acceptodds
Under review as a conference paper at ICLR 2027

Can Generative Retrievers Learn Semantic IDs Without Forgetting How to Speak?

Abstract

Generative retrieval (GR) enables end-to-end retrieval by generating document semantic identifiers (SIDs). However, retrieval-only fine-tuning can over-specialize pretrained language models to SID prediction, substantially distorting their natural-language distribution and limiting their suitability for interactive systems that must both retrieve documents and generate natural-language responses. We introduce SpeakGR, a dual-objective framework that learns SIDs while preserving language generation. It combines supervised SID learning with speak-preserving regularization: an on-policy distillation objective that aligns the current model with a frozen copy of the original model on student-generated prefixes using forward KL over the original text vocabulary. We further propose Adaptive SpeakGR, which dynamically adjusts the preservation strength based on observed language drift. Compared with SFT-only, SpeakGR reduces WikiText-2 forward KL by 81.3-93.8% on MS MARCO and 81.2-85.2% on Natural Questions (NQ) while retaining effective retrieval across three different LLMs. Adaptive SpeakGR further improves retrieval over SpeakGR in most settings while maintaining substantially lower language drift than SFT-only.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generative retrieval, language preservation, semantic IDs, catastrophic forgetting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zS2jOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies an interesting problem in generative retrieval: learning semantic IDs can harm the original language generation ability of LLMs. To address this issue, the authors propose SpeakGR, which combines semantic ID learning with a language-preserving objective. They further introduce Adaptive SpeakGR, which dynamically adjusts the preservation strength during training. Experiments across different models and datasets show that the proposed methods can significantly reduce language drift while maintaining competitive retrieval performance.

Critical strengths
  1. Preserving language generation capabilities during semantic ID learning is an important but relatively underexplored problem.
  2. The proposed method is simple, intuitive, and well motivated. I appreciate that such a straightforward approach can effectively address the problem without requiring major changes to the model architecture.
  3. The experiments cover multiple models and datasets, showing clear improvements in language preservation. The comparisons with existing methods and additional analyses also provide solid support for the proposed approach.
  4. I believe this work has broader implications for generative retrieval research. It highlights that retrieval performance alone may not be sufficient to evaluate generative retrievers, and that preserving the original language capabilities of LLMs is also important.
  5. This paper also opens up a promising direction for building unified retrieval and generation models. The ability to learn retrieval without significantly affecting language generation could be valuable for future LLM-based information retrieval systems.
Critical weaknesses
  1. It would be interesting to explore whether the proposed method can be extended to other semantic ID-based tasks, such as generative recommendation, in future work.
  2. The experiments mainly focus on relatively small LLMs. It would be interesting to see whether similar results can be achieved with larger models.
  3. A few qualitative examples of language generation before and after fine-tuning could help readers better understand the benefits of the proposed method.
Questions for authors
  1. As a possible future direction, have the authors considered evaluating more complex language tasks, such as reasoning, long-form generation, or multi-turn conversations?
  2. Could the authors briefly discuss how the proposed method might be used in a unified retrieval-and-generation framework?
  3. Do the authors expect similar results with larger LLMs or different semantic ID construction methods?
3 comments
AuthorsOpenReview ↗

Response to Reviewer zS2j

We sincerely thank the reviewer for the encouraging assessment of our motivation and contributions. We appreciate your recognition of language preservation as an important, underexplored problem, the simplicity and effectiveness of our approach, and the experimental support across models and datasets. We also thank you for highlighting the broader value of evaluating language preservation alongside retrieval and the potential for unified retrieval and generation. Your constructive suggestions motivated the larger-model experiments and qualitative examples below.

Response to Weaknesses

W1. Generative recommendation.

A natural extension is to predict item SIDs from user interaction histories while retaining the same speak-preserving regularization over the original text vocabulary. We added this future direction to Appendix B.

W2. Larger models.

We added Qwen3-8B/NQ experiments. To train the larger backbone efficiently, we used LoRA for both SFT and SpeakGR, with matched settings and preservation weight 1.

ModelR@10 (%)Forward KLPPL
Base—0.00014.06
SFT77.423.816319.93
SpeakGR71.370.09312.56

SpeakGR reduces language-distribution drift, with R@10 lower by 6.05 percentage points. These results extend the evidence for language preservation to an 8B backbone under LoRA (Appendix G.2).

W3. Qualitative examples.

These excerpts come from six fixed prompts evaluated on Qwen3-1.7B Base, SFT, and SpeakGR (preservation weight 1). All use greedy decoding over the original text vocabulary and a 1,024-token limit per reply. Outputs are verbatim; omissions are editorial (Appendix G.5).

Example 1: story generation, first sentence.

Prompt: Write a coherent 450 to 550 word story about a librarian repairing an old clock. Use the headings Beginning, Discovery, and Resolution. Keep the protagonist named Mira throughout.

ModelOutput excerpt
BaseMira had always been a quiet person, content with the rhythm of her life in the small, dusty library.
SFT. (omitted)
SpeakGRMira had always been fascinated by the quiet hum of the library’s old clock.

Story headings and remaining text are omitted; the full Base/SpeakGR stories contain narrative oddities.

Example 2: dialogue, first two turns.

User, turn 1: For this conversation, answer in English and finish every reply with the exact sentence "The code is amber." Remember that my favourite fruit is pear. Acknowledge in one sentence before the required ending.

User, turn 2: What is my favourite fruit? Answer with the fruit and keep following my earlier formatting rule.

ModelTurn 1Turn 2
BaseI like pears, and my favorite fruit is pear. The code is amber.My favorite fruit is pear. The code is amber.
SFT. (omitted). (omitted)
SpeakGRI love pears, and the code is amber. The code is amber.My favorite fruit is pear. The code is amber.

Each model uses its own complete dialogue history. In both examples, omitted SFT content is only punctuation/whitespace through the token limit. Base and SpeakGR retain the ending in these two turns but misattribute the user's preference as their own.

Response to Questions

Q1. Reasoning and interaction.

On 1,319 GSM8K questions, Qwen3-1.7B Base, SFT, and SpeakGR achieve 79.15%, 0.15%, and 78.39% flexible numeric-answer accuracy. This is exploratory evidence of arithmetic retention in this LoRA setting (Appendix G.1). W3 provides corresponding long-form and dialogue examples.

Q2. Unified retrieval and generation.

As discussed in W1, language preservation could support interaction alongside SID prediction. For document retrieval, one model could generate a SID, resolve it to a document, and answer using that evidence. We study the preservation aspect; establishing system-level benefits remains future work (Appendix B).

Q3. Scale and SID construction.

Please see W2 for the larger-model results. Our study focuses on language preservation with mainstream RQ-VAE-based SIDs; comparing alternative constructions is outside its present scope (Appendix B).

Thank you for your detailed rebuttal. I believe the authors have adequately addressed my previous concerns and questions.

AuthorsOpenReview ↗

Thank you for your constructive feedback

Thank you for your thoughtful follow-up. We are glad that we have addressed your concerns and questions. Your constructive suggestions helped us strengthen the paper through additional evaluations and clearer presentation. We sincerely appreciate your time and careful review.

Reviewer_Nt62OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates language forgetting during generative retrieval training and proposes SpeakGR, which uses on-policy distillation to preserve the pretrained model’s language behavior while learning semantic IDs. An adaptive variant is also introduced to dynamically adjust the preservation strength. Experiments on MS MARCO and NQ show that the proposed methods reduce language distribution drift while maintaining reasonable retrieval performance.

Critical strengths
  1. The paper studies a problem that has received relatively limited attention in GR. The observation that retrieval fine-tuning can quickly affect the pretrained language distribution provides a reasonable motivation for this work.

  2. The proposed approach is relatively simple and effective. The results show considerable reductions in language drift across different models and datasets.

  3. The experiments cover multiple language model backbones and retrieval datasets. I also appreciate that the paper reports both retrieval performance and language preservation, which helps illustrate the trade-off between these two objectives.

Critical weaknesses
  1. The motivation emphasizes the need to preserve the pretrained model’s language capabilities for downstream applications, such as question answering and interactive retrieval. However, these capabilities are not actually evaluated. The reported metrics show that the model’s token distributions remain closer to the original model, but it is unclear whether this translates into better preservation of its actual language capabilities. This leaves me somewhat unconvinced by the connection between the motivation and the experimental results. Some evaluations on downstream tasks would make the work more convincing.

  2. The main preservation objective of the method builds on existing on-policy distillation techniques. The methodological novelty seems somewhat limited. It is also unclear how much of the improvement comes specifically from using student-generated prefixes. A comparison with conventional KL distillation on fixed text prefixes would help clarify the contribution.

Questions for authors
  1. Have the authors examined whether the reduced language distribution drift also leads to better preservation of downstream language capabilities, such as question answering or instruction following?

  2. Table 7 shows that a fixed weight of 0.1 performs better than the default adaptive variant. Would Adaptive SpeakGR still provide benefits when compared with a fixed weight matched to its average adaptive coefficient?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Nt62

We sincerely thank the reviewer for recognizing the motivation of studying language forgetting in generative retrieval, the effectiveness of our simple approach, and the value of evaluating retrieval and preservation together. Your insightful suggestions prompted additional downstream language evaluations, a fixed-prefix control, and a matched-average-weight comparison. These additions strengthen the empirical evidence and clarify the contribution of language preservation and the role of adaptive weighting.

Response to Weaknesses

W1. Language preservation and language tasks.

Our focus is preserving pretrained language behavior during SID specialization. WikiText forward KL, perplexity, and top-1 agreement directly assess this objective. Following your insightful suggestion, we added exploratory tests of instruction following, arithmetic, and QA. In our full-parameter experiments, retention on these tasks was more limited. For these additional exploratory experiments, we used LoRA for efficient training. Both Qwen3-1.7B/NQ adapted models below use matched training settings and the same checkpoint across metrics within each row.

ModelR@1 (%)Forward KLIFEval (%)GSM8K (%)QA EM (%)
Base—0.00065.4379.1524.02
SFT32.816.11312.750.150.06
SpeakGR27.270.11558.9678.3919.81

SpeakGR retains more downstream performance than SFT: GSM8K remains close to Base, while IFEval and QA show partial retention. R@1 is 5.54 percentage points lower. QA uses identical passages from a full-parameter SFT retriever. We report IFEval strict prompt accuracy (541 prompts), GSM8K flexible numeric-answer accuracy (1,319 questions), and QA exact match (3,610 questions). Both adapted models use terminal checkpoints from one matched training seed.

These task evaluations complement the WikiText evidence for language preservation. We thank you for this extension, added in Appendix G.1; its exploratory scope is clarified in the limitations in Appendix B.

W2. Contribution and fixed-prefix control.

Our contribution is a systematic study of language preservation during SID specialization, across six backbone–dataset settings. The preservation objective builds on established on-policy distillation.

The requested control uses full-distribution KL on fixed text prefixes. It matches SpeakGR's initialization, seed, retrieval data/order, prompt pool, preservation weight 1, training schedule, and checkpoint selection. Both use the frozen base model's next-token distributions; only the distillation prefixes differ (Appendix G.3). Following Appendix A.1 of the original paper, each preservation continuation uses up to 32 tokens during training.

Prefix sourceR@1 (%)GSM8K (%)QA EM (%)
Fixed text25.8972.2518.81
SpeakGR student27.2778.3919.81

Fixed-prefix KL is effective; student prefixes give higher scores in this comparison. An additional control with nearly matched numbers of supervised tokens shows the same pattern (Appendix G.3). These results use one training seed.

Response to Questions

Q1. Capability preservation.

Please see W1 and Appendix G.1 for the task evaluations and their scope.

Q2. Matched-average fixed weight.

Thank you for this useful suggestion. We trained a new full-parameter adaptive reference and matched fixed-weight control on Qwen3-0.6B/MS MARCO, separate from the main-table run. The fixed weight, 0.3496, is the average coefficient actually applied over the 10,200 updates leading to the adaptive checkpoint. Data, prompt schedule, distillation-token counts, and checkpoint selection are matched; both select stage-3 step 200.

Weight policyR@1 (%)R@10 (%)Forward KL
Matched fixed24.0164.981.495
Adaptive25.1263.861.701

The fixed control performs well. Adaptive has a higher R@1 point estimate, while fixed weighting has higher R@10 and lower KL. The R@1 difference's paired 95% interval is -1.73 to +3.84 percentage points and includes zero. Adaptive offers feedback-based adjustment as drift changes; its stage-average weights were 0.013, 0.442, and 0.010. We present it as an optional mechanism. This suggested control sharpened the paper's insight into preservation weighting (Appendix G.4).

Thank you for your detailed response and additional experiments.

The new downstream evaluations provide useful evidence of language capability preservation. However, these experiments use LoRA, while the main experiments use full-parameter fine-tuning. These two settings are not directly comparable, so I still have some reservations.

The fixed-prefix comparison addresses my concern about the effectiveness of on-policy distillation.

For Adaptive SpeakGR, the matched-weight comparison does not show a clear advantage over fixed weighting. This makes the claimed benefits of the adaptive mechanism less convincing. It would be more accurate to tone down these claims.

AuthorsOpenReview ↗

Follow-up Response to Reviewer Nt62

Thank you for the helpful follow-up. We are glad that the fixed-prefix comparison addresses your concern about on-policy distillation, and appreciate your suggestions on the remaining points.

The original evaluations directly assess language preservation under full-parameter tuning. On held-out WikiText, forward KL measures distributional drift from the pretrained model, perplexity measures language-modeling quality, and top-1 agreement measures consistency with its next-token predictions. Their consistent improvements across six backbone–dataset settings provide complementary evidence for preserving pretrained language behavior during SID specialization.

Both SFT and SpeakGR use LoRA in the added comparison. With matched training settings, adding the preservation objective improves IFEval from 12.75% to 58.96% and GSM8K from 0.15% to 78.39%. This supports the objective's contribution within that configuration. We agree that these downstream results should remain specific to this setting. Downstream task retention under full-parameter tuning is more limited, as clarified in Appendix B.

We have toned down the claims for Adaptive SpeakGR. The matched-average comparison does not establish an advantage over fixed weighting. We now present adaptation as an optional feedback mechanism and specify that the original retrieval gains are relative to the default fixed-weight configuration. We discuss the matched-average comparison in Section 5.3 and Appendix G.4. The central contribution remains language preservation during SID specialization, which is also supported by fixed-weight SpeakGR.

Thank you again for helping us strengthen the paper. We hope these clarifications and revisions address your remaining concerns.

Reviewer_jaU4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies catastrophic forgetting in LMs that are finetuned for generative retrieval (GR) with semantic IDs. The authors use a three-stage training pipeline: document indexing, pseudo-query training, and supervised query finetuning. They observe that retrieval-only finetuning changes the model's base distribution over text. To mitigate this, the authors propose SpeakGR, which adds an auxiliary loss in the form of the forward KL divergence between the frozen base model’s next-token distribution and its own, computed at each sampled prefix over the original text vocabulary. The authors also propose an adaptive variant that adjusts the KL weight λ in an online fashion. The experiments cover three pretrained models (Qwen3-0.6B/1.7B, and Gemma-3-1B-IT) on MSMARCO and Natural Questions, and show that the new aux loss can help mitigate catastrophic forgetting while preserving retrieval performance.

Critical strengths

The forgetting of language generation ability does occur while finetuning to generate semantic IDs, and the authors conduct many experiments across different hyperparameter settings to illustrate this in well-known benchmarks while demonstrating the improvements due to their method.

Critical weaknesses

There is relevant related work that hasn’t been cited. The problem of performing generation and retrieval jointly has been studied in a number of papers, e.g., CorpusLM (Li et al, 2024) and "Direct Retrieval-augmented Optimization" (Shi et al, 2025). CorpusLM is conceptually quite close: it trains one model on a weighted sum of a generative-retrieval loss and natural-language generation losses, so retrieval specialization is regularized by continued language supervision.

I am not entirely convinced by the motivation of this paper. Catastrophic forgetting is not necessarily a problem in generative retrieval, since the retrieval model and the generation model are usually separate. In most RAG applications, the retrieval is implemented as tool use, and the retriever does not need to be the same model that handles the generation task. The authors are interested in preserving the probability distribution in the original LM, but it isn't clear why this is important. For example, it isn't clear that merging the retriever with the generation model would be beneficial versus keeping them separate, since this paper doesn't contemplate that experimental line. The authors concede in Appendix B that the retrieve-then-generate setting is motivating but not actually demonstrated in this paper. Ultimately, if we intend to use the GR model for GR and nothing else, then forgetting isn't a problem. In addition, the broader question of mitigating catastrophic forgetting in multi-task learning has been addressed extensively by prior work, and GR is just one more example.

Adding the auxiliary loss does help preserve the model's original text distribution, but it can also degrade retrieval performance. In Table 1, we observe mixed results in MSMARCO and a significant degradation in NQ for Qwen (for the largest Qwen3-1.7B model, SFT-only Recall@1 is 48.08% versus SpeakGR 40.56% and AdaSpeakGR 43.87%). It's not clear that we would want to preserve the original model's distribution if the retrieval performance degradation is serious.

Questions for authors

The models examined in this work were quite small, starting with 600M models and going up to 1.7B in the largest setting. Small models may be especially prone to catastrophic forgetting, see "Investigating Continual Pretraining in Large Language Models" (Yildiz et al, 2024). CorpusLM, which was written in 2024, used a 7B model with QLoRA. Is the forgetting simply an artifact of the small model sizes used? Or perhaps it is an artifact of tuning the entire model instead of using LoRA? See "LoRA Learns Less and Forgets Less" by Biderman et al. (2024).

4 comments
AuthorsOpenReview ↗

Response to Reviewer jaU4

We sincerely thank the reviewer for the thoughtful comments and for recognizing our experiments on language forgetting during SID training and the improvements from SpeakGR. Your insightful questions helped us clarify how our motivation and target problem of language preservation differ from those of related studies, and motivated additional experiments on model size and adaptation method. We address each concern below.

Response to Weaknesses

W1. Related work: motivation and target problem.

We have added the suggested references in Section 2 and clarified the distinct research questions.

CorpusLM integrates retrieval and generation to improve corpus-grounded task performance; our motivation is to address language forgetting caused by SID specialization, with language preservation relative to the pretrained backbone as the target.

DRO addresses end-to-end RAG optimization by jointly training a document selector and a separate generator; SpeakGR addresses language preservation within the SID-generating backbone.

Our question is how to learn SID retrieval while retaining the backbone's pretrained speaking behavior, using the frozen base model as the preservation reference (Sections 1 and 3–4).

W2. Motivation for language preservation.

We study retaining pretrained natural-language predictive behavior while learning SID retrieval. WikiText evaluates this directly through forward KL, perplexity, and top-1 agreement. Section 1 describes a common interface for SIDs and text and the possibility of reusing one backbone. Language preservation is an initial step toward studying retrieval and language interaction within one model.

Separate retriever–generator systems are effective solutions, but leave open whether one backbone can acquire retrieval while retaining its speaking behavior. Reuse could reduce the need to maintain separate model weights in memory. Appendix B states this scope, and Appendix G extends the evaluation to language tasks.

W3. Retrieval–preservation trade-off.

The Qwen/NQ recall reduction is a meaningful trade-off, already reported in Table 1 and discussed in Section 5.2.1. Section 5.3 and Appendix F examine how preservation strength changes this balance.

A dedicated retriever can prioritize retrieval accuracy; a backbone intended for retrieval and language tasks must also be assessed for what it retains. Our contribution makes language preservation measurable and studies how to control this trade-off. Appendix G.1 adds exploratory task-level evidence, with its scope stated in Appendix B.

Response to Questions

Q1(a). Model size.

Yıldız et al. examine model-size effects in continual domain-adaptive pretraining; we study specialization to SID prediction.

Following your suggestion, we added Qwen3-8B/NQ experiments. To train this larger backbone efficiently, we used LoRA for both SFT and SpeakGR with matched settings. WikiText forward KL is 3.816 for SFT and 0.093 for SpeakGR, with R@10 of 77.42% and 71.37%, respectively (Appendix G.2). Forgetting and the preservation effect therefore also occur beyond the original small models.

Q1(b). Adaptation method.

Biderman et al. compare LoRA and full fine-tuning in programming and mathematics; our question is whether LoRA alone preserves language behavior during SID training.

Our additional experiments use LoRA for training efficiency. The matched comparisons in Appendices G.1–G.2 show that retrieval-only LoRA still reduces language performance, while adding SpeakGR improves retention under the same adaptation setting. This supports a role for the preservation objective beyond LoRA alone, within the evaluated configurations. We have cited both suggested studies in Section 2 and appreciate the questions that helped clarify this scope.

I thank the authors for responding to my review and for adding Appendix G. They have addressed the questions I listed above. My concern remains the scientific motivation, where the authors and I unfortunately disagree. The new results in Appendix G (Table 8) show that using SpeakGR does degrade retrieval performance significantly. The loss in retrieval performance is quite large (~9 points in recall@10) for the sake of keeping the model close to the base LM, and SpeakGR trails the base model anyway on language tasks. It isn't clear why this tradeoff needs to be accepted at all, when a system that separates retrieval and generation would perform better.

Thank you for acknowledging that our responses and Appendix G have addressed your questions. We agree that separate retrieval and generation systems are a strong practical choice. As emphasized in our Introduction, our motivation is to explore highly integrated, unified models in which retrieval becomes an additional capability alongside language generation and reasoning. A shared autoregressive interface could coordinate language and SID generation, enable potential reasoning over the SID space, and reuse one backbone across these functions.

Realizing this vision is challenging and requires substantial research. Preserving language functionality during SID learning is a basic prerequisite, and our paper takes an initial step by studying this obstacle and exploring how to mitigate it. We are encouraged that the other two reviewers also recognized the value of investigating language preservation in this setting. We respect your different perspective on the broader motivation and have sought to explain its rationale and scope as clearly as possible. We hope this clarifies why we consider unified models a worthwhile research direction alongside effective modular systems.

One further point worth considering is the system resource requirements. Maintaining two separate models introduces additional computational and memory overhead, making the overall framework less efficient. This is particularly problematic in resource-constrained scenarios, such as edge-based retrieval systems, where limited memory and computational resources make deploying multiple models costly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions and Rebuttal Progress

Dear Area Chair,

Thank you for overseeing the review process. We would like to highlight the central contribution of SpeakGR and the substantial progress made in addressing the reviews.

  • SpeakGR brings an important preservation question into generative retrieval. Learning semantic IDs changes the pretrained language model itself, making the retention of its language behavior a consequential research question. Our work systematically identifies, measures, and mitigates this forgetting. It broadens the evaluation of generative retrievers to include what the backbone retains while acquiring retrieval, with implications for models intended to support both retrieval and language interaction. Reviewers zS2j and Nt62 both recognized the motivation for studying this underexplored problem.

  • The contribution is supported by a simple method and consistent empirical evidence. Our original full-parameter experiments span three backbones and two retrieval datasets, demonstrating substantial preservation benefits through forward KL, perplexity, and top-1 agreement alongside retrieval performance. Reviewer zS2j highlighted the method's simplicity, motivation, and experimental support. Reviewer Nt62 also recognized its effectiveness and appreciated the evaluation of retrieval and preservation together. Fixed-weight SpeakGR directly establishes the empirical basis of this contribution.

  • New experiments directly address the principal reviewer requests. We added downstream language evaluations, an 8B backbone experiment, qualitative examples, a fixed-prefix distillation control, and a matched-average-weight comparison. Reviewer zS2j explicitly confirmed that their previous concerns and questions were adequately addressed. Reviewer Nt62 acknowledged the value of the downstream evaluations and confirmed that the fixed-prefix comparison addressed their concern about on-policy distillation.

  • A distinct motivation and research focus. In response to Reviewer jaU4's question about related work, we explained the differences in motivation and target problem. CorpusLM integrates retrieval and generation to improve corpus-grounded task performance. DRO jointly optimizes a document selector and a separate generator for RAG. SpeakGR is motivated by language forgetting caused by SID specialization and targets retention of the backbone's pretrained speaking behavior, using the frozen base model as the preservation reference. We added the suggested references and clarified this distinction in Section 2. We also clarified the scope of the LoRA evaluations and moderated the adaptive claims.

We have carefully addressed every reviewer comment through additional experiments, explicit explanations, or concrete manuscript revisions. We hope the above contributions, evidence, and revisions will be helpful in your final decision. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.