acceptodds
Under review as a conference paper at ICLR 2027

HyperBridge: Effectively Adapting Hypernetworks across LLM Architectures

Abstract

Hypernetworks have recently been explored to internalize long contexts into Large Language Model (LLM) parameters, where a hypernetwork maps documents to parameter updates that enable the adapted LLM to answer relevant queries without repeatedly processing the original context. However, obtaining such a hypernetwork requires substantial data and training compute. Beyond that, the resulting hypernetwork is tightly coupled to its corresponding LLM: its internal modules must match the backbone’s hidden dimensions and number of layers. This limits practical usage under constrained resources since repeating the same training process for each newly introduced LLM is prohibitively expensive. Thus, we ask: Under limited data and compute, can we transfer a trained source hypernetwork across backbones and enable more effective target hypernetwork training? To answer this, we conduct a systematic cross-architecture analysis using representational similarity analysis. We reveal a clear separation in transferability across the hypernetwork: its intermediate representations exhibit substantial shared structure across backbones and can be aligned via a linear mapping. In contrast, its parameter generator remains strongly architecture-specific. Guided by these findings, we propose HyperBridge, a cross-architecture hypernetwork transfer framework that aligns input representations, preserves transferable weights, and re-learns backbone-specific heads. We evaluate HyperBridge across multiple LLM architectures and QA datasets, showing up to 28.5% higher performance than training from scratch and remaining effective across data budgets.

Primary area:
generative models
Keywords:
hypernetworks, context internalization, LLMs

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

The paper studies “cross-architecture transfer of context-internalization hypernetworks”. A context-internalization hypernetwork is a neural network that maps a document into LoRA parameter for a frozen LLM without normal LoRA training. A trained hypernetwork is tightly coupled to its backbone (input projection must match hidden dims; output generator must match layer count/dims), so adapting to a new LLM would normally require retraining from scratch, which is expensive in data and compute. The authors ask whether a trained source hypernetwork can be reused to make target hypernetwork training more effective under limited resources.

They first conduct a systematic cross-architecture analysis on a 600-document probe set across Gemma2-2B, Qwen3-4B and Mistral-7B, yielding three observations: (i) input projections across backbones are misaligned but can be recovered by a linear map; (ii) Perceiver encoder/decoder representations are substantially shared/transferable; (iii) the LoRA generation head is backbone-specific (near-full-rank spectrum).

Based on these findings they propose HYPERBRIDGE: ALIGN (fit a linear map and fold it into the target input projection), PRESERVE (initialize target Perceiver from source, plus low-rank LoRA adapters), and RE-LEARN (layer-interpolate the post-Perceiver, re-initialize the generation head), trained with the standard KL-distillation + L1 objective. Validation covers all six source→target directions at 5K/10K/15K/20K training steps on SQuAD/DROP/ROPES (with FineWeb-Edu pretraining), reporting up to 28.5% relative gains over from-scratch baselines, with ablations on input alignment, Perceiver adaptation, and data composition.

Critical strengths
  1. Clear, well-motivated problem and a genuinely informative analysis. Introducing cross-architecture hypernetwork transfer as a problem is timely and practical, and the RSA/CKA-driven decomposition into "align / preserve / re-learn" components turns the analysis directly into an actionable recipe. This makes the paper read as an honest empirical study of where hypernetwork knowledge is reusable, which is consequential because it can guide future backbones without repeating the full pipeline.

  2. Consistent, budget-aware empirical validation. The evaluation covers all six transfer directions, multiple datasets, and four training budgets; HyperBridge improves over from-scratch Native on most directions and budgets (14.7% at 5K, up to 28.5% on Mistral), with ablations isolating each design choice (input alignment, Perceiver adaptation, data composition). The robustness across data budgets directly supports the "limited resources" motivation.

Critical weaknesses
  1. The analysis conclusions are tied to the specific Doc-to-LoRA architecture and their generality across the broader space of context-internalization approaches is not established. The paper fixes the hypernetwork decomposition into the "input projection P + Perceiver E + LoRA generator G" trio, and all three core observations (projection linearly alignable, Perceiver preservable, generator backbone-specific) are derived on this one specific architecture. However, context internalization admits other implementations: peer context→LoRA hypernetworks (e.g., SHINE) may use different compressors; task-conditioned hypernetworks (HyperLoRA/ZHyper, etc.) do not use a Perceiver at all; Shin et al. 2025, Qi et al. 2025, and Kujanpää et al. 2025 pursue per-context optimization/self-distillation rather than hypernetworks; and Caccia et al. 2025 follow a fundamentally different plug-n-play "knowledge module + deep context distillation" paradigm. The paper neither studies the sensitivity/robustness of its findings to architectural choices nor discusses how the framework would transfer to these alternatives. Consequently, the claimed generality of "which hypernetwork components are reusable across backbones" is only supported for Doc-to-LoRA-style LoRA-generating hypernetworks, which weakens its value to the broader community and, in turn, the significance of the contribution.

