, and the clarifications on the evaluation sets and the dual-critic ablation. We also revised parts of the text for clarity, including the examples the reviewer pointed out.

We will add a more detailed global comment listing the changes at the end of the discussion phase. We hope the revised version addresses the reviewer's concerns, and we would be grateful if they could reconsider their score. Thank you!

Thank you for your responses and updated draft. I am changing my score from 2 -> 3

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

BRM uses features from an existing SAC critic to improve a preference-based reward model. It concatenates a frozen copy of the critic's first-layer state-action features with the raw state and action. At each feedback round, it refreshes those features, reinitializes the reward heads, and trains them on accumulated preferences. The idea is to draw on the critic's access to the full replay buffer when labeled comparisons are scarce. Across twelve continuous-control tasks, the reported task-normalized IQM is 0.85, compared with 0.55 for QPA; additional experiments also improve PEBBLE and RUNE. Analyses of feature source, capacity, depth, resets, label noise, and held-out preference prediction explore where the gains come from and when transfer helps.

Critical strengths

The main attraction is that BRM gets substantial gains from a simple change to an existing pipeline. It reuses a representation the agent already learns, without additional environment interaction or an auxiliary representation-learning objective. The twelve-task, ten-seed comparison and improvements across three backbones give the empirical result breadth. These results carry most of the weight in my weak-accept recommendation.

The feature-source and capacity controls also make a useful case that the representation matters. In particular, the held-out SAC-buffer preference test provides evidence beyond return curves: door-unlock accuracy rises from 0.41 to 0.72. Reporting tasks with little improvement and the weaker transfer from deeper critic features helps establish the limits of the approach.

Critical weaknesses

The missing REED comparison is the largest gap in the evaluation. The related-work section identifies it as the closest alternative for learning reward representations from unlabeled transition structure, yet it is not among the seven baselines. Random features, actor features, and larger raw-input networks are useful controls, but they do not establish BRM's quality/cost advantage over a representation learned specifically for the reward model. That comparison matters particularly for such a simple proposal.

I am less persuaded by some of the mechanistic explanations than by the performance results. B.2 describes the dual-critic experiment as breaking the critic–reward loop, but D.2 says the feature-producing critic still learns from the reward model's output; only the actor's critic is decoupled. The reward-to-feature-critic-to-reward loop therefore remains. A non-significant difference at ten seeds, with the reported wide errors, also does not establish that feedback is harmless. Similarly, the raw-input skip path permits a fallback in principle but does not demonstrate that training chooses it when critic features are poor. Narrowing these explanations would address the concern without undermining the main empirical result.

Questions for authors
  1. Can REED and BRM be compared under matched feedback, interaction, and reward-training budgets? If that is not feasible, please explain the specific obstacle. This is the comparison most likely to affect my assessment of the representation-transfer contribution.

  2. Does the feature-producing critic in BRM-DC still receive rewards from the head that consumes its features, as D.2 indicates? If so, B.2 should describe the experiment as testing actor–critic feature coupling, rather than removing the reward–critic feedback loop.

  3. Are preference data, evaluation segment pairs, and head-reset schedules matched in the feature-source and held-out-accuracy comparisons? Please include the held-out pair counts and uncertainty for Table 3. Since the policies can produce different full and recent buffers, a shared evaluation set matters for interpreting the results as reward generalization.

2 comments

We thank the reviewer for their thoughtful comments, and for noting the simplicity of the method and the size of its gains. We answer the main points below.

