Off-Policy Evaluation for Human Feedback

Off-policy evaluation (OPE) is important for closing the gap between offline training and evaluation of reinforcement learning (RL), by estimating performance and/or rank of target (evaluation) policies using offline trajectories only. It can improve the safety and efficiency of data collection and policy testing procedures in situations where online deployments are expensive, such as healthcare. However, existing OPE methods fall short in estimating human feedback (HF) signals, as HF may be conditioned over multiple underlying factors and is only sparsely available; as opposed to the agent-defined environmental rewards (used in policy optimization), which are usually determined over parametric functions or distributions. Consequently, the nature of HF signals makes extrapolating accurate OPE estimations to be challenging. To resolve this, we introduce an OPE for HF (OPEHF) framework that revives existing OPE methods in order to accurately evaluate the HF signals. Specifically, we develop an immediate human reward (IHR) reconstruction approach, regularized by environmental knowledge distilled in a latent space that captures the underlying dynamics of state transitions as well as issuing HF signals. Our approach has been tested over two real-world experiments, adaptive in-vivo neurostimulation and intelligent tutoring, as well as in a simulation environment (visual Q&A). Results show that our approach significantly improves the performance toward estimating HF signals accurately, compared to directly applying (variants of) existing OPE methods.

Paper

References (83)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer cuXN6/10 · confidence 3/52023-06-28

Summary

This paper proposes an off-policy evaluation method that utilizes sparse human feedback. Experiments are conducted on two real-world environments and one simulation environment.

Strengths

The idea of incorporating human feedback into policy evaluation is interesting and significant. The proposed framework is compatible with existing methods. The experiments are comprehensive. Especially, real-world experiments are appreciated.

Weaknesses

My major concerns are described in the Questions section. It would be appreciated if the authors address them in detail and I’m happy to raise my score.

Questions

1. Are 4 patients enough for validating the method? I’m wondering how many participants other similar methods adopt. 2. The proposed method additionally requires offline human feedback data. Are these comparing baselines also using the extra human data? If not, what did you do to make fair comparisons?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Please see the questions.

Reviewer 1oyy5/10 · confidence 3/52023-07-07

Summary

The authors examine an off-policy evaluation scenario where we have an environment with an intrinsic environment reward $r$, and an extrinsic human reward $r^H$. A further assumption is made that the human reward is only observed at the end of the episode, and comes as an rating on the total return of the episode. This is done to line up with cases where human ratings may only be clear after the fact. The goal of the paper is to do off-policy evaluation in this modified class of MDPs. Two baseline approaches for doing this are considered. The first is to treat the human return as a reward that is achieved at the last timestep of the episode, with $r = 0$ at all earlier steps. The second is to set the human reward $r^H = r$, where $r$ is the original known environment reward, then add an offset $G^H - G$ term to the last timestep, where $G^H$ = total human return and $G$ = total environment return. This paper proposes a method to learn timestep level $r^H$ directly. First it argues that given accurate timestep level $r^H$, the variance of the policy return estimate is equal or smaller than the estimate in the sparse MDP where all human reward is allocated to the final timestep. The reward learning method is then done based on two objective terms. The total reward must equal the human label of return for each trajectory, and the reward for individual $(s,a)$ should be similar to reward for other $(s',a')$ that have similar embeddings. This then leaves determining the embeddings. This is done via a VAE to learn a latent space, and the the nearest K-neighbors of a given $(s,a)$ embed are treated as similar examples. Once a reward estimate is learned, any off the shelf OPE method can be used.

Strengths

The paper makes an important observation that human feedback tends to be given at the end of an episode or after the fact, and is usually not given at the per-timestep level. It is clear that if per-timestep labels can be learned, then using them to estimate policy performance should give better results than estimating just total episode return (at least assuming learning per-timestep labels is a well-formed problem)

Weaknesses

I don't like that the paper proposes an "OPEHF framework", yet the bulk of the paper is really about learning a reward function / learning to distribute a return estimate over past trajectories. Nothing about the reward function learning is inherently tied to off-policy evaluation, it could theoretically be done online too, it's just something that can be done off-policy as well. The idea of return decomposition in the paper seems very similar to the RUDDER: Return Decomposition for Delayed Rewards paper, albeit with a different use case (in that paper, the environment naturally has a large sparse reward, which is decomposed to provide partial reward signal for an RL agent instead of needing to learn to reachieve the large sparse reward seen before.) It seems very relevant, in that the RUDDER paper proposes a number of return decomposition methods, none of which are considered in this paper. The embedding of $(s,a)$ is said to come from a bi-directional LSTM, in order to leverage information from both previous and subsequent steps. Doesn't this break the Markov assumption of MDPs? Since the rewards $r(s,a)$ now may depend on future states? How does the method handle cases where $r(s,a)$ may not be well defined? For example, two human raters have differing opinions on the score of a given trajectory. Overall I do not feel the paper is benchmarked sufficiently to be accepted. Edit: see rebuttal chain, have decided to increase score based on discussion of concerns.

Questions

Are there examples of ways the original environment reward can be used, aside from the fusion baseline mentioned? I am trying to understand if it makes sense to even view this as a new class of MDPs, vs treating it like an MDP with one reward function that is only observed at the last timestep. It seems like the method proposed here does not use original env reward in any way besides training the encoder-decoder in the VAE?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Yes

Reviewer KHGy6/10 · confidence 3/52023-07-07

Summary

This paper discusses a framework that learns to decompose human feedback reward, which is usually sparse and only given at the end of the episode, to intermediate step reward. They formulate a reward decomposition objective that considers original environment trajectory information using a variational auto-encoding representation learning objective as an implicit constraint. The authors show that their framework can reduce MAE for multiple OPE algorithms on a few domains. The core contribution of this paper is composed of a few known techniques, but the final result is an enhanced OPE with a lower MAE. The real-world experiments are the highlight of this paper.

Strengths

The paper has offered two very interesting real-world experiments to demonstrate the effectiveness of the proposed framework and show how it can enhance existing OPE algorithms. Indeed, a lot of existing OPE algorithms often work well with dense rewards, and these algorithms do not have an additional component that can infer intermediate rewards. I find the real-world experiments very refreshing and quite convincing.

Weaknesses

The methodology proposed in this paper is not very novel — the authors assembled a few known techniques in the AI/ML community. Also, I’m not entirely sure some of the content in the paper is necessary. The authors can correct me if I’m wrong or misunderstood the contribution, but some parts of the paper do not seem like original contributions but simply restatements of known results. Line 101: Problem 1. The authors indicated one of the issues is “the total number of offline trajectories AND $r_t^H$ being unknown.” However, in the later sections, I think only $r_t^H$ was solved. It doesn’t seem like the proposed framework would change much if “the total number of offline trajectories” is known or unknown. Can authors provide more insight on how this affects their framework design? Specifically, how would they change the algorithm if the number of trajectories is known? I sincerely suggest dropping this if this is not a serious factor that affects your algorithm design. Line 121 Proposition 1: The rescaled OPE estimator has a lower variance than the original OPE estimator. I think this is a known result. For example, weighted importance sampling has a lower variance than importance sampling. Similarly, weighted PDIS has a lower variance than PDIS. See Thomas [1] Chapter 3 for consistent weighted PDIS. Can the authors explain and clarify how they made a novel contribution beyond what’s provided in [1]? Otherwise, I suggest the authors cite the previous work so readers will not get confused. [1] Thomas, P. S. Safe Reinforcement Learning. PhD thesis, University of Massachusetts Amherst, 2015b.

Questions

1. FQE is an OPE algorithm that tries to do reward decomposition. However, the experiment still showed that using the decomposed reward produced from OPEHF, FQE-OPEHF can have a lower MAE than just using FQE alone. Can the authors comment on why this would be the case? 2. The objective is on the sum of discounted rewards being the same, but there is no constraint on reward from each timestep — this means “theoretically” we can have a neural network predict -10000 on one timestep and 9999 on the next step, and still have the sum of reward to be -1. Can the authors comment on why this degenerate case hasn’t happened, and can authors provide some analysis/insight/summary data on what the decomposed intermediate human feedback rewards look like?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors addressed the limitations of their approach.

Authorsrebuttal2023-08-16

Mid-point check-in

As we are stepping into the 2nd half of the discussion period, should the reviewer have any follow-ups, we will try out best to address them in time. If satisfied, we would greatly appreciate the reviewer to update the reviews/acknowledge our responses. We sincerely thank the reviewer again for the efforts devoted to the review process, allowing the work to be thoroughly evaluated and discussed.

Reviewer ochq7/10 · confidence 2/52023-07-27

Summary

The paper introduces a framework for Off-Policy Evaluation for Human Feedback (OPEHF). Off-policy evaluation allows the evaluation of learned policies on offline recorded data, which is especially valuable in scenarios where online deployment to evaluate policies is dangerous or expensive. Especially in scenarios such as healthcare, human feedback is very valuable and apparently often not very correlated with environmental rewards, making the combination of these two aspects very interesting. The paper considers the challenging case where human feedback is only available at the end of an episode without any per-step immediate human reward (IHR). To utilize existing OPE approaches, the per episode human reward is first reconstructed to a sequence of IHR by optimizing the sequence such that the cumulative discounted sum equals the human rewards, while additionally regularizing the reconstructed IHRs to align well with a latent representation of a state-action pair sequence. The resulting approach is evaluated on 2 real world scenarios, showing interesting results.

Strengths

- The paper tackles a problem with realistic applications and proposes an interesting approach to do so. - It's well written with an extensive supplementary. Even though this is sadly not my area of expertise, I felt the paper was made to be understandable. - Results look favorable compared to simpler baselines and the method does not rely on environmental rewards being correlated with the human feedback, for which the paper gives multiple interesting example use cases. - Code is provided.

Weaknesses

- While I very much like the actual longer term studies with real humans, overall the empirical evidence still is not extremely convincing. I guess this is an inherent problem with human feedback being expensive to collect. Nevertheless, some additional experiments would of course be very valuable.

Questions

- In Proposition 1 you assume that the expected value of the rewards is positive. This is obvious in the considered scenarios, but would this approach generalize to other scenarios where human feedback could indeed be negative? - To what extend does the VLM-H hyperparameter tuning play a role here? How robust is the overall approach to less optimal setups here? - If you could discuss further evaluation venues for the proposed method, that would be interesting, but honestly I quite like the overall paper and find it to be an interesting read. If no further bigger issues are brought up by the other reviewers, I am willing to up my rating, I would just like to keep it a little conservative given this is very much outside of my expertise.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Some limitations are discussed and no other obvious limitations or major negative societal impacts come to mind.

Authorsrebuttal2023-08-16

Mid-point check-in

As we are stepping into the 2nd half of the discussion period, should the reviewer have any follow-ups, we will try out best to address them in time. If satisfied, we would greatly appreciate the reviewer to update the reviews/acknowledge our responses. We sincerely thank the reviewer again for the efforts devoted to the review process, allowing the work to be thoroughly evaluated and discussed.

Reviewer ochq2023-08-16

Post rebuttal opinion

Thank you for answering all the questions and concerns. I was mainly curious to see if reviewer 1oyy would still reply. Even though this is not my area of expertise, I think you addressed the concerns thoroughly. Unless reviewer 1oyy is willing to spearhead the rejection of this paper, I am recommending this paper to be accepted and I upped my rating accordingly.

Authorsrebuttal2023-08-16

Thank you

We sincerely thank the reviewer for the prompt response, as well as the patience and efforts dedicated to evaluating our work thoroughly.

Reviewer cuXN2023-08-14

Response to the authors

Thanks for the authors' detailed responses. I'm happy to raise my score to 6.

Reviewer 1oyy2023-08-17

Apologies for delay, was addressing another deadline. > R1: We respectfully disagree that our work is more of a reward shaping work, nor can be adapted to the online learning setup, from two perspectives. > The motivation for decomposing human returns in our case, follows from the analysis in proposition 1 that proper decomposition can help reduce variance and lead to more accurate OPE estimation over policies that lead to varied performance. This is rather different than in existing reward shaping works that try to smooth over sparse rewards, and mainly focus on improving the convergence and/or optimality properties of the optimal policy that can be learned out of the shaped reward functions. > The focus of our work is to estimate the feedback/returns that are directly provided by human participants, toward the cases where the performance pre-existing RL policies need to be demonstrated before they get deployed to human participants, e.g., as required in healthcare systems. So the problem setup of OPEHF cannot be changed to online. Moreover, the use of bi-directional LSTM requires all trajectories to be fully observed a priori, and cannot be used online as well (see more in R3 below). I would argue that generally, improving the convergence the policy via reward shaping is *because* reward shaping reduces the variance of the policy update. Perhaps the more accurate way to put it is that both OPEHF and general reward shaping are trying to reduce variance, but one is for evaluation and the other is for improving gradient updates. I think that overall this is not an important concern though and am willing to move on. > On RUDDER The authors are correct that RUDDER's return decomposition view only ensures the optimal policy's return is unchanged. I was more curious at the different variants RUDDER proposes for return decomposition (see Section 3 of paper: integrated gradients, layer-wise relevance propagation, etc.) Could the authors clarify if OPEHF will maintain the ranks of different policies in the limit of infinite data? It was not clear to me that was a proposed contribution. > R3: The bi-directional LSTM is used to map the end-of-episode human return back to immediate human rewards (IHRs) over the entire horizon, for all the offline trajectories being considered. Specifically, it is used only after these trajectories are already observed (i.e., no online deployment nor adding new trajectories). Such a case falls under the typical sequence modeling setup where RNNs/LSTMs are mainly used. As an analogy, the use case here is somewhat reminiscent of how decision transformers [3, 4] are used to model the decision process. I think decision transformers are not a great analogy here. In decision transformers, we condition on the return to go, but this does not change the reward function $r(s,a)$ of the underlying MDP. That reward function is still Markovian / does not depend on the future. Although at training time DT is aware of the future (knows the true return to go), at inference time it is not aware of the future and we pick a return-to-go to condition on at $t=0$. In OPEHF, the reward $r(s,a)$ does depend on the future, via the bi-directional part of the LSTM. And, this is all fine for off-policy evaluation where we do not do online execution, but I would argue it's not exactly a reward function anymore. You would literally be unable to use the $r(s,a)$ during live inference if I understand correctly. To be clear: for the purposes of OPE, the bi-directional LSTM is probably helpful. Since the human giving the preference label sees the entire trajectory $\tau$ before rating the trajectory. I just would not describe it as $r(s,a)$ when it is more like $r(s,a|\tau)$ in practice. > Challenges in getting more than 1 rater for a scenario. I will quote the reply to Q4 from reviewer ochq. > Unlike frameworks like RLHF for language model training [5], where one can hire 3rd-party raters to provide feedback simply by following a predefined rubric (without the need of domain expertise); in human-centric domains like healthcare or education, professionals that are thoroughly trained (e.g., doctors, professors, instructors) over years may only be eligible to rate, other than the participants themselves. I agree with all of this, but think of the question as, *in a hypothetical scenario where there is more than 1 human rater, what would happen?* I am not expecting experiments for this, I understand the setting in this paper has 1 human rater and more would be difficult. I am more looking for discussion on if the method could handle multiple $r(s,a)$ labels gracefully or not. If it cannot it would be a good limitation to mention. > Evaluation To be specific, I was interested in comparisons to other methods of return decomposition. i.e. the proposal of this work is, "reconstruct IHR, here is 1 way to do it" and I wanted to see > 1 way of doing it (from the existing data, to avoid needing to get more)

Authorsrebuttal2023-08-18

Author's responses

We thank the reviewer for getting back to us in time. Please see our detailed responses below. Q7: I would argue that generally, improving the convergence the policy via reward shaping is because reward shaping reduces the variance of the policy update ...... **I think that overall this is not an important concern though and am willing to move on.** R7: Thank you for aligning with us that our work does not simply fall within the scope of reward shaping, echoing on our R1 above. Q8: **The authors are correct that RUDDER's return decomposition view only ensures the optimal policy's return is unchanged.** I was more curious at the different variants RUDDER proposes for return decomposition (see Section 3 of paper: integrated gradients, layer-wise relevance propagation, etc.) Could the authors clarify if OPEHF will maintain the ranks of different policies in the limit of infinite data? It was not clear to me that was a proposed contribution. R8: Reward decomposition is part of RUDDER’s methodology, with the final goal as constructing a **transformed MDP that has expected future rewards equal to zero**, from which Q-values can be captured smoothly. This is helpful for policy optimization as the agent mainly need to act following argmax of leanred Q values there to find the optimal policy, hence RUDDER only guarantees that the optimal policy under the pre- and post-transformation MDPs are equivalent. However, **it dose not guarantee that, under the transformed MDP, the redistributed rewards can properly facilitate downstream OPE estimators**. As this would require one to show that not only the optimal policy between two MDPs are equivalent, but all the sub-optimal policies can result in the same return distribution before and after transformation; because policies with varied scale of returns will be evaluated by OPE [1,2 above], not only just the optimal ones. **In contrast, our framework did not rely on any MDP transformations, thus would not bounded by this limitation.** To this end, we believe that it would require non-trivial efforts to integrate the return decomposition method in RUDDER into the OPEHF framework. However, as we mentioned, we appreciate the reviewer sharing an interesting read, and we would be more than happy to discuss in the camera-ready how similar ideas in RUDDER can be adapted toward OPEHF in the future, if accepted. We believe one of the (future) adaptation could be using this idea to further improve FQE estimators, as part of the downstream OPE estimator we considered. Q9: In OPEHF, the reward $r(s,a)$ does depend on the future, via the bi-directional part of the LSTM. And, this is all fine for off-policy evaluation where we do not do online execution, but I would argue it's not exactly a reward function anymore. You would literally be unable to use the $r(s,a)$ during live inference if I understand correctly. **To be clear: for the purposes of OPE, the bi-directional LSTM is probably helpful.** Since the human giving the preference label sees the entire trajectory before rating the trajectory. I just would not describe it as $r(s,a)$ when it is more like $r(s,a|\tau)$ in practice. R9: Thank you for alingning with us on the purpose of using bi-directional LSTM for RILR. We wanted to emphasize that the immediate human rewards $r^\mathcal{H}(s_t,a_t)$ follows Markov property, as defined in Section 2.1. However, with $r^\mathcal{H}(s_t,a_t)$ being unknown, the RILR uses bi-directional LSTM to approximate $\hat r^\mathcal{H}(s_t,a_t)$ which may depend on future states. We agree with the reviewer that we can correct the notation $\hat r^\mathcal{H}(s_t,a_t)$ to be $\hat r^\mathcal{H}(s_t,a_t|\tau)$ in the camera-ready. Since we are considering OPE where we only work with a fixed offline dataset, we would not encounter the cases that require ‘live inference’. Q10: **I agree with all of this**, but think of the question as, in a hypothetical scenario where there is more than 1 human rater, what would happen? R10: Our work specifically consider the case where the participants directly provide feedback, as defined in Section 2.1. It is mentioned as part of our future work where we are building a platform to allow professionals to rate the trajectories as well, and there definitely need some extra work to adapt the current OPEHF frameowrk for that, as it is beyond the scope of this paper.

Authorsrebuttal2023-08-18

Author's response (cont.)

Q11: To be specific, I was interested in comparisons to other methods of return decomposition. i.e. the proposal of this work is, "reconstruct IHR, here is 1 way to do it" and I wanted to see > 1 way of doing it (from the existing data, to avoid needing to get more) R11: Besides the discussion in R8, we wanted to further emphasize that most existing reward decomposition work **decompose the delayed environmental returns**, leveraging the fact that some state-action pairs may contribute more to the final returns, following typical environmental dynamics in, for example, games and robotics. However, the human returns in OPEHF is de-coupled from the environmental returns, which is justified in experiments that low correlations are found between them in general. Multiple works have found that it does not simply follow the way how environmental returns would be composed, as it depends on many factors even outside of scope of what the environmental dynamics could capture, such as the mood when a participant provides feedback and more other underlying confounders that are non-trivial to capture [5,6 above]. We believe that our current work is both technically and experimentally complete, and solves an important empirical problem within the scope of OPE (as pointed out by reviewers KHGy, ochq and cuXN). As we are, to the best of our knowledge, the first work that introduce the framework for OPEHF, we tried to make our approach straightforward, to maximize the comaptibillity with existing OPE frameworks. Some components in our framework could potentially be further improved by leveraging the ideas from related communities, and we appreciate the reviewer pointing to us one possible future avenue. Again, we greatly appreciate the reviewer’s efforts facilitating meaningful discussions with us, and we are delighted to see that the reviewer has aligned with us on many of the questions above. We are always more than happy, and grateful, to have the chance addressing any further comments. We wish the reviewer all the best for the endeavor catching up the other deadline.

Authorsrebuttal2023-08-18

Any follow-ups?

As the discussion window is closing in the next 2 days. We are wondering if the reviewer would have any follow-ups over our latest response. We would also greatly appreciate the reviewer to **reconsider the assessment and recommendation** of our work, as we have seen that the reviewer had agreed with us on many points above (as highlighted when we quote the reviewer's responses).

Reviewer 1oyy2023-08-19

> R7: Thank you for aligning with us that our work does not simply fall within the scope of reward shaping, echoing on our R1 above. To be clear: I still pretty strongly *disagree* on this point. I just think it is not relevant to discussion on the quality of the paper. Please don't imply that I agree here. > R8: On RUDDER and OPEHF: I don't think the transformed MDP in RUDDER is particularly important. The transformed MDP accumulates reward-so-far into an extra feature just to keep the Markovian assumption true. But if done in an OPE way, then we do not need to keep that Markovian assumption, right? That is what the authors are arguing. RUDDER does not guarantee the transformed reward facilitates OPE - but then, neither does the author's method. Both are about how the VAE-based embedding works in practice. > R9: On Markovian reward: If the authors mention the distinction here that the approximate $\hat{r}$ is not a true reward function, is conditional on $\tau$, and is not usable for "live inference", then I am satisfied on this point. (I do think there is an interesting question here about if $r(s,a)$ is even a suitable model for human reward - see http://ai.stanford.edu/blog/robomimic/ for an argument that human decision making is often not Markovian due to unobserved state. Might be of interest to authors.) > R10: More than 1 human rater: Please include this as a limitation of the work as well. > R11: Other return decomposing methods. Here, the authors say that > However, the human returns in OPEHF is de-coupled from the environmental returns, which is justified in experiments that low correlations are found between them in general. Multiple works have found that it does not simply follow the way how environmental returns would be composed, as it depends on many factors even outside of scope of what the environmental dynamics could capture, such as the mood when a participant provides feedback and more other underlying confounders that are non-trivial to capture [5,6 above] If human returns are decoupled from environmental returns because of factors outside the scope of environmental dynamics, how is that handled by the proposed methods in OPEHF? The OPEHF embeddings are based solely on fitting from the environmental dynamics sequence $(s,a,s,a,\cdots)$ to the human return label $R$. I think at a theory level it's all equally intractable, but then deep learning finds a way. Which is to say that, this argument that other return decomposing methods are less relevant does not feel strong to me. ----------------------------- I thank the authors for replying to my comments so far. I'd like to let them know that I'll be busy tomorrow and won't be available to respond (but they are welcome to write a reply for other reviewers to read if they wish). I plan to bump my 3 to a 5, I still have a number of reservations about the work but am feeling more borderline about it than the 1st pass.

Authorsrebuttal2023-08-20

Thank you

We thank the reviewer for sharing these thoughtful comments, and we indeed appreciate the feedback/thoughts from different aspects. We will definitely reflect them in our camera ready version, if accepted.

Reviewer KHGy2023-08-21

Thank you for the rebuttal -- the authors have discussed the points I've raised, and I'm satisfied by the response. I will keep my score as is.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC