acceptodds
Under review as a conference paper at ICLR 2027

Improve Large Language Model Systems with User Logs

Abstract

Scaling training data and model parameters has long driven progress in large language models (LLMs), but this paradigm is increasingly constrained by the scarcity of high-quality data and diminishing returns from rising computational costs. As a result, recent work increasingly focuses on continual learning from real-world deployment, where user interaction logs provide a rich source of authentic human feedback and procedural knowledge. However, learning from user logs is challenging due to their unstructured and noisy nature. Vanilla LLMs often struggle to distinguish useful feedback signals from noisy user behavior, and the disparity between user log collection and model optimization (e.g., the off-policy optimization problem) further strengthens the problem. To this end, we propose UNO (User log-driveN Optimization), a unified framework for improving LLM systems (LLMsys) with user logs. UNO first distills logs into semi-structured rules and preference pairs, then employs query-and-feedback-driven clustering to manage data heterogeneity, and finally quantifies the cognitive gap between the model’s prior knowledge and the log data. This assessment guides the LLMsys to adaptively filter out noisy feedback and construct different modules for primary and reflective experiences extracted from user logs, thereby improving future responses. Extensive experiments show that UNO achieves state-of-the-art effectiveness and efficiency, significantly outperforming Retrieval Augmented Generation (RAG) and memory-based baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Continual Learning, Large Language Models, User Logs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2DonOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes UNO, a framework for improving LLM response quality using user interaction logs collected from real-world LLM systems. Specifically, UNO extracts feedback from user logs, converts it into rules and preference pairs, and performs clustering based on queries and rules. It then measures the cognitive gap of each cluster and either directly trains an Expert LoRA or trains a Critic LoRA that generates feedback to improve responses, which are then utilized through routing. Through experiments, this paper demonstrates that UNO outperforms existing RAG, memory-based methods, and training-based baselines.

Critical strengths
  • Addressing an important problem: The problem addressed in this paper is highly important for the development of LLMs, and this work provides a valuable milestone toward solving it.
  • Reasonable design choices supported by experimental results: The design choices that constitute the proposed method are reasonable and clearly presented. Furthermore, the effectiveness of each component is well demonstrated through experimental results, including Table 1.
  • Addressing various concerns through extensive experiments: Efficiency issues, applicability to real-world data, and effectiveness in online settings are important aspects that need to be addressed given the characteristics of the proposed method. This paper effectively demonstrates the effectiveness of the proposed method in these aspects through extensive experiments, providing comprehensive evidence of its advantages.
Critical weaknesses
  • Concerns about generalization to diverse tasks: I have concerns regarding whether the proposed method can generalize to diverse tasks, particularly regarding its hyperparameters and its reliance on embedding-based clustering.
    • Threshold hyperparameter: In the proposed method, the number and granularity of clusters are determined by the distance threshold, which directly affects performance. Although MemoryBench and WildFB cover relatively diverse domains and tasks, I am concerned about whether the proposed method would require hyperparameter tuning whenever it is applied to new settings as it becomes more widely used.
    • Concerns about environments involving multiple tasks: In realistic LLM systems, diverse and more tasks and user feedback coexist within a single environment, where many, different items may be densely distributed in the embedding space. Therefore, I am concerned about whether the proposed method can still operate effectively in such settings.
      • Furthermore, since the current experiments involve relatively few clusters (2–4 per task), I believe additional experiments are needed to verify whether clustering and routing can still work effectively when more diverse logs are mixed together and the number of clusters increases.
Questions for authors

After reviewing the Appendix and code, it appears that the experiments on MemoryBench performed clustering and routing separately for each task group (e.g., Short-Long, etc.). Could you provide results when clustering and routing are performed jointly across all task groups?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2Don

Thank you for the careful review and thoughtful questions. We address your questions below and have updated the manuscript accordingly.

1. Hyperparameter generalization

Thank you for raising this question. All experiments in our paper share the same default hyperparameters without any per-setting tuning. Specifically, key hyperparameters such as the clustering distance threshold, the outlier routing threshold, the cognitive gap threshold τ∗\tau^*, and the verifier win-rate threshold γ\gamma (App. C.3) are kept identical across both base models, all four task groups, the joint experiment below, and WildFB, despite its very different distribution. We now state this explicitly in Sec. 4.3.

Tuning is unnecessary for two reasons. First, the key hyperparameters lie in wide stable ranges (Figs. 9 to 11). For example, the distance threshold has a clear plateau, and performance drops only at the extremes (a threshold of 2 with too few samples per cluster, or no clustering at all). Second, the two-stage filtering turns most hyperparameter errors into extra cost rather than lower quality. Noisy clusters that pass the gap filter are still blocked by the verifier, and poorly trained small clusters are absorbed by the Reflective Path, which is why full UNO is more robust than UNO-Single in Fig. 11.

2. Joint clustering over all task groups (and Question)

Following your suggestion, we merged the training logs of all four task groups and ran clustering, training, and routing jointly with Qwen3-8B and the same default hyperparameters. The per-task runs produce 2 to 4 clusters each (12 in total, Table 5), and the joint run automatically produces 11 clusters with the same threshold, without any manual adjustment.

Qwen3-8B, Norm-ScoreShort-ShortShort-LongLong-ShortLong-LongAverage
Base72.3574.4346.9463.4164.28
UNO-Single (per task group)75.8276.3646.8163.7765.69
UNO-Single (joint)75.6575.1146.9865.9265.92
UNO (per task group)76.2677.0949.9964.2366.89
UNO (joint)77.4876.3651.0767.7068.15

Joint UNO outperforms Base and all baselines in Table 1 on every task group, raising the four-group average from 66.89 to 68.15 (and from 65.69 to 65.92 for UNO-Single). This suggests positive transfer from feedback patterns shared across tasks, and it confirms that clustering and routing generalize to a more diverse mixed setting without retuning. For much larger scales, App. E discusses two-stage clustering (coarse ANN clustering followed by fine hierarchical clustering) and multi-LoRA serving with vLLM. The results are added to App. L (Table 20).

Thank you again for the helpful suggestions, and we are glad to discuss any further questions.

Reviewer_2DonOpenReview ↗

Thanks for your response and good work

All of my concerns have been addressed (especially regarding the joint UNO), and I believe this work is important and well-conducted. Therefore, I would like to maintain my recommendation to “accept” this work. To avoid any misunderstanding, I would like to clarify that my decision to maintain my recommendation without raising the score does not mean that the authors’ rebuttal failed to address my concerns; rather, it reflects my continued support for the acceptance of this work.

AuthorsOpenReview ↗

Thank you for your continued support of our work

Thank you for taking the time to read our response, and we are glad that our rebuttal and the joint UNO experiment have addressed all of your concerns. Your suggestion has meaningfully strengthened our paper, and we truly appreciate your positive assessment and continued support of our work.

Reviewer_D6uKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces UNO, a framework for improving LLM systems using user interaction logs. UNO distills user logs into structured rules, clusters the resulting query-rules pairs, and assesses the cognitive gap between the model's prior knowledge and each cluster based on semantic distance. Depending on the estimated cognitive gap, either a trained LoRA expert module is used for response generation at inference time, or a trained critic LoRA is used to provide feedback and refine the response generated by the base LLM. Experiments show that UNO outperforms both RAG and memory-based baselines.

Critical strengths
  1. The methodological design is interesting and well-motivated. Dividing user logs into training data for improving the model's generation capability or for criticizing and refining the base model's generations, based on the semantic gap between the logs and the model's prior knowledge, is intuitive and is shown to be effective through the experimental results.

  2. The empirical evaluation is comprehensive. The paper compares UNO against a broad set of baselines, and the analysis covers both effectiveness and efficiency.

Critical weaknesses
  1. UNO depends on model-generated rules based on the query; however, it is unclear how the quality of these generated rules is controlled. Given that these rules are used for both clustering and cognitive gap estimation, if the generated rules are unreliable due to, e.g., the base model capability, the validity of UNO could be compromised.

  2. The potential impact of user-specific preferences is not sufficiently considered. Even for the same query, different users may have different preferences and thus result in different rules; however, during cognitive gap estimation, rule set prediction is based solely on the query. A discussion of how to distinguish user preferences from noise, and how user preferences may affect the effectiveness of UNO, would help address this concern.

  3. The evaluations primarily rely on norm-score and z-score metrics. While the evaluation is comprehensive, a human study with real users would better capture real-world user needs and further validate the effectiveness of UNO.

Questions for authors

Please refer to weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer D6uK (Part 1/2)

Thank you for the careful review and constructive comments. We address each concern below and have updated the manuscript accordingly.

W1: About generated rules and their quality

Thank you for raising this concern. First, we appreciate the opportunity to clarify that UNO uses two kinds of rules. The rules RiR_i, used for clustering, preference pairs, the verifier, and the Critic, are extracted from the full session (qi,yi,Fi)(q_i, y_i, F_i) that contains the user's real feedback, which only requires summarizing requests the user has already made. Only the predicted rules RiLLMR_i^{LLM} are generated from qiq_i alone, and they are used only to compute the cognitive gap. If the base model is weak in a domain and predicts RiLLMR_i^{LLM} poorly, the gap becomes larger and the cluster is routed to the Reflective Path. Errors in RiLLMR_i^{LLM} are therefore fail-safe.

In our submission, we already manually checked rule quality in App. F.9 (50 sessions, 0% unfaithful rules). Following your suggestion, to further verify rule quality across all task groups, we extended this human annotation to 200 sessions (50 per task group, all 15 subsets, 597 rules extracted by Qwen3-8B). Two annotators not involved in this work independently labeled whether each rule is hallucinated and whether each session misses user feedback.

Annotator 1Annotator 2Agreement
Non-hallucinated rules99.3% (593/597)98.8% (590/597)98.5%
Sessions with missed feedback2.5% (5/200)6.5% (13/200)94.0%

Consistent with App. F.9, hallucinated rules are very rare, and missed feedback only reduces how much feedback is used without introducing errors. Existing results in the appendix also show that small rule errors are tolerated. Replacing 10% of the rules with irrelevant ones barely changes performance (Table 11) and a noisy cluster misjudged as low-gap is still blocked by the verifier, which gives consistent results when replaced with GPT-5.2 (Table 13). We have added the new annotation results to App. F.9 and App. I (Table 18).

AuthorsOpenReview ↗

Response to Reviewer D6uK (Part 2/2)

W2: Individual user preferences versus noise

Thank you for raising this question. We have added a discussion of this point in App. J.

UNO targets system-level continual learning rather than user-level personalization, and we have distinguished these two problems in Sec. 2.1. Its goal is to improve the overall performance of the deployed system on future requests, Eq[V(π′(⋅∣q))]\mathbb{E}_{q}[\mathcal{V}(\pi'(\cdot \mid q))] (Sec. 3.1), and all users share the same modules. From this perspective, user feedback contains three components: noise unrelated to the task, needs and procedural knowledge shared across users, and preferences that belong only to individual users. A personal preference is reasonable for that user, but it does not generalize to others, and writing it into the shared weights could hurt their experience. Therefore, just like noise, it should be kept out of the shared weights.

The cognitive gap helps reduce the likelihood that purely personal preferences enter direct weight updates. The gap compares the feedback-derived rules RiR_i with the rules RiLLMR_i^{LLM} generated from the query alone, which represent the model's general prior on the problem. Feedback driven purely by personal taste tends to deviate from this prior and receive a larger gap, making it less likely to enter the Primary Path that updates the weights. Since the gap of a session is the minimum novelty over its rules (App. C.3), a session with at least one general improvement is not flagged as high-risk because of a single personal preference.

The verifier filters preferences at the group level. Each validation session is judged against its own user's rules, and an Expert is deployed only if its win rate over all validation users exceeds γ\gamma, so only improvements that help a clear majority can pass. If preferences conflict within a cluster, an Expert that favors one side stays near a 0.5 win rate and is rejected, and the cluster falls back to the Reflective Path. Moreover, on our experiments on WildFB, whose real users have diverse preferences, both Fig. 3 and our human evaluation show that UNO clearly outperforms the baselines (see W3).

W3: Human evaluation

Thank you for this suggestion. Norm-Score and Z-Score only normalize scores across subsets. The underlying scores are the official metrics of each source dataset (Tables 16 and 17), and many of them were validated against human judgments. For example, the WritingBench critic used by MemoryBench agrees with human preferences 84% of the time [1]. On WildFB (Fig. 3), WildReward-8B was also directly validated against human preferences in its original paper, both on human-annotated preference benchmarks (RewardBench, RM-Bench, and PPE Human) and on a manually verified WildChat user-satisfaction set [2].

To further verify alignment with human judgments on our outputs, we conducted an additional human evaluation on WildFB. Two professional annotators who were not involved in this work blindly ranked three anonymous responses for 150 randomly sampled WildFB test queries from the user's perspective, with ties allowed: the base response, A-Mem (the strongest baseline in Fig. 3), and UNO (Qwen3-8B), in a random order for each annotator. Pairwise win rates (ties count as 0.5) are shown below.

Annotator AAnnotator B
UNO vs. original response0.7700.893
UNO vs. A-Mem0.6300.743
A-Mem vs. original response0.7100.783

Both annotators prefer UNO over the original responses and over A-Mem. The annotators agree on 89.3% of the pairs where both expressed a preference, and their system ranking (UNO > A-Mem > original response) matches WildReward (Fig. 3). We have added these results to App. K (Table 19).

Thank you again for the valuable comments, and we are glad to discuss any further questions.

[1] Wu et al., WritingBench, arxiv: 2503.05244

[2] Peng et al., WildReward, arxiv: 2602.08829

Dear Reviewer D6uK,

We hope this message finds you well, and we truly appreciate the time and effort you have dedicated to reviewing our paper. We completely understand that this is a busy period with many responsibilities.

Whenever your schedule permits, we would be very grateful to hear whether our response and revisions have addressed your concerns, in particular our clarification and expanded human audit of 200 sessions on rule quality (App. I, Table 18), the discussion on individual user preferences (App. J), and the new blind human evaluation on 150 WildFB queries conducted following your suggestion (App. K, Table 19).

If you find that our clarifications and new evaluations have resolved your concerns, we would deeply appreciate it if you could consider raising your score. Otherwise, please let us know if any questions remain, and we will be very happy to clarify further before the discussion period concludes. Thank you once again for your thoughtful feedback.

Reviewer_NU8eOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes UNO, User log-driven Optimisation framework for continually improving the generation performance of LLMs with user logs. It leverages the valuable feedbacks and procedural knowledge from user logs by first converting the unstructured, noisy free texts to semistructured rules. Then clusters are formulated based on the rules as well as the original texts. Cognitive gaps are introduced to decide distance between the LLM’s inherit knowledge and the clusters. Finally, a dual-path mechanism is leveraged for different cognitive gap threshold. Extensive experimental results verify the effectiveness and efficiency with the proposed method compared with other state-of-the-art strategies.

Critical strengths
  1. The paper identifies a fundamental challenge in distinguishing useful feedback from irrelevant, misleading, or potentially harmful signals from raw LLM systems termed as the Signal-or-Noise Dilemma, providing a useful information for developing noise-robust optimisation strategies.
  2. Rather than directly retrieving historical interactions or relying exclusively on dialogue histories, UNO converts unstructured user feedback into actionable rules, which provide an intermediate representation fro preference construction, training and response refinement.
  3. The paper points out that the distinct gap (Cognitive gap in the paper) may exits between the user queries and the LLM inherit knowledge, which is leveraged to guide the choice between direct optimisation and critique-based refinement.
  4. UNO generally improves the aggregate normalized scores over the base models across the four task regimes which is encouraging.
Critical weaknesses
  1. Rule extraction seems to depend heavily on the capabilities and quality of the base LLM. The paper does not sufficiently clarify how the quality an fidelity of extracted rules can be reliably assessed beyond the proposed downstream verification process. Furthermore, extracted rules can vary considerably in specificity: a general rule may apply across multiple queries, whereas a specific rule may be relevant only to a narrow context. How does UNO account for this distinction when calculating the cognitive gap? Could general rules lead to misleading gap estimates or inappropriate cluster assignments? A more detailed analysis of rule quality and its impact on downstream optimization would strengthen the method.
  2. The paper discusses batch-wise adaptation, but the treatment of newly emerging clusters remains insufficiently clear.
  3. Cognitive gap depends on assumptions about the distributions of high-quality and noisy data. Especially, valuable feedback may introduce knowledge beyond the base model's capabilities, while feedback consistent with the model's existing understanding may still be incorrect. A discussion may need to strengthen this contribution.
  4. The use of John Dewey’s philosophy is somewhat far-fetched, since CoT can also be considered to be related to John Dewey’s philosophy.
  5. What are the differences between UNO-Single and SFT? As without reflective module, only conservative lora is kept. But the performance seems very different.
  6. Some important related papers are missing: a. Tucker, A.D., Brantley, K., Cahall, A. and Joachims, T., 2024, July. Coactive Learning for Large Language Models using Implicit User Feedback. In ICML (pp. 48809-48822). b. Ye, T., Dong, L., Dong, Q., Wu, X., Huang, S. and Wei, F., 2026. Online experiential learning for language models. arXiv preprint arXiv:2603.16856.

Minor: The two asterisks make the results look a bit messy; I suggest removing them.

Questions for authors

Please see my explanation in the section on weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer NU8e (Part 1/2)

Thank you for the careful review and thoughtful questions. We respond to each point below and have updated the manuscript accordingly.

W1: Rule quality, and general versus specific rules

Thank you for raising this question. In our submission, App. F.9 already includes a manual check of rule quality (50 samples, 0% unfaithful rules). To further verify rule quality across all four task groups, we extended this human annotation to 200 sessions (50 per task group, all 15 subsets, 597 rules extracted by Qwen3-8B; App. I, Table 18). Two independent annotators not involved in this work labeled 99.3% (593/597) and 98.8% (590/597) of the rules as non-hallucinated (98.5% agreement), and found missed user feedback in only 2.5% (5/200) and 6.5% (13/200) of sessions (94.0% agreement). Together with Table 11 (10% irrelevant rules barely change performance) and Table 10 (the gains disappear without real user rules), this shows that the rules are faithful and that small errors are tolerated.

General and specific rules do not need to be treated differently. The gap is computed within a single query. Each extracted rule is compared only with the rules RiLLMR_i^{LLM} predicted for the same query (Sec. 3.3.3, App. C.3), so whether it also applies to other queries plays no role. The gap measures novelty relative to the model's prior rather than generality. A general rule that the model would produce itself (for example, "keep the structure clear") is already within its knowledge, so a low gap is correct, while a general rule that the model would not produce still receives a high gap. The same holds for specific rules. Empirically, the cluster-level gap predicts training failure well (Table 7). For clustering, query and rule embeddings are weighted equally (App. A.2.1) and the query part anchors the domain, so general rules do not merge different domains.

W2: New clusters in batch-wise adaptation

Thank you for asking about this detail. As described in Sec. 4.3 of the submission (with criteria in App. C.3), we perform clustering over all user logs and adopt new centroids if cluster counts change, then retrain using the full logs. Otherwise, we keep the original centroids and perform continual training. New clusters appear when all logs are re-clustered. If the number of clusters changes, every cluster, including the new ones, goes through the cognitive gap assessment, training, and verifier gating again on the full logs. Otherwise, the original centroids are kept and training continues on the new logs.

W3: Assumptions behind the cognitive gap

Thank you for raising this thoughtful point. Both cases you describe are closely related to the Signal-or-Noise Dilemma in Sec. 3.3.3, and the cognitive gap is designed as a probabilistic tool rather than a per-sample judgment. Under Assumption 3.1, high-quality feedback is more likely than noisy feedback to align with the model's prior in the low-gap region. Given this assumption, Proposition 3.2 provides a conditional upper bound on the posterior noise probability P(N∣gi≤τ)P(N \mid g_i \leq \tau). In the high-gap region, valuable knowledge beyond the model's capability and noise look alike to the model, so the risk cannot be controlled. UNO therefore neither discards such feedback nor writes it into the weights, but uses it through the Critic in the Reflective Path. Feedback that agrees with the model's prior but is wrong corresponds to a low-probability event under this conditional bound.

This motivation is supported by prior findings. Dataset cartography (Swayamdipta et al., 2020, cited in App. B.1) shows that hard-to-learn regions mix mislabeled and genuinely hard samples that a model cannot tell apart, and OEL (Ye et al., 2026) finds that knowledge from a stronger model helps a smaller model less, likely because it exceeds the smaller model's capability. Finally, the gap is only a pre-filter. An Expert that does not beat the base model on validation is rejected, and such rejected clusters indeed bring no downstream gain (App. F.1, Table 8). UNO's consistent gains over all baselines and the correlation between the gap and training failure (Table 7) further support this design.

AuthorsOpenReview ↗

Response to Reviewer NU8e (Part 2/2)

W4: Dewey's philosophy

Thank you for this suggestion. The analogy only provides the names "primary" and "reflective". We have softened it to "loosely inspired by" and moved it to a footnote in revision.

W5: UNO-Single versus SFT

Standard SFT cannot be applied to raw logs directly, because the logs contain no corrected target responses. The SFT baseline in Table 1 is SFT w/ UNO Data (Sec. 4.2, App. C.2), which already uses the improved responses from UNO's pipeline. Even so, the two methods differ substantially.

  • SFT trains one global adapter, while UNO-Single trains one Expert LoRA per cluster. Each query is routed to its cluster at inference time, and Table 2 shows the drop without clustering.
  • SFT writes all data into the weights, while UNO-Single filters out noise. UNO-Single pre-filters risky clusters with the cognitive gap and deploys an Expert only if it beats the base model under the verifier, otherwise using the base model (App. F.1, F.3).
  • The training objective is different. SFT uses NLL and UNO-Single uses DPO + NLL by default.

As a result, global SFT is hurt by noisy logs and falls below the base model on several tasks, while UNO-Single keeps the gains of useful clusters and isolates the noisy ones (Table 1).

W6: Related work

Thank you for pointing out these two papers. We have added both to Sec. 2.3. CoRLL (Tucker et al., 2024) treats user edits as weak preferences (yˉ≻y\bar{y} \succ y) for online DPO. UNO builds preference pairs in a similar spirit, but its feedback is free-form text that may be irrelevant or noisy, so it first distills rules and verifies before updating the weights. OEL (Ye et al., 2026) distills experiential knowledge from trajectories into the weights through on-policy context distillation. Like UNO, it finds distilled knowledge more useful than raw trajectories and finds that knowledge from a stronger model helps less, but it learns from clean environment feedback in single-task games, whereas UNO handles noisy and heterogeneous human feedback across many tasks. OEL is also close to SDPO (Buening et al., 2026), which we already compare with (Tables 1 and 9), as both train the model to match itself conditioned on extra context (the user's follow-up in SDPO and accumulated knowledge in OEL).

Thank you again for the constructive comments, and we are happy to answer any further questions.

I thank the authors for their responses to my questions and concerns. The clarifications provided have adequately addressed the issues raised in my review. Overall, I am satisfied with the authors' responses and believe that the paper makes a valuable contribution. I therefore maintain my decision as "accept".

AuthorsOpenReview ↗

Thank you for your feedback and continued support

Thank you very much for taking the time to read our response and for confirming that our clarifications and revisions have addressed your concerns. We truly appreciate your thoughtful feedback and continued support of our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Area Chairs and Reviewers,

Thank you very much for your time and constructive reviews. We are grateful that all reviewers have confirmed that their concerns were addressed and support our work, and we especially appreciate their recognition of our study as an important and valuable contribution. For ease of reference, we briefly summarize our key clarifications, new experiments, and blue highlighted manuscript revisions below:

  1. Rule Quality Audit (Reviewer D6uK, Reviewer NU8e): Clarified that rules RiR_i are distilled from real user feedback while RiLLMR_i^{LLM} only computes the cognitive gap in a fail safe way. Beyond the 50 session check in App. F.9 and noise ablations (Table 11), a new blind audit over 200 sessions across all 4 task groups found 99.3% / 98.8% non hallucinated rules (App. I, Table 18).
  2. WildFB Human Evaluation (Reviewer D6uK): Clarified that underlying source metrics and WildReward were already validated against human preferences. In a new blind human evaluation, two independent annotators preferred UNO over original responses and over A-Mem with high agreement (App. K, Table 19).
  3. Hyperparameter Generalization & Joint Clustering (Reviewer 2Don): Explicitly noted in Sec. 4.3 that all settings share identical default hyperparameters without tuning. Merging all four MemoryBench task groups under the same threshold automatically yielded 11 clusters and raised the four group average Norm-Score from 66.89 to 68.15 (App. L, Table 20).
  4. Further Discussions, Clarifications, and Related Work (Reviewer D6uK, Reviewer NU8e): Added App. J on filtering personal preferences via the cognitive gap and group level verifier, clarified batch wise cluster updates, the assumptions behind the cognitive gap, and UNO-Single vs. SFT, added CoRLL and OEL to Sec. 2.3, and moved the softened naming analogy to a footnote in Sec. 1.

Thank you once again to the Area Chairs and Reviewers for helping us strengthen our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.