Q1 (REED comparison): We thank the reviewer for raising this important point. REED was initially omitted since it is not among the baselines of the recent PbRL methods we compare to, but we agree that it is an important comparison for this work. We now provide REED results on 11 tasks with 10 seeds, under the same feedback budget, training steps and reward-training budget as the other baselines.*

MethodIQM [95% CI]vs. QPA
BRM0.861 [0.819, 0.893]+49%
REED0.677 [0.616, 0.733]+17%
QPA0.579 [0.510, 0.642]
TaskREEDBRMBest other baseline
walker-walk890 ± 45929 ± 18QPA 751
cheetah-run589 ± 57704 ± 22PPE 600
quadruped-walk297 ± 85838 ± 46PPE 623
quadruped-run355 ± 48640 ± 45PPE 473
walker-run638 ± 11684 ± 62QPA 617
door-unlock91.4 ± 5.294.4 ± 5.4PEBBLE 82.9
button-press56.7 ± 8.591.6 ± 5.2QPA 72.6
door-open64.5 ± 10.869.7 ± 15.2STAR 76.0
window-open95.5 ± 2.887.8 ± 9.8QPA 73.6
sweep-into62.4 ± 12.247.8 ± 13.4RIME 39.3
drawer-open41.0 ± 14.667.1 ± 14.9PEBBLE 71.4

* humanoid-stand is not included since it runs for 2M steps with 10k feedback; all IQMs above are computed on the remaining 11 tasks.

In aggregate, BRM reaches an IQM of 0.86 against 0.68 for REED, and the two 95% confidence intervals do not overlap. In the DeepMind Control Suite, REED remains behind BRM on all five tasks, and sometimes behind the other baselines (cheetah-run, quadruped-walk, quadruped-run). In MetaWorld, results are closer: REED is the best method on window-open and sweep-into, though neither gap with BRM is significant (Welch's t-test, p > 0.4), and remains behind BRM on the other four. BRM is significantly ahead on quadruped-walk, quadruped-run and button-press. Overall, REED is less consistent, with large gaps on some tasks (quadruped-walk, button-press), while BRM reaches these results without training an extra objective as in REED. We will add REED to Table 1 and to the related work discussion in the paper.

Q2 (BRM-DC): We thank the reviewer for this point, and we agree that our phrasing in B.2 could be improved for clarity. As the reviewer points out, the feature critic is still trained on the reward model's output, and D.2 does not aim to separate that part, as it is the essence of BRM. However, the goal of this experiment is to understand whether feature sharing only improves the reward modeling (Table 3) or if sharing features with the reward also improves the Q-function, as shown in DeepMDP (Gelada et al., 2019) and other deep RL methods that add auxiliary reward learning tasks. To separate the two, BRM-DC adds a second critic whose only goal is to provide features for the reward model. This critic is trained in a standard way but does not feed SAC's own critic and policy. This way, the reward model does not share any feature with SAC's critic. Cutting this link does not change performance (IQM 0.86 for BRM-DC vs. 0.87 for BRM), which suggests the gain comes mainly from the reward model. We will rephrase B.2 to match and describe the D.2 setup more clearly in the revised appendix.

Q3 (matched evaluation in Table 3): We thank the reviewer for this question. Each method gets the same feedback budget and query schedule, but its labels come from its own policy, as is standard in PbRL evaluation. Matching the labels across methods would remove the effect of the reward model on what gets queried. For head reset, QPA trains as published, and BRM's heads are reset at every feedback round. Each evaluation uses 200 random pairs of 50-step segments per seed over 5 seeds.

The full and recent sets come from each run's own buffer. They ask whether a reward model trained on a limited number of labels is accurate on the rest of the buffer its agent trains on, which is already a generalization question. The SAC set is the shared one. Here, SAC refers to a separate agent trained on the true reward. Both methods (QPA and BRM) use this SAC buffer for each seed, with pairs drawn independently, and it is the set that supports the distribution-shift claim. We will add the standard errors and the pair counts and this distinction to the revised Table 3 and Section 5.4.

SetMethodwalker-walkcheetah-rundoor-unlockdrawer-open
SACQPA0.68 ± 0.090.93 ± 0.010.41 ± 0.050.65 ± 0.03
SACQPA + BRM0.78 ± 0.040.95 ± 0.010.72 ± 0.040.62 ± 0.07

We look forward to hearing from you and providing any other clarifications during the discussion period. Thank you!

Thank you for the additional experiments and clarifications. The REED comparison addresses my main evaluation concern, and the Table 3 details answer my questions about sample counts and uncertainty. I would keep the mechanistic conclusions narrow because a non-significant difference does not rule out feedback effects. I would be happy to support your work toward acceptance and therefore maintain my recommendation.

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

This paper addresses the data asymmetry problem in preference-based reinforcement learning(PbRL). It proposes BRM, which reuses frozen first-layer features from the SAC critic as extra inputs to the reward model. The method requires no additional environment interaction and can be plugged into existing PbRL pipelines. Experiments across continuous control benchmarks show improved performance and better OOD reward prediction. The paper also conducts extensive ablations and robustness tests.

Critical strengths

The paper includes a rich set of carefully designed control ablations, which effectively rule out confounding factors such as model capacity and random features, strengthening the credibility of its empirical findings.

Critical weaknesses
  1. The novelty of this work is somewhat limited. The core operation is a standard transfer-learning recipe. The skip connection, ensemble heads, and per-round reinitialization are all off-the-shelf components. The paper's real contribution is an empirical finding rather than a method, and the thorough ablations only establish that the trick works, not that it is new.
  2. The main result (0.85 vs. 0.55) is not a single-variable comparison. QPA runs at its published config (K=1, 20× cropping augmentation) while BRM uses K=3 without augmentation, so the gap bundles ensemble size and augmentation changes. The clean comparison (unified K=3, no augmentation) exists only in the appendix, at 4 tasks with 5 seeds.
Questions for authors
  1. The critic features fed into the reward model are frozen during reward model training. Does the quality of these pre-extracted critic features degrade as the SAC policy keeps updating throughout training?
  2. Reward model predictions supervise critic learning, and critic features are used by the reward model. Have you tested whether this loop can lead to catastrophic drift under high label noise?
  3. Why does the second critic layer, which shows higher linear correlation with true reward, yield worse transfer performance for reward modeling? Can you provide further diagnostic evidence?
2 comments

We thank the reviewer for their thoughtful comments, in particular on the ablations and how they support our findings. We address each point below.

W1 (Novelty): We thank the reviewer for this point. We agree that the components are standard, and that transfer learning can be considered a standard tool in ML. We believe however that it is less common in PbRL, where most methods still train the reward model from scratch on the preference labels. BRM shows important gains with a small modification, and we see this simplicity as a strength, a point that reviewer quTf also makes.

Beyond empirical improvements, we believe that another important takeaway message of the paper is that research on feature sharing and transfer in PbRL should not be restricted to the last hidden layer, as is often the case in the RL literature, because of the non-stationary reward property of PbRL which we discuss further in our answer to Q3. We therefore see the novelty less in the components themselves and more in how and where the features are transferred.

W2 (Single-variable comparison): We agree that the main comparison is not single-variable. In the main result, we compare each method with its published configuration. For QPA, this indeed means K=1 (single reward head) with crop augmentation. Figure 3, a task-normalized aggregation of Appendix B.1's results, shows that QPA's reward MLP with K=3, no augmentation, and parameters matched to BRM is worse than the original configuration (IQM 0.43 vs 0.50). Improvements over PEBBLE and RUNE in Table 2, which do not use cropping and use 3 reward heads, are additional evidence that the gain does not come from additional capacity or cropping.

Q1 (Staleness of frozen features): We want to clarify this point. Each copy of the critic features is used only between two feedback rounds. When new feedback is obtained, we refresh these features by taking a copy of the current features from the critic, as shown in Algorithm 1, and retrain the reward heads on them. Between two rounds the reward does not change, so the copied features do not degrade. They can only fall behind the live critic until the next refresh.

Q2 (Feedback loop under label noise): We appreciate the observation about the circular dependency between the critic and the reward model. In the main results, we only see the positive side of it, and the skip connection can provide a safe fallback when the features are worse. The label noise point is also an interesting one. In Appendix D.3, we compare BRM and QPA on 4 environments with 10 and 20% chance of randomly flipping preference labels: while both methods degrade, BRM remains ahead on every cell. While we didn't study an even higher noise level, we note that at 20% noise, QPA's success on drawer-open is already near zero (0.2%).

Q3 (Layer 2 more correlated, worse to transfer): This is a good question. In Figure 7, the second layer of the critic often appears to be more correlated with the true reward. However, its correlation rises and dips (below the correlation of the first layer) during feedback on door-unlock and drawer-open. In PbRL, such a dip can be critical, since a worse reward model can break the policy, and new queries then come from worse behavior, which makes it hard to recover.

A second reason is how the layers change across rounds. The critic's targets change with each new reward model, and the second layer, closer to the output, generally follows these changes more than the first. The first layer then gives the reward model a more stable input across rounds. This goes along with Yosinski et al. (2014), where earlier features are more general and transfer better.


We look forward to hearing from you and providing any other clarifications during the discussion period. Thank you!

Thanks for the additional experiments and detailed explanations. I think they definitely strengthen the empirical part of the paper. However, my concerns about the limited methodological novelty and the mismatched settings in the main headline result still remain. I wanted to update the score to between 2 and 3, but that option isn't available, I had to keep my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal

We thank the reviewers for their constructive and positive feedback. We summarize below the main changes in the revised manuscript and where each reviewer stands.

Key points of agreement

  • Using only a simple change, BRM obtains substantial gains when added to an existing PbRL pipeline, using features that are already learned by the critic (Reviewer quTf).
  • Ablations are carefully designed and rule out confounding factors. (Reviewers pTQj, quTf)
  • The generalization experiment gives evidence beyond return curves (Reviewer quTf), and the paper is clearly introduced and well motivated (Reviewer mK2c).

Addressed concerns

  • REED comparison (Reviewers quTf, mK2c). We added REED as an 8th baseline on all the 12 tasks, including the one that was not in the early rebuttal results (humanoid-stand). It is the strongest baseline with an IQM of 0.63, but remains behind BRM which reaches 0.85 with non overlapping confidence intervals. Reviewer quTf stated that this addition adresses their main concern.
  • Generalization experiment evaluation sets (Reviewers quTf, mK2c). We clarified how the evaluation sets for the generalisation experiment are built and added standard errors and requested pair counts. Reviewer quTf stated that these details answered their questions.
  • Frozen features, label noise and layer depth (Reviewer pTQj). We clarified these points using existing results from the paper (Algorithm 1, Appendix D.3, Figure 7). The reviewer thanked us for the detailed explanations.
  • Dual-critic, skip connection, backbone (Reviewers quTf, mK2c). We clarified these points descriptions in our rebuttal responses and in the paper.

Changes in the revised manuscript (edits are highlighted in red).

  • REED is now added in Table 1, Figure 2 and in the significance tests in the appendix.
  • SAC trained on the true reward as a reference in Figure 2 (Appendix A.4).
  • As mentionned above, clarification of the evaluation sets of the generalization experiment (Section 5.4) and the description of the dual-critic ablation (Section 4.2, Appendix B.2 and D.2), skip connection (Appendix B.2) and BRM's backbone (Section 5.1).
  • Some minor writing changes throughout, as suggested by Reviewer mK2c.

After the updated manuscript, Reviewer mK2c raised their score from 2 to 3, and Reviewer quTf supports acceptance. Reviewer pTQj found that the additional experiments "definitely strengthen the empirical part of the paper", with novelty as their main remaining reservation. Reviewer pTQj also wanted to raise their score but no option was available between 2 and 3.

We thank the reviewers again for their engagement during the discussion, and the Area Chair for considering our work and this rebuttal in the final evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.