  2. Limited generality and modest, noisy margins. Only small backbones (2B–7B), 512-token documents, extractive-QA ROUGE metrics, and no significance/variance reporting temper the headline result; several gains (especially DROP) are small and some Native@20K cells rival the transferred model. The "up to 28.5%" figure, while correct, sits atop a heterogeneous set of much smaller improvements, and the mismatch of evaluation metrics across targets makes cross-backbone comparisons fragile. The causal generality of the analysis (e.g., does the linear-alignability observation hold for larger or non-decoder-only backbones?) is untested.

Questions for authors

(Clarifying) Are the reported numbers from a single run or multiple seeds? If multiple, please report mean/variance; if not, are the observed gaps (often a few points on DROP) statistically meaningful?

8 comments

We thank the reviewer for the thoughtful feedback. We address these concerns as follows.

(W1) Specificity to Doc-to-LoRA

We thank the reviewer for the comment, but would like to clarify several important distinctions.

  • We note that the review references several works without providing proper citations. We encourage the reviewer to provide precise references, particularly when using prior literature to substantiate a criticism of our paper.

  • The review conflates distinct CI paradigms. Several CI frameworks the author mentioned do not involve transferring a pre-trained context-to-parameter hypernetwork. Consequently, they do not face the cross-backbone generator-transfer problem addressed by HyperBridge.

  • Our philosophy extends beyond Doc-to-LoRA. For SHINE [1], we can naturally align backbone-specific memory states, preserve transferable M2P Transformer components, and re-learn target-specific modules. While the exact operations differ, the underlying principle of representation-guided selective transfer remains applicable. More details please refer to our response to Reviewer 3HSp (Q1).

(W2) Limited generality and noisy gains

We clarify each of the points as follows:

  • Backbone scale and evaluation scope: We use the publicly available Doc-to-LoRA checkpoints and follow its 512-token training and evaluation setup for consistency. This backbone scale is also comparable to SHINE [1], which evaluates 0.6B–8B Qwen3 models.
  • Transfer gains: The following table reports relative gains (%) to provide a more comprehensive picture of HyperBridge's gain over Native training. We additionally report mean performance with variance over three random seeds in our response to (Q1) below.
Transfer Direction5K10K15K20K
Qwen3-4B → Gemma2-2B+20.8%+17.2%+9.7%+7.8%
Mistral-7B → Gemma2-2B+22.7%+18.8%+10.4%+8.4%
Gemma2-2B → Qwen3-4B+14.3%+10.4%+8.6%+7.2%
Mistral-7B → Qwen3-4B+7.8%+10.8%+11.4%+8.8%
Gemma2-2B → Mistral-7B+27.0%+33.0%+26.9%+23.7%
Qwen3-4B → Mistral-7B+21.6%+28.4%+20.9%+18.8%
Mean over 6 directions+19.1%+19.8%+14.6%+12.4%
  • Evaluation metrics: In paper, we follow Doc-to-LoRA's evaluation protocol (Section 5.1). We additionally provide both ROUGE-L F1 and recall for all target backbones as follows. The gain of HyperBridge holds.
TargetMethodROUGE-L F1ROUGE-L recall
Gemma2-2BNative0.390 ± 0.0320.402 ± 0.034
Gemma2-2BQwen3-4B →0.421 ± 0.0100.431 ± 0.010
Gemma2-2BMistral-7B →0.423 ± 0.0100.435 ± 0.012
Qwen3-4BNative0.120 ± 0.0430.420 ± 0.011
Qwen3-4BGemma2-2B →0.197 ± 0.0110.451 ± 0.006
Qwen3-4BMistral-7B →0.181 ± 0.0210.458 ± 0.005
Mistral-7BNative0.207 ± 0.0130.437 ± 0.011
Mistral-7BGemma2-2B →0.224 ± 0.0020.541 ± 0.006
Mistral-7BQwen3-4B →0.213 ± 0.0180.519 ± 0.016
  • Generality: We have studied six transfer directions across three distinct LLM families with descent model sizes. Extending our analysis to non-decoder-only architectures is beyond the scope of our work.

(Q1) Clarification question

The reported number in the paper is from a single run. To further address the reviewer's concern, we replicate our main experiment results in Table 1 with two additional random seeds and record mean ± std over three seeds. Mean metrics over three datasets are reported. As shown in the following tables, HyperBridge consistently outperforms Native across all target backbones and training budgets, demonstrating robust transfer gains.

TargetMethod5K10K15K20K
Gemma2-2BNative0.314 ± 0.0230.354 ± 0.0250.380 ± 0.0340.390 ± 0.032
Gemma2-2BQwen3-4B →0.380 ± 0.0110.415 ± 0.0170.416 ± 0.0020.421 ± 0.010
Gemma2-2BMistral-7B →0.386 ± 0.0030.421 ± 0.0060.419 ± 0.0100.423 ± 0.010
TargetMethod5K10K15K20K
Qwen3-4BNative0.358 ± 0.0060.395 ± 0.0080.410 ± 0.0120.420 ± 0.011
Qwen3-4BGemma2-2B →0.409 ± 0.0060.436 ± 0.0070.445 ± 0.0070.451 ± 0.006
Qwen3-4BMistral-7B →0.386 ± 0.0040.438 ± 0.0040.457 ± 0.0090.458 ± 0.005
TargetMethod5K10K15K20K
Mistral-7BNative0.366 ± 0.0070.388 ± 0.0080.424 ± 0.0130.437 ± 0.011
Mistral-7BGemma2-2B →0.465 ± 0.0120.517 ± 0.0060.537 ± 0.0050.541 ± 0.006
Mistral-7BQwen3-4B →0.445 ± 0.0040.499 ± 0.0270.512 ± 0.0240.519 ± 0.016

[1] SHINE: A Scalable In-Context Hypernetwork for Mapping Context to LoRA in a Single Pass, ICML 2026.

Reviewer_tDYAOpenReview ↗

Re: We note that the review references several works without providing proper citations.

All the works I mentioned can be found in the References in your paper:

(Shin et al. 2025, Qi et al. 2025, and Kujanpää et al. 2025):

  • Haebin Shin, Lei Ji, Yeyun Gong, Sungdong Kim, Eunbi Choi, and Minjoon Seo. Generative prompt internalization. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 7338–7363, 2025.
  • Siyuan Qi, Bangcheng Yang, Kailin Jiang, Xiaobo Wang, Jiaqi Li, Yifan Zhong, Yaodong Yang, and Zilong Zheng. In-context editing: Learning knowledge from self-induced distributions. In International Conference on Learning Representations, volume 2025, pp. 77563–77585, 2025.
  • Kalle Kujanp¨a¨a, Pekka Marttinen, Harri Valpola, and Alexander Ilin. Effi cient knowledge injection in LLMs via self-distillation. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=drYpdSnRJk.

(Caccia 2025):

  • Lucas Caccia, Alan Ansell, Edoardo Ponti, Ivan Vuli´c, and Alessandro Sordoni. Training plug-n-play knowledge modules with deep context distillation. arXiv preprint arXiv:2503.08727, 2025.

We thank the reviewer for noticing that these works come from our paper's Introduction, where readers can clearly see that they are mentioned only as background knowledge on context internalization for LLMs.

We would like to emphasize that using these background citations to criticize a hypernetwork-based method for not applying to non-hypernetwork approaches might reflect a clear misunderstanding of the scope of our work.

Reviewer_3HSpOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies transferring Doc-to-LoRA hypernetworks [1], which map documents to LoRA adapters for context internalization (CI), between LLM backbones to reduce the cost of training a new hypernetwork for each model. The authors break the Doc-to-LoRA architecture into several modules and analyze which should be aligned, adapted, or retrained when moving to another LLM. They use this analysis to develop HyperBridge, which uses a linear mapping to align input representations, adapts the perciver via LoRA, and trains a new output head to generate weights for the target. They run experiments to test whether transferring a trained hypernetwork helps when adapting to a new backbone and which parts of the transfer strategy account for the gains. The results show that reusing learned components can improve question-answering performance with less target training, although the benefit depends on the source and target models.

Critical strengths
  • (S1) Useful analysis of what transfers when moving to a new backbone. Breaking the hypernetwork architecture into an input projection, preciver, and a LoRA generator gives a clear way to study transfer between backbones. The representation analysis in Section 3 and the ablations in Table 3, provides evidence that some learned knowledge can be reused, and offers guidance on how to reuse existing Doc-to-LoRA hypernetworks.

  • (S2) Practical value of transfer. The empirical analysis makes a convincing case that transfer can improve learning when target data is limited. The benefits across different model pairs and data budgets suggest that existing hypernetwork checkpoints provide a useful starting point for new backbones.

Critical weaknesses
  • (W1) Impact while the problem of hypernetwork adaptability across model architectures is real, I don't believe that it is among the most important issues needed to tackle in order to make context internalization hypernetworks useful in practice. Specifically, generalization to unseen documents, composiblity, qulity of steering etc. are more pressing issues. I don't view this as a detremental weakness, as this is still and interesting and important problem to solve on the path of making hypernetworks usefull at scale, but it is a weakness nevertheless given the current state of the field and reduces the impact of the paper.
Questions for authors
  • (Q1) Applicability to other CI hypernetwork designs. What motivated choosing Doc-to-LoRA [1] specifically, and which findings would you expect to extend to other hypernetwork-based CI architectures? For example, [2] uses a Transformer to mix memory states across LLM layers and reshapes its outputs into LoRA weights. [3] constructs updates from outer products of hidden states via projections and SVD normalization. Neither has the same Perceiver and output-head you analyze in the paper, so the conclusions about how to adapt each modules do not carry over directly. How would you expect the the analysis and transfer strategy to change in these settings?

  • (Q2) Transfer between checkpoints of the same model. While transfer across architectures is a harder problem, I expect that in practice, shipping a CI hypernetwork alongside a model would make adaptation across checkpoints of that model more immediately useful. For example, after additional training or at different stages of post-training. Could your method enable more efficient adaptation in this setting, and which parts of the transfer procedure could be simplified?


[1] Charakorn et al. Doc-to-LoRA: Learning to Instantly Internalize Contexts. arXiv:2602.15902, 2026.

[2] Liu et al. SHINE: A Scalable In-Context Hypernetwork for Mapping Context to LoRA in a Single Pass. ICML, 2026.

[3] Chen et al. Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass. ICLR, 2025.

3 comments

We thank the reviewer for recognizing our work. We further address the comments as follows.

(W1) Impact of the work

We thank the reviewer for this thoughtful perspective. We would like to further clarify the impact of our work from two perspectives:

  • Broader implications for CI: Beyond cross-architecture transfer, our work reveals a useful principle: identifying reusable knowledge through representation-level analysis, and selectively adapting architecture-dependent components. We believe this philosophy could also inform the challenges highlighted by the reviewer. Some examples include:

    • Generalization: Backbone-invariant representations should provide a useful inductive bias for learning more robust context representations that better generalize to unseen documents.

    • Composability: A shared representation space could enable combining information from multiple contexts before generating parameter updates, avoiding direct composition of potentially incompatible LoRA weights at the output level.

  • Necessity of cross-architecture transfer: Beyond issues of normal single-faimly hypernetworks, we envision two scenarios where HyperBridge is particularly valuable:

    • One-Large-to-many-small transfer: A CI hypernetwork trained on a large, powerful backbone can be adapted to many smaller LLMs through HyperBridge for resource-constrained deployment.
    • Heterogeneous multi-agent systems (MAS): Specialized agents powered by different LLM architectures collaborate while requiring shared context-internalization capabilities. HyperBridge allows these agents to reuse learned CI knowledge without training a separate hypernetwork from scratch for every backbone.

(Q1) Applicability to other hypernetwork designs

Thank you for the insightful question. Firstly, we chose Doc-to-LoRA mainly because it provides a clean repo with public pre-trained (and well-trained) checkpoints across several LLM backbones, which makes a controlled cross-backbone analysis possible for us.

For adapting HyperBridge to other CI designs, we emphasize that our core philosophy is not tied to the specific architecture used by Doc-to-LoRA: we aim to identify transferable representations and backbone-specific components, then selectively preserve or adapt them. For example, this philosophy could extend to SHINE [2]: we could (i) ALIGN the target memory states to the source M2P hidden space through a learned projection, (ii) PRESERVE the source M2P Transformer with lightweight LoRA update, and (iii) RE-LEARN the backbone-specific Meta-LoRA & memory embeddings, with a target-specific output projection to recover the required LoRA dimensions.

For Generative Adapter [3], which follows another paradigm that does not employ a separately trained hypernetwork, our module-level transfer strategy and analysis might need more adjustment. We believe our representation-guided transfer philosophy should provide a potential starting point.

(Q2) Transfer within the same model

We agree that adapting hypernetworks across different versions of the same backbone LLM might be a pretty interesting use case. While HyperBridge targets the cross-architecture setting, its core principles should translate to this setting as a special case.

For example, guided by HyperBridge’s philosophy of modular preservation, we can preserve the Perceiver modules (which should capture higher-level, invariant context representations) and focus on lightweight tuning on the output heads to adjust the backbone's shift. We can also consider a similar activation-based correction to fix the context misalignment across checkpoints given a few calibration documents.

Overall, we believe this should be an interesting direction to pursue building on top of this work.

Reviewer_3HSpOpenReview ↗

Reply to authors' response

I thank the authors for the clarifications. I have no further questions.

Thank you for the feedback. We will be happy to answer any further question if there is any.

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

This paper investigates transferring context-internalization hypernetworks between different large language model architectures. They generate document-conditioned LoRA parameters, allowing an LLM to answer queries without repeatedly including the original document in its input. The method is evaluated across six transfer directions involving Gemma2-2B, Qwen3-4B, and Mistral-7B, with SQuAD, DROP, and ROPES as evaluation datasets.

Critical strengths
  1. HyperBridge provides a concrete way to reuse previously trained components instead of restarting the entire training procedure.
  2. The component-level experimental design helps identify where transferred initialization is useful.
Critical weaknesses
  • The evidence for the claimed stage-wise transferability decomposition is incomplete
  • Report the efficiency of the proposed method
Questions for authors
  1. What directly demonstrates that the generator is architecture-specific?
  2. Which part of HyperBridge is necessary for successful transfer?
  3. What are the total costs for HyperBridge?
  4. Are the 600 alignment-probe documents disjoint from final evaluation examples?
  5. How stable are the observed transfer gains?
2 comments

We address the reviewer's concerns as follows.

(W1) Stage-wise transferability

We emphasize that stage-wise transferability claims are well supported by representation-level analysis (Section 3) and downstream ablation studies (Section 5.3):

  • Input Projection: Figure 2(a) shows that cross-backbone projection misalignment can be recovered through linear mapping. Consistently, the alignment ablations (Tables 2) show that alignment improves downstream performances.

  • Perceiver: Figure 2(b) shows consistently high cross-backbone RSA/CKA across Perceiver layers. Table 3 further validates the benefit of transferring Perceiver: frozen Perceiver mostly outperform full-parameter tuning or frozen re-initialized Perceiver.

  • Generation Head: LoRA generator is inherently architecture-specific: the output shape of the LoRA generator needs to respect the backbone model's layer count and hidden dimensions. Additionally, we conduct more ablation studies in the response to (Q1) below.

(W2)(Q3) Efficiency

We report the runtime of Native and HyperBridge. The experiment setup follows Table 1.

TargetNative 20K stepsHyperBridge 20K stepsSpeedup
Gemma2-2B7.4 h5.9 h1.2–1.3×
Qwen3-4B9.2 h5.7 h1.5–1.6×
Mistral-7B7.8 h5.7 h1.3–1.4×

The one-time alignment overhead of HyperBridge ranges from 1.7 to 8.7 minutes, which is negligible compared to training cost.

(Q1) Why generator is architecture-specific

Following our response to (W1), we conduct additional ablations on the two generator components: the post-Perceiver (PP) and linear generation head. We use Qwen3-4B as the target and report mean ± std over three seeds, averaged across 3 datasets. The conclusion is as follows.

  • Post-Perceiver: Random initialization performs comparably to source initialization adopted by HyperBridge, suggesting limited benefit from transferring its weights.

  • Linear head: Copying source weights offers no advantage, while freezing them substantially degrades performance, demonstrating the need for architecture-specific learning.

SourcePP Init.5K10K15K20K
—Native0.358 ± 0.0060.395 ± 0.0080.410 ± 0.0120.420 ± 0.011
Gemma2-2BHyperBridge0.409 ± 0.0060.436 ± 0.0070.445 ± 0.0070.451 ± 0.006
Gemma2-2Bw/ random PP0.399 ± 0.0080.447 ± 0.0110.462 ± 0.0140.464 ± 0.015
Mistral-7BHyperBridge0.386 ± 0.0040.438 ± 0.0040.457 ± 0.0090.458 ± 0.005
Mistral-7Bw/ random PP0.395 ± 0.0080.436 ± 0.0170.453 ± 0.0140.454 ± 0.014
Source ModelHead Init.5K10K15K20K
—Native0.358 ± 0.0060.395 ± 0.0080.410 ± 0.0120.420 ± 0.011
Gemma2-2BHyperBridge0.409 ± 0.0060.436 ± 0.0070.445 ± 0.0070.451 ± 0.006
Gemma2-2Bcopied head, learnable0.395 ± 0.0060.439 ± 0.0130.457 ± 0.0120.459 ± 0.014
Gemma2-2Bcopied head, