acceptodds
Under review as a conference paper at ICLR 2027

Learning User Representations to Improve Simulation Fidelity

Abstract

User simulators enable interactive training and evaluation of AI assistants at scale. Yet current simulators do not reproduce the diverse behaviors of real users very well, leading to assistants that struggle to generalize to underrepresented users. We identify user representations as an overlooked bottleneck in user simulation. These are natural-language descriptions that simulators are conditioned on to reproduce a particular user's behavior. A representation's quality becomes apparent in hindsight, through the behavior it induces in a simulator. To evaluate and optimize user representations, we formulate a conversation reconstruction task: given an original conversation, we generate a user representation, condition a simulator on it, simulate a conversation, then measure fidelity by comparing user behavior in the original and simulated conversations. This comparison also gives hindsight feedback, which identifies where the representation falls short and how it can be improved. We introduce hindsight distillation, an online training procedure that distills feedback-conditioned improvements into the representation generator, enabling it to generate improved representations directly from the original conversation. Across 700 conversations spanning seven datasets, hindsight feedback improves user representations at test time, and hindsight distillation yields UserRepGen-4B, whose generated representations enable more faithful simulation than those generated by GPT-5.5. Beyond conversation reconstruction, the learned representations also improve fidelity for unseen tasks, better preserve the population-level diversity of real users, and increase user re-identification accuracy.

Primary area:
datasets and benchmarks
Keywords:
user representation, user simulation, simulation fidelity, hindsight distillation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_z5cAOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies how learned user representations can improve the behavioral fidelity of LLM-based user simulators. It separates representations into task and behavior descriptions and introduces a conversation reconstruction framework, where simulated interactions are compared with original conversations to generate hindsight feedback. The authors use this feedback to train UserRepGen-4B through online supervised fine-tuning. Experiments across seven datasets report improvements in reconstruction fidelity, with additional analyses examining cross-task transfer, population-level behavioral diversity, and user identification across conversations.

Critical strengths

The paper addresses a meaningful problem in user simulation: improving simulation fidelity by refining user representations rather than solely optimizing the simulator itself. Separating task descriptions from behavioral descriptions provides a clear representation interface that could potentially be reused across tasks.

The paper is clearly and concisely written, making it easy to follow.

Critical weaknesses

First, the paper primarily relies on GPT-5.4 mini to assess whether simulated users resemble real humans, without sufficiently establishing its ability to judge behavioral fidelity. The evaluation emphasizes linguistic style, formatting, response length, and turn count, making it difficult to distinguish genuine behavioral improvements from adaptation to the scoring rubric. Higher scores do not necessarily imply more realistic human behavior. The reliance on a relatively weak judge is particularly concerning given the lack of evaluation using stronger models.

Second, although the authors report human evaluation, the study in the appendix mainly tests agreement between human preferences and the LLM judge’s rankings, while excluding cases where the judge assigns equal scores. This does not substitute for user studies and A/B testing with real participants, nor does it demonstrate that simulated users respond as real users would in new situations. Insufficient reporting of annotator numbers, inter-annotator agreement, and verifiable annotation records also makes the reliability of this evidence difficult to assess independently.

Finally, although the paper includes supplementary analysis using an existing distributional-gap metric, its main results still rely on an author-designed conversation reconstruction protocol. The paper lacks validation on independent evaluation resources and comparisons with existing user-profile construction, behavioral representation, and specialized simulation methods and models. Furthermore, the experiments do not sufficiently disentangle the contributions of training data, prompt design, representation length, and the proposed algorithm. Consequently, the claimed advantage over GPT-5.5 remains limited to the authors’ evaluation protocol and does not establish a genuine advantage in realistic user simulation.

Questions for authors

The authors may not have had sufficient time to complete a more thorough evaluation. I encourage them to prioritize addressing the concerns raised in the Weaknesses section. If additional experiments and explanations adequately resolve these concerns, I would be willing to raise my score.

One additional question concerns Section 7.2: with 700 samples per group and 500 clusters, the estimated JS divergence may be sensitive to finite-sample effects and clustering randomness. Could the authors report results across different random seeds and cluster counts, along with confidence intervals, to verify that the reported reduction in the distributional gap is robust?

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/n)

Thank you for your careful review of our paper! We are encouraged that you recognize the importance of optimizing user representations and highlight that our paper is clearly and concisely written. We agree with your concerns regarding our evaluations, and report additional experiments and explanations in the sections below.


W1. Judge reliability and adaptation to the scoring rubric

Higher scores do not necessarily imply more realistic human behavior.

We understand the reviewer's concern about GPT-5.4 mini's ability to judge behavior fidelity. We put extensive effort into carefully designing our evaluation to ensure that it measures realism well. Firstly, we developed the criteria and behavioral dimensions that were important to this task, and examined whether the judge's rationales and scores reflected the intended comparisons. As described in Appendix C, these dimensions cover communication style, engagement, request formulation, responses to assistant outputs, and context provision. Importantly, our evaluation compares simulated behavior with the original user behavior, rather than simply asking whether a conversation appears plausible.

We would also like to mention that evaluating user simulators with rubrics and LLM-judges is a well-established practiced for evaluating realism and behavioral fidelity in user simulation research:

  • OdysSim [3] (NeurIPS 2026) uses GPT-5 nano as its default judge for LLM-scored tasks in the SoUL-Index, which is a suite of 24 benchmarks that includes UserLM, SimulatorArena, MirrorBench, CoSER, and HumanLM's Humanual-Chat. All five of these compare simulations against real human behavior.
  • MUSE [4] (EMNLP 2026) uses DeepSeek-V3 to evaluate generated user responses.
  • Turing-RL [5] (EMNLP 2026) uses Claude Sonnet 4.6 for to evaluate Turing distinguishability, similarity to real user responses, and response originality.
  • HumanLM [6] (ICML 2026) uses Claude Haiku 4.5 to evaluate alignment between simluated and real user responses and states, and GPT-5 mini during training.
  • ThinkingUS [7] (ACL 2026) uses DeepSeek-V3 to judge the realism of user simulators on reference-base dand intrinsic criteria.
  • UserLM [8] (ICLR 2026) uses GPT-4o to evaluate intent adherence.
  • SimulatorArena [9] (EMNLP 2025) uses GPT-4o to assess similarity between simulated and real user messages.
  • USP [10] (ACL 2025) uses GPT-4o to evaluate profile consistency and expression of subjective characteristics.
  • CoSER [11] (ICML 2025) uses GPT-4o to evaluate character fidelity.

Nevertheless, we share concerns with the reviewer regarding only relying on an LLM judge. For this reason, we complement it with two evaluations that do not use the the judge's fidelity score:

  • Preserving population-level behavioral diversity (Section 7.2): We apply a previously proposed distributional-gap metric to compare the behavioral distributions of original and simulated conversations.
  • Distinguishing users across conversations (Section 7.3): We present a multiple-choice task that tests whether behavior descriptions capture characteristics that identify the same user across different conversations.

The reliance on a relatively weak judge is particularly concerning given the lack of evaluation using stronger models.

We agree with the reviewer, and re-evaluated the conversations with four additional judges: Qwen3.5-122B-A10B, gpt-oss-120b, GPT-6 Luna, and GPT-6 Sol. These are also reported in our responses to the other reviewers. We report Best-of-1 / Best-of-3 fidelity:

GeneratorGPT-5.4 miniQwen3.5-122B-A10Bgpt-oss-120bGPT-6 LunaGPT-6 Sol
Qwen3.5-4B (base)3.41 / 4.422.81 / 3.713.30 / 4.293.07 / 3.842.97 / 3.77
+ GRPO3.57 / 4.592.90 / 3.833.45 / 4.413.15 / 3.953.12 / 3.90
+ SFT4.64 / 5.683.74 / 4.834.36 / 5.453.87 / 4.793.84 / 4.75
+ OPSD4.46 / 5.593.57 / 4.724.11 / 5.333.69 / 4.693.75 / 4.65
+ OSFT (UserRepGen-4B)4.95 / 6.023.89 / 5.084.50 / 5.734.05 / 4.984.05 / 4.92
Qwen3.5-9B3.68 / 4.683.02 / 3.973.61 / 4.623.22 / 4.073.18 / 4.04
Qwen3.5-27B3.76 / 4.833.04 / 3.993.69 / 4.763.34 / 4.223.30 / 4.16
Qwen3.5-122B-A10B3.52 / 4.552.86 / 3.853.45 / 4.463.14 / 4.003.13 / 3.98
GPT-5.4 mini3.73 / 4.713.06 / 4.023.57 / 4.543.31 / 4.093.24 / 4.04
GPT-5.53.89 / 4.903.15 / 4.123.77 / 4.793.38 / 4.203.37 / 4.23

Our conclusions hold under every judge, and the relative orderings of the representation generators are nearly identical. The table below reports Spearman rank correlations with the original GPT-5.4 mini judge:

JudgeBest-of-1Best-of-3
Qwen3.5-122B-A10B0.990.98
gpt-oss-120b0.990.98
GPT-6 Luna1.000.98
GPT-6 Sol0.990.98
AuthorsOpenReview ↗

Rebuttal by Authors (2/n)

W2. Human Validation of LLM Judges

We agree with the reviewer that our human validation study is not a substitute for user studies or A/B testing with real participants. We would like to clarify that the purpose of this study is to validate the judge's fidelity scores. Specifically, we test whether the judge's rankings agree with human assessments of how faithfully user simulators reproduce the original user behaviors.

To do so, we carefully constructed a validation set of 100 pairwise comparisons. Each example contains an original conversation and two simulated conversations, together with a preference that selects the more faithful simulated conversation. The authors selected these preferences through detailed analysis and discussion. We believe that it is necessary for the authors to do this task because it requires expertise and familiarity with the task and the behavior criteria in Appendix C. In particular, it requires distinguishing behavior fidelity from other metrics, such as general conversation quality (a more fluent simulation is not necessarily more faithful to the original user). During labeling, the simulations were presented in randomized order, with their source and scores hidden. Most importantly, we would like to highlight that the comparisons are grounded in observed user behavior, rather than the general impression of whether a conversation "looks human". Our validation therefore tests whether the judge's rankings agree with these carefully selected preferences. We will update Appendix B to be more clear about these details.


W3. Comparisons with Prior Work and Experiment Controls

The paper lacks validation on independent evaluation resources and comparisons with existing user-profile construction, behavioral representation, and specialized simulation methods and models.

We than the reviewer for suggesting these comparisons. We have conducted additional experiments with representations based on prior work, also reported in our responses to Reviewers eedj and uo1b:

  1. Task only: our task description, without a behavior description.
  2. Intent (UserLM): the user's overall goal, generated using UserLM's intent-generation prompt.
  3. Task + Persona (PersonaHub): our task description paired with a PersonaHub-style description of demographic and biographical attributes.
  4. Task + Implicit Profile (USP): our task description paired with an adapted USP profile that describes the user separately from the task.

We generate each representation with both Qwen3.5-4B and GPT-5.5. For the conversation reconstruction task, we report Best-of-1 / Best-of-3 fidelity:

GeneratorTask onlyIntent (UserLM)Task + Persona (PersonaHub)Task + Implicit Profile (USP)Task + Behavior (ours)
Qwen3.5-4B3.04 / 3.842.86 / 3.613.01 / 3.783.23 / 4.053.41 / 4.42
GPT-5.53.16 / 3.983.04 / 3.773.09 / 3.843.51 / 4.393.89 / 4.90
UserRepGen-4B————4.95 / 6.02

We additionally report results for all three tasks in Section 7. For simulating users in new scenarios, we add PersonaHub-style personas and adapted USP profiles to the four conditions. For the added baselines, the persona or profile replaces the behavior description, and the task is held fixed across conditions:

Generator(A)(B)(C)(D)
Task + Persona (PersonaHub)3.30 / 4.023.23 / 3.953.16 / 3.783.18 / 3.92
Task + Implicit Profile (USP)3.30 / 4.023.32 / 4.033.49 / 4.203.57 / 4.45
Qwen3.5-4B (base)3.30 / 4.023.46 / 4.163.37 / 4.173.80 / 4.68
+ SFT3.31 / 4.014.36 / 5.374.61 / 5.705.32 / 6.35
+ OPSD3.57 / 4.254.10 / 5.204.37 / 5.514.95 / 5.96
+ OSFT (UserRepGen-4B)3.43 / 4.164.48 / 5.504.69 / 5.805.45 / 6.59

For preserving the behavioral diversity at the populaton level, we include results for Intent, Task + Persona, and Task + Implicit Profile:

GeneratorRepresentationJSSelf-BLEU
Qwen3.5-4BIntent (UserLM)0.3890.411
Qwen3.5-4BTask + Persona (PersonaHub)0.3400.376
Qwen3.5-4BTask + Implicit Profile (USP)0.3500.366
Qwen3.5-4B (base)Task + Behavior (ours)0.3150.362
+ GRPOTask + Behavior (ours)0.3380.353
+ SFTTask + Behavior (ours)0.2700.319
+ OPSDTask + Behavior (ours)0.3080.333
+ OSFT (UserRepGen-4B)Task + Behavior (ours)0.2730.316
AuthorsOpenReview ↗

Rebuttal by Authors (3/n)

And finally, for distinguishing users across conversations, we replace the descriptions with personas or profiles in the same user-identification tasks:

GeneratorRepresentationAccuracy
-Full conversation (baseline)44.22
Qwen3.5-4BPersona (PersonaHub)34.47
Qwen3.5-4BImplicit Profile (USP)38.55
Qwen3.5-4BBehavior (ours)38.78
+ GRPOBehavior (ours)41.50
+ SFTBehavior (ours)43.54
+ OPSDBehavior (ours)50.34
+ OSFT (UserRepGen-4B)Behavior (ours)51.02

Across all four of these tasks, UserRepGen-4B consistently outperforms the added representation baselines. It leads to higher simulation fidelity for conversation reconstruction and for simulating a user in unseen scenarios, lower distributional gaps and lexical homogeneity at the population-level, and higher same-user identification accuracy.

Why we keep the simulation method and models fixed. In Section 3, we explain how recent works focus on optimizing the user simulator (i.e. the method or model), while representation generation is typically done with a general-purpose LLM. In our work, we focus on optimizing the user representation that the simulator will be conditioned on, and optimizing the model that generates these user representations. In our evaluation setup, we intentionally keep the simulator fixed so that we can isolate the effects of the representation. Since only the representation changes, and the simulator stays the same, it is a controlled setting where any differences in simulation fidelity can be attributed to the representation.

Lastly, we train using an ensemble of simulation environments based on Qwen3.5-4B and Gemma3-4B, and evaluate using GPT-5.4 mini. Using different models during training and evaluation is intended to reduce overoptimizing to a single simulator, and ensuring that the learned representations generalize to simulators not seen during training.


Q2. Robustness of JS divergence to clustering and finite sample size

Thank you for raising this question. Section 7.2 adopts the distributional gap metric introduced by [10], and follows the same configuration. [10] validates their evaluation approach by reporting sensitivity analyses across embedding models, clustering algorithms, clustering counts, and find that simulator rankings remain largely stable. Nevertheless, we agree that robustness should also be assessed in our setting. In particular, we use 700 conversations per group while the original work uses 5,000. We repeat the ablations for clustering counts to assess whether our results remain consistent across different cluster sizes. The table below reports Spearman correlations between the JS-divergence rankings at each cluster count and the default setting of k=500:

Cluster count501002501000
Spearman correlation with k=5000.830.930.950.98

The rankings achieve strong correlations across the cluster counts, particularly for k >= 100. The lower correlation at k=50 indicates more sensitivity to coarse clustering, but the overall agreement supports the stability of the rankings across different cluster counts.

AuthorsOpenReview ↗

Rebuttal by Authors (4/n)

References

[1] Zhou, X., et al. (2026). OdysSim: Building foundation models for human behavior simulation. In The Fortieth Annual Conference on Neural Information Processing Systems.

[2] Liu, Z., et al. (2026). MUSE: Multi-domain Chinese user simulation via self-evolving profiles and rubric-guided alignment. In Findings of the Association for Computational Linguistics: EMNLP 2026.

[3] Wang, Y. S., et al. (2026). Learning user simulators with Turing rewards. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing.

[4] Wu, S., et al. (2026). HumanLM: Simulating users with state alignment beats response imitation. In Forty-third International Conference on Machine Learning.

[5] Wu, X., et al. (2026). Thinking alignment of scenario-oriented user simulation. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

[6] Naous, T., et al. (2026). Flipping the dialogue: Training and evaluating user language models. In The Fourteenth International Conference on Learning Representations.

[7] Dou, Y., et al. (2025). SimulatorArena: Are user simulators reliable proxies for multi-turn evaluation of AI assistants? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.

[8] Wang, K., et al. (2025). Know you first and be you better: Modeling human-like user simulators via implicit profiles. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

[9] Wang, X., et al. (2025). CoSER: Coordinating LLM-based persona simulation of established roles. In Forty-second International Conference on Machine Learning.

[10] Ge, T., et al. (2024). Scaling synthetic data creation with 1,000,000,000 personas.

[11] Mehri, S., et al. (2026). Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors. In The Fortieth Annual Conference on Neural Information Processing Systems.


Thank you again for your constructive feedback. We hope that our additional experiments and explanations address the weaknesses and concerns that were raised. We would be happy discuss any remaining questions or concerns, and appreciate the time and effort the reviewer has put into our work.

Reviewer_uo1bOpenReview ↗
Rating
2
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper addresses the representation generator as the bottleneck for user simulation, and formulate a reconstruction task to evaluate this capability. It proposes hindsight distillation, and compare various method for this task, showing OSFT performs the best.

Critical strengths
  1. This paper addresses the representation generator as the bottleneck for user simulation. This is an unrecognized view that deserves more interest.
  2. The paper compares various training methods for representation generation, and gives a detailed analysis of why OSFT performs well compared with SFT and other on-policy methods.
Critical weaknesses
  1. Conversation reconstruction is inappropriate for evaluating representation generation, so the main analyses can be misleading. As stated in Section 3, the primary metric is how well a simulator reproduces the very conversation C that the representation was generated from. This is an autoencoder-like reconstruction objective. Using it for training may be fine, but using it as the main evaluation is not: the metric is maximized by a representation that simply scripts C, and nothing in the evaluation penalizes this (the low ROUGE-L of the resulting output and guiding prompt not to copy do not inherently address the caveat of evaluation). This questions the generalizability of the proposed model and framework. The experiments that actually test capabilities that can be used in real scenarios are treated as "broader utility" in Section 7, and there the gains are small. The Section 7 settings, where the target conversation is held out, should be the main evaluation, and additional analysis should be done on expansion to OOD distributions or tasks.

  2. The evaluation is insufficient to show that representation generation is the bottleneck. All compared methods are representation generators under the same paradigm with the simulator fixed, which does not show that representation generation, rather than the simulator, is the bottleneck, which is the main claim of the paper. Comparison or additional analysis with prior work that trains the simulator itself is fundamental for this paper. Additionally, the evaluation uses the same small model, GPT-5.4 mini, as the simulator, assistant, and judge. Using an identical model for simulation and judging can introduce unexpected bias even if the results look sound and are human validated.

Questions for authors

See weaknesses. Please let me know if there is any misunderstanding of this paper, especially regarding the task/evaluation formulation.

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

This paper investigates the task of user simulation and introduces a conversation reconstruction framework to improve simulation fidelity. Given an original conversation, the framework generates a user representation, conditions a user simulator on this representation, and evaluates simulation fidelity by comparing user behaviors in the original and simulated conversations. The authors further propose Hindsight Distillation, which leverages hindsight feedback obtained from this comparison process to improve user representation learning. Experiments on 700 conversations collected from seven datasets demonstrate the effectiveness of the proposed approach.

Critical strengths

S1. The paper addresses an important problem in user simulation. The idea of formulating user simulation as a conversation reconstruction task is intuitive and well-motivated. In particular, Figure 1 clearly illustrates the overall motivation and framework.

S2. To address the high cost of obtaining hindsight feedback, the authors propose distilling feedback-conditioned improvements into the representation generator, allowing it to generate improved user representations without requiring additional simulation at inference time. The experimental results demonstrate consistent improvements over several baseline approaches.

Critical weaknesses

W1. The methodological novelty and contributions are not sufficiently clear.

Although the authors discuss several related works on user modeling and user simulation, the paper lacks direct experimental comparisons with existing approaches and does not sufficiently clarify its distinctions from prior studies. For example, UserLMs (Naous et al. 2026) and USP (Wang et al., 2025) have also explored user representation modeling. In particular, USP appears to adopt a similar conversation reconstruction paradigm. The authors argue that existing approaches, such as UserLMs, struggle to capture individual user behaviors and preserve user diversity, but these claims are not supported by direct experimental comparisons.

Furthermore, although the authors introduce structured user representations, the paper does not sufficiently elaborate on their concrete content, construction principles, and differences from existing user representations. The proposed Hindsight Distillation also largely builds upon existing distillation and online supervised fine-tuning techniques, without introducing substantial task-specific modifications. These issues weaken the methodological novelty and the clarity of the paper's contributions.

W2. The experimental setup and evaluation protocol require further clarification.

Although the authors conduct experiments on seven datasets, several important details of the evaluation protocol remain unclear. In particular, GPT-5.4 mini is used as the user simulator, assistant, and judge during evaluation. While the general evaluation framework is described, it would be helpful to provide a more detailed explanation of how representations generated by different models are integrated into the same simulation environment and how evaluation consistency is ensured.

Moreover, using the same underlying model for user simulation, assistant responses, and fidelity evaluation raises concerns about potential evaluation biases and the reliability of the reported results. Although the authors provide a human validation study, these concerns are not fully addressed.

Additionally, most comparison models are general-purpose LLMs rather than specialized user modeling or simulation approaches. It is unclear whether these general-purpose models are evaluated under sufficiently comparable settings, particularly regarding prompting strategies and representation generation. More details are needed to establish the fairness of these comparisons.

W3. The experimental evidence for generalization and robustness remains insufficient.

In Section 7, the authors evaluate the proposed approach in terms of generalization to new scenarios, population-level behavioral diversity, and user distinguishability. However, these experiments mainly compare different training variants of the proposed framework, without including independent third-party user modeling or simulation methods as reference baselines.

Moreover, the performance differences among the evaluated variants are relatively modest in some experiments, particularly regarding cross-conversation generalization. Without comparisons against existing approaches or additional human evaluations, it is difficult to determine whether these numerical improvements correspond to meaningful differences in actual simulation behavior.

This raises concerns about the robustness of the experimental findings and whether the observed improvements can be generalized beyond the specific evaluation framework adopted in this paper.

Questions for authors
  1. Comparison with existing approaches: Could the authors provide a more explicit comparison with existing user modeling and simulation approaches, such as UserLMs and USP, in terms of both methodological design and experimental performance? In particular, what distinguishes the proposed conversation reconstruction framework and representation learning strategy from these prior works?

  2. User representation design: Could the authors elaborate on the design of the structured user representations, including their specific components, the information captured by each component, and the principles used to distinguish task-specific information from user-specific behavioral characteristics?

  3. Downstream utility: Could the authors provide evidence of the practical effectiveness of the proposed approach in downstream applications? For example, would using the learned user representations for interactive assistant training or evaluation lead to measurable improvements in assistant performance?

Comments on the paper

1 comment

Thank you for the response and additional experiments. Some concerns, including judge dependence and comparisons with prior representations, have been addressed. However, my central concern regarding the evaluation setting remains unresolved.

The authors’ response for W1 cites prior simulator evaluations. In those settings where representations are fixed and exclude target responses, reconstruction can meaningfully evaluate the simulator. Here, however, the representation generator receives the entire evaluation target, allowing the objective to reward conversation-specific reconstruction scripts without establishing transferable user behavior.

My concern is not whether hacking actually occurs, but whether the evaluation distinguishes these two possibilities. The analyses in Section 6.4 and Appendix E do not resolve this limitation. I acknowledge that Section 7 partially addresses this concern and appreciate its expansion. However, these evaluations should form the main experimental foundation of the paper. Stronger validation of transfer and further analysis of representation generation as a bottleneck are needed to establish a sufficiently substantive contribution.

I recognize the value of this work, but substantial revision is needed to align the evaluation with the claims. I therefore retain my overall score (2:Weak reject), while raising soundness from 1 to 2, and informativeness from 2 to 3 in recognition of the additional evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.