acceptodds
Under review as a conference paper at ICLR 2027

SAGE: Online Continual Knowledge Distillation for Small Language Models

Abstract

Small language models are increasingly deployed in resource-constrained settings because of their computational efficiency, yet their limited capacity often leads to performance degradation when they encounter previously unseen data. Although knowledge distillation can transfer capabilities from larger teacher models, conventional approaches are typically confined to offline training and therefore cannot provide continuous supervision after deployment. Here, we introduce SAGE, an online continual distillation framework that enables a deployed small language model to selectively acquire knowledge from a teacher model in streaming environments. SAGE comprises three coordinated modules. The Trigger identifies inputs beyond the student’s competence boundary by combining teacher–student discrepancies without requiring labels. The Buffer uses streaming clustering and stability filtering to organize triggered samples into semantically coherent knowledge regions. Finally, the Debugger performs a two-stage search over low-rank parameter updates to derive effective adaptation configurations for targeted knowledge injection. Experiments show that SAGE substantially improves both sample efficiency and adaptation accuracy, providing a practical framework for continuously enhancing small language models after deployment.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Online Distillation, Low-Rank Adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WKtZOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper frames online continual distillation as a post-deployment learning problem, introducing SAGE: a closed-loop framework that identifies competence-boundary violations, clusters emerging deficiencies into stable knowledge regions, and injects targeted teacher supervision via efficient low-rank adaptation.

Critical strengths
  • The paper identifies and solves the boundary saturation for distilling SLMs in streaming environments.
  • The proposed SAGE integrates a label-free Trigger, streaming clustering Buffer, and a two-stage low-rank adaptation Debugger to handle streaming data without retraining the whole model.
  • Extensive experiments demonstrate that SAGE improves performance and enables cumulative gains across multiple models.
Critical weaknesses
  • The performance relies heavily on the quality of the teacher signal, which means noisy teacher feedback can degrade downstream adaptation. The three benchmarks used in the article is limited and lacks evaluation for complex reasoning tasks.
  • While real-time triggers and buffering add minimal latency, the Debugger's cluster optimization remains computationally expensive
  • The presentation of the experimental results feels unstructured and lacks a cohesive layout.
Questions for authors
  • SAGE combines standard anomaly detection, off-the-shelf clustering, and LoRA tuning. What is the distinct breakthrough over a naive pipeline coupling these off-the-shelf modules, and why is this specific design uniquely necessary for LLMs?
  • How is SAGE generalized to the complex tasks? In a realistic post-deployment streaming scenario where user inputs arrive as unstructured raw text, how can SAGE guarantee task atomicity without an expensive, pre-processed teacher pipeline?
4 comments
AuthorsOpenReview ↗

Response: Framework Novelty

We thank the reviewer for the SAGE's framework-level novelty, teacher dependence, and generalization to complex tasks. And we provide additional CoT and preprocessing results.

Q1: What distinguishes SAGE from a naive combination of existing modules, and why is its design necessary for LLMs?

Response to Q1. We agree that anomaly detection, clustering, and LoRA are established techniques. However, SAGE's contribution lies in coordinating these components to transform deployment-time experiences into persistent model capabilities, rather than simply combining existing algorithms (Sec. 1, p. 2, L070–089). The key distinctions are twofold:

  • Framework-level novelty. SAGE establishes a coordinated Detect–Organize–Inject process: Trigger identifies what to learn through teacher–student discrepancies, Buffer determines when to adapt through cluster stability, and Debugger determines how to update through cluster-specific LoRA optimization (Sec. 3, pp. 4–7).

  • Necessity for LLMs. Conventional offline distillation leaves deployed SLMs unable to adapt when evolving inputs exceed their learned capabilities. Addressing this boundary saturation requires three capabilities: detecting emerging deficiencies, distinguishing persistent knowledge gaps from transient noise, and incorporating targeted knowledge without retraining the entire model (Sec. 1, p. 2, L054–089). SAGE realizes these requirements through its three coordinated modules(Trigger, Buffer and Debugger), enabling SLMs to learn from deployment-time interactions and progressively acquire domain-specific capabilities under distribution shifts. Table 3 (p. 8) further supports this design: Debugger improves EM from 80.64% to 94.65% over fixed LoRA with the same 315 samples.

W1: How does teacher signal quality affect SAGE's adaptation performance?

Response to W1. We acknowledge that SAGE's adaptation performance depends on teacher quality. We use ground-truth supervision to simulate an ideal teacher and estimate the performance upper bound (Sec. 4.1, p. 7, L351–356). Table 5 (p. 17) shows that GSM8K EM decreases from 99.80% with GT to 88.65% with GPT-5 and 73.06% with DeepSeek. This gap partly reflects limitations in current teacher capabilities, which may narrow as LLMs improve, while robustness to noisy supervision remains future work.

AuthorsOpenReview ↗

Response: Complex-Task Generalization, and Task Atomicity**

W1 and Q2.1: Can SAGE generalize to more complex reasoning tasks?

Response to W1 and Q2.1. We provide three complementary results to examine SAGE's generalization to complex reasoning tasks, covering experience selection, progressive capability acquisition, and structured adaptation:

  1. Cross-domain experience selection (Appendix C, p. 16). Trigger serves as the gatekeeper for knowledge acquisition, determining which incoming experiences enter the Buffer for subsequent adaptation. On EntailmentBank, WorldTree, and ARC, it achieves approximately 1.0 AUC using TriviaQA-calibrated parameters (Figure 9), supporting its ability to identify emerging capability gaps across domains.

  2. Progressive reasoning adaptation (Figure 3, p. 8). Curriculum-based distillation enables SAGE to acquire increasingly complex reasoning capabilities. Level-2 reasoning EM improves from nearly 0% to approximately 81%, while previously learned Level-1 performance remains above 99%, demonstrating capability expansion with knowledge retention.

  3. Adaptation to complex tasks (Table A). Since complex reasoning may involve entangled supervision, we use CoT decomposition to construct atomic learning experiences (Appendix E, p. 24). We further evaluate SAGE and baselines under matched supervision, as shown in Table A. SAGE (GT) with CoT achieves 95.70% and 87.14% accuracy on PubMedQA and LexGLUE, outperforming CA-LoRA with identical CoT supervision by 7.41 and 9.05 percentage points, respectively.

Table A. CoT Supervision and SAGE Adaptation (LLaMA2-7B).

MethodTraining DataPubMedQA AccPubMedQA F1LexGLUE AccLexGLUE F1
LoRAOriginal66.6064.1335.2935.29
LoRACoT72.9172.0642.8441.97
CA-LoRAOriginal84.8084.7072.2175.50
CA-LoRACoT88.2987.6778.0978.92
LwFOriginal58.2056.4539.8038.50
LwFCoT65.5464.4246.3145.47
SAGE (GPT)Original78.0378.1175.4674.96
SAGE (GT)Original87.5187.3084.2184.08
SAGE (GPT)CoT84.7383.6777.2277.87
SAGE (GT)CoT95.7095.3787.1487.63

Together, these results support SAGE's potential to extend to more complex tasks through transferable experience selection, progressive learning, and structured knowledge adaptation.

Q2.2: How does SAGE ensure task atomicity from unstructured streaming inputs without costly preprocessing?

Response to Q2.2. We clarify that CoT decomposition is a one-time, task-specific preparation step, rather than a pipeline executed for every incoming query. Specifically, we first identify task-relevant examples through keyword matching, then apply LLM-assisted CoT decomposition to construct atomic reasoning steps (Appendix E, p. 24, L1250–1259). We evaluate its cost and quality in Table B below. This approach achieves 97.50% and 94.70% atomicity on PubMedQA and LexGLUE, demonstrating its applicability across both domains.

Table B. Cost and Quality of Task-Specific CoT Decomposition.

DatasetSource SamplesTraining SamplesPreprocessing Time (s)LLM Time (min)Atomicity (%)
PubMedQA211,000780244597.50
LexGLUE45,00064483894.70

Also, decomposition makes complex tasks more manageable for SLMs by providing fine-grained supervision. For less-structured tasks, curriculum-based incremental adaptation, as demonstrated on GSM8K (Figure 3, p. 8), offers a complementary strategy for progressively introducing more challenging reasoning tasks.

AuthorsOpenReview ↗

Gentle Reminder: Request for Feedback on Our Rebuttal

Thank you again for your thoughtful review. We have posted a detailed response addressing each of your concerns, together with the additional experiments and clarifications you requested. As the discussion period closes in 1 days, we would greatly appreciate your feedback on whether our responses resolve your questions. We are happy to answer any remaining concerns promptly.

If our revisions have addressed your main points, we would be grateful if you could consider updating your assessment.

Thank you for your time and effort.

AuthorsOpenReview ↗

Gentle Reminder

With only 9 hours remaining in the discussion period, we would be glad to answer any further questions you may have. If our responses have resolved your main concerns, we would appreciate your consideration of an updated assessment. Thank you for your time and constructive feedback.

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

In this work, the authors propose SAGE, an online continual knowledge distillation framework designed to mitigate post-deployment boundary saturation in small language models (SLMs). SAGE operates via a three-module closed loop: a Trigger module that identifies competence-boundary violations without ground-truth labels, a Buffer module that organizes streaming anomalies into stable knowledge clusters via density clustering and stability filtering, and a Debugger module that conducts a two-stage hyperparameter search to inject teacher knowledge into modular LoRA adapters. Evaluated across multiple model families (Qwen2.5, LLaMA-2, Mistral) and domains (GSM8K, PubMedQA, LexGLUE), SAGE demonstrates consistent gains in adaptation accuracy and sample efficiency while effectively preventing catastrophic forgetting.

Critical strengths

 The problem of adapting deployed SLMs to streaming distribution shifts online is practical and well-motivated.  The three-stage pipeline (Trigger, Buffer, Debugger) is intuitive, and isolating updates into cluster-specific LoRA adapters is a clean way to avoid catastrophic forgetting.  The empirical evaluation is extensive, covering five student backbones, five teacher sources, and thorough ablations on each module.

Critical weaknesses

 The Trigger formulation contradicts the paper's motivation of selective teacher invocation. While the introduction states the teacher is invoked only when the student needs help, computing the trigger score AS_t in Section 3.1 requires generating the full teacher response y_t^T for every incoming sample to calculate BLEU, ROUGE-L, and embedding similarity. This negates the efficiency benefit of deploying an SLM, and the 24.01 ms trigger latency in Table 2 clearly omits teacher inference time.  Since the Debugger saves a separate LoRA adapter for each stable cluster, the number of adapters could grow substantially in long open-ended streams. The paper does not analyze memory overhead or potential routing degradation as the adapter pool scales to hundreds or thousands, nor is it clear how unlabeled test queries are routed without oracle source tags.  Although the Debugger runs asynchronously, training N random configurations plus up to 9k grid refinements per cluster entails substantial compute, which seems at odds with the resource-constrained scenarios motivating this work.

Questions for authors

 How can the Trigger be evaluated in 24.01 ms if the teacher must generate y_t^T beforehand? Does this include teacher inference/API latency, and if not, what is the true end-to-end trigger latency?  If the teacher already generates responses for every input x_t to compute the Trigger score, what is the practical incentive for deploying the student model rather than directly serving the teacher's answers?  What is the expected memory and retrieval overhead when the system accumulates a large number of adapters over long deployments, and does routing accuracy degrade as cluster density increases?  How does the system handle continuous concept drift where previously learned clusters become obsolete or conflict with newly observed streaming data?

3 comments
AuthorsOpenReview ↗

Response: Clarification of Trigger Cost and Deployment Motivation

We thank the reviewer for the feedback. We clarify the teacher-dependent latency and deployment motivation of SAGE, and provide additional analyses of adapter scalability, Debugger computation, and distribution drift.

W1 and Q1-Q2: The Trigger requires teacher responses for every input. Does the reported 24.01 ms include teacher inference, and what is the actual end-to-end latency?

Response to W1 and Q1–Q2. We acknowledge that the reported 24.01 ms Trigger latency measures only local anomaly-score computation, excluding teacher inference and API latency. We therefore conduct an additional end-to-end efficiency evaluation, with results reported in Table A below. GPT-SAGE incurs 2408.01 ms mean latency (P95: 2579.48 ms), including 1054.40 ms of teacher latency, compared with 1168.64 ms for student-only inference. These costs are now reported ( Table 2 (a) p. 7).

Table A. End-to-End Latency and Deployment Cost Comparison.

Component / StrategyLocal Latency (ms)Teacher Latency (ms)E2E Mean (ms)E2E P95 (ms)Teacher Calls / 1KAPI Cost / 1K ($)
Student Generation1168.6401168.642361.6500
Teacher Generation01054.40--10000.026
Student Only1168.6401168.642361.6500
GPT-SAGE1353.61†1054.402408.012579.481000†0.026
GT-SAGE1353.6101353.612509.7200

W1 and Q3: If the teacher is invoked for every input, what is the practical benefit of deploying an SLM instead of directly serving the teacher?

Response to W1 and Q3. Response to W1 and Q3. We acknowledge that SAGE aims to enable continual capability improvement of deployed SLMs rather than accelerate individual requests, while reducing teacher dependence during adaptation remains future work. Our motivation rests on two aspects:

  • Learning from deployment-time experience. Conventional offline training cannot continuously incorporate new deficiencies exposed by distribution shifts. SAGE addresses this gap by selectively collecting informative experiences through Trigger, organizing persistent deficiencies through Buffer, and consolidating teacher knowledge into reusable adapters through Debugger (Sec. 1, p. 2, L070–089). Unlike direct teacher serving, this process transforms deployment-time feedback into persistent model improvements.

  • Practical value of SLM deployment. SLMs are more suitable for local, resource-constrained, and domain-specific applications where deploying large models may be impractical. Through targeted adaptation, SAGE enables specialized SLMs to accumulate domain knowledge and reuse acquired capabilities during subsequent student-only inference (Sec. 3.3, p. 7, L330–334).

AuthorsOpenReview ↗

Response: Adapter Scalability, Debugger Computation, and Distribution Drift

W2 and Q4.1: How do memory overhead, retrieval latency, and routing accuracy scale as the number of clusters and adapters increases?

Response to W2 and Q4.1. SAGE does not keep all LoRA adapters in memory. Instead, an in-memory metadata pool maintains cluster embeddings, keywords,(Sec. 3.2, p. 5, L233–241, Eq. 9) and adapter information for query routing (Sec. 3.3, p. 7, L331–332), while adapter checkpoints are stored externally and loaded on demand. This avoids keeping the entire adapter pool in memory. Adapter routing requires only 9.62 ms (Table 2, p. 7). Therefore, memory overhead is not the primary bottleneck for scalability as the adapter pool grows.

W2 and Q4.2: Does routing accuracy degrade as the number of accumulated adapters and cluster density increase?

Response to W2 and Q4.2. We acknowledge that routing accuracy degrades as the number of clusters and adapters increases. We therefore evaluate routing scalability across different adapter pool sizes, with results reported in Table B below. As the pool grows from 30 to 1500 adapters, routing accuracy decreases from 100% to 33.49%, while P95 routing latency increases from 12.90 to 21.29 ms and adapter switching remains stable. These results indicate that routing accuracy, rather than retrieval overhead, is the primary scalability limitation. The results are also reported (Table 8 p. 23). This limitation may be mitigated through hierarchical routing and cluster pruning, which we plan to investigate in future work.

Table B. Scalability of Cluster Routing and Adapter Retrieval.

ClustersAdaptersRoute Acc (%)P95 Route (ms)P95 Switch (ms)Storage (GB)
1030100.0012.908.780.3099
5015087.5514.038.951.7021
10030068.7219.129.133.6147
500150033.4921.299.2516.9834

Q5: How does SAGE handle concept drift in the presence of conflicting or outdated clusters?

Response to Q5. SAGE addresses concept drift through cluster-aware adapter replacement. It monitors embedding and cluster-center deviations, replacing outdated adapters only when newly accumulated experiences stabilize; otherwise, existing adapters are retained. Unmatched samples form new clusters for adaptation (Sec. 3.2, p. 5; Sec. 4.2, p. 8, L411–420). Figure 4 (p. 8, L421–430) validates this mechanism: following distribution shift, SAGE recovers current-domain accuracy from approximately 50% to 91%, while the variant without replacement declines to 45%. Meanwhile, old-domain accuracy remains at 85–90%, demonstrating effective drift adaptation with knowledge retention.

AuthorsOpenReview ↗

Gentle Reminder: Request for Feedback on Our Rebuttal

Thank you again for your thoughtful review. We have posted a detailed response addressing each of your concerns, together with the additional experiments and clarifications you requested. As the discussion period closes in 1 days, we would greatly appreciate your feedback on whether our responses resolve your questions. We are happy to answer any remaining concerns promptly.

If our revisions have addressed your main points, we would be grateful if you could consider updating your assessment.

Thank you for your time and effort.

Reviewer_ef33OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAGE adapts small language models continually using teacher feedback. It detects teacher-student disagreement, groups selected examples into semantic clusters, and learns cluster-specific LoRA adapters for later retrieval. The experiments cover mathematical, biomedical, and legal tasks with multiple student models and both ideal and model-generated teacher signals. They also include component ablations and cross-round retention tests.

Critical strengths
  1. Delayed clustering and separate adapters give the framework a plausible way to specialize repeatedly while limiting interference with previously acquired capabilities (§3). The cross-round evaluation tests this motivation.

  2. The evaluation includes actual teacher outputs, smaller students, component ablations, and retention on held-out tasks. Table 3 compares adaptation methods using the same 315 selected examples. This control helps distinguish the effect of adaptation from the effect of data selection.

Critical weaknesses
  1. The accuracy and extraction metrics need to be reconciled. In Table 3, the GPT All-to-Buffer setting has 81.27% EM and 63.13% Numeric Extraction Rate; Table 6 includes 100% EM with 98.79% NER. If both rates use the same test set and a correct numerical answer requires successful extraction under the same rule, EM cannot exceed NER. Different denominators or parsers could explain the values, but §4.1 does not specify them. Extraction rates vary across methods, so accuracy conditional on successful extraction cannot be read as accuracy over all test inputs. Until the scoring rule is clear, the main performance comparisons are difficult to interpret.

  2. The biomedical and legal comparisons leave the source of the gains unclear. Section 4.1 applies CoT decomposition to SAGE for PubMedQA and LexGLUE and describes the comparison methods as using the original datasets or static clustering. Appendix E states that this preprocessing substantially improves SAGE. The description leaves unclear whether the methods are evaluated on identical tasks and how much of the gain comes from adaptation versus the additional structured supervision. The GSM8K data-matched experiment does not answer these questions for the other two domains.

Questions for authors
  1. What are the exact EM and NER definitions, denominators, and extraction rules? Please reconcile the examples above and clarify whether the main EM results measure accuracy over every held-out question, counting extraction failures as incorrect. If they do not, can that accuracy be calculated from the existing outputs?

  2. For PubMedQA and LexGLUE, are all methods evaluated on the same original held-out questions, with splitting performed before decomposition and related derivatives kept together? Are results already available from a baseline using the same decomposed training examples and teacher targets? Please explain what the available evidence establishes about the individual modules as well as the complete pipeline.

5 comments
AuthorsOpenReview ↗

Response: Clarification of EM, NER, and GSM8K Evaluation

We thank the reviewer for the constructive feedback. We clarify the GSM8K evaluation metrics and provide CoT-matched comparisons and component ablations on PubMedQA and LexGLUE.

W1 and Q1: The reported EM exceeds NER in some settings, raising concerns about metric definitions, how are these results reconciled?

Response to W1 and Q1. We clarify that the originally reported EM measures accuracy conditional on successful numerical extraction (Sec. 4.1, p. 7, L358–361). Let NN, MM, and CC denote the numbers of test questions, successfully extracted predictions, and correct predictions, respectively. The metrics are defined as

EMcond=CM,NER=MN,EMfull=CN=EMcond×NEREM_{cond} = \frac{C}{M}, \quad NER = \frac{M}{N}, \quad EM_{full} = \frac{C}{N} = EM_{cond} \times NER

Also, the specific extraction rules implemented in our evaluation code are as follows: the parser first matches the "final answer is" pattern and otherwise searches the last five generated lines for a valid numerical answer. Extraction failures are excluded from conditional EM and MAE. For example, GPT All-to-Buffer (Table 3, p. 8) has N=499N=499, M=315M=315, and C=256C=256, yielding 81.27% conditional EM, 63.13% NER, and 51.30% full-set EM. GPT w/o Buffer achieves 41.48% full-set EM rather than 72.13%. Similarly, the reported 100% EM with 98.79% NER (Table 7, p. 22) corresponds to 98.79% full-set EM. Overall, we report EM and NER separately to distinguish numerical correctness from answer extractability.

AuthorsOpenReview ↗

Response: Disentangling the Contributions of CoT Supervision and SAGE Adaptation

Q2.1: How are PubMedQA and LexGLUE split, and are all methods evaluated on the same held-out questions?

Response to Q2.1. We clarify that both datasets are split into training, validation, and test sets at the original-question level before CoT decomposition (Sec. 4.1, p. 7, L366–367). CoT decomposition is applied only to training data (Appendix E, p. 24, L1254–1260), and all derived examples from the same original question remain within the same split, preventing cross-split leakage. All methods are evaluated on the same original, undecomposed held-out questions under identical evaluation protocols.

W2 and Q2.2: Are SAGE's gains attributable to CoT supervision or its adaptation mechanisms? What do matched-supervision comparisons and component ablations establish?

Response to W2 and Q2.2. We conduct two additional experiments on PubMedQA and LexGLUE: **CoT-matched baseline comparisons (Table 9, p. 23) and component-level ablations (Table 6, p. 19) **.

In the first experiment, LoRA, CA-LoRA, LwF, and SAGE use identical CoT-decomposed training examples and GT targets. CoT supervision improves LoRA accuracy by 6.31 and 7.55 points on PubMedQA and LexGLUE, respectively, confirming the benefit of structured supervision. Under matched supervision, SAGE further outperforms CA-LoRA by 7.41 and 9.05 points, demonstrating that its gains extend beyond CoT preprocessing (Table A).

Table A. Disentangling CoT Supervision and SAGE Adaptation (llama2-7B, GT).

MethodTraining DataPubMedQA AccPubMedQA F1LexGLUE AccLexGLUE F1
LoRAOriginal66.6064.1335.2935.29
LoRACoT72.9172.0642.8441.97
CA-LoRAOriginal (Static Clustering)84.8084.7072.2175.50
CA-LoRACoT88.2987.6778.0978.92
LwFOriginal58.2056.4539.8038.50
LwFCoT65.5464.4246.3145.47
SAGE(GPT)Original78.0378.1175.4674.96
SAGE(GT)Original87.5187.3084.2184.08
SAGE (GPT)CoT84.7383.6777.2277.87
SAGE (GT)CoT95.70↑95.37↑87.14↑87.63↑

In the second experiment, ablations under matched CoT supervision (Table 6, p. 19) show that All-to-Buffer and removing Buffer reduce accuracy to 89.74%/85.67% and 88.54%/85.78%, respectively, highlighting selective admission and structured organization. Replacing Debugger with LoRA causes the largest drops of 22.12/18.90 points, confirming the importance of adaptive updates. Consistent with Table 3 (p. 8), these results demonstrate the complementary roles of all three modules.

Table B. Component Ablation on CoT-Decomposed PubMedQA and LexGLUE (llama2-7B, GT).

MethodPubMedQA AccPubMedQA F1LexGLUE AccLexGLUE F1
w/o Trigger (Random, GT)49.8033.2414.3313.55
w/o Trigger (Random, GPT)48.5021.1613.7813.07
w/o Trigger (All-to-Buffer, GT)89.7489.2785.6785.98
w/o Trigger (All-to-Buffer, GPT)82.4381.6876.2475.87
w/o Buffer (GT)88.5487.8785.7885.05
w/o Buffer (GPT)81.9281.4475.4175.03
w/o Buffer (HDBSCAN, GT)90.9390.4886.6286.48
w/o Buffer (HDBSCAN, GPT)82.6682.1476.9076.37
w/o Debugger (LoRA, GT)73.5873.4468.2468.72
w/o Debugger (LoRA, GPT)65.4364.8662.7162.20
SAGE (GT)95.70↑95.37↑87.14↑87.63↑
SAGE (GPT)84.7383.6777.2277.87
AuthorsOpenReview ↗

Gentle Reminder: Request for Feedback on Our Rebuttal

Thank you again for your thoughtful review. We have posted a detailed response addressing each of your concerns, together with the additional experiments and clarifications you requested. As the discussion period closes in 1 days, we would greatly appreciate your feedback on whether our responses resolve your questions. We are happy to answer any remaining concerns promptly. If our revisions have addressed your main points, we would be grateful if you could consider updating your assessment. Thank you for your time and effort.

Comments on the paper

1 comment

Thank you both for your feedback (since you use "we" in your review, we also use "both" here to refer to both you and your AI assistant).

In your previous review, you stated that our work "is practical and well-motivated," yet in this round you question its practical value, which is highly contradictory.

May I ask which specific point led you to make this contradictory statement?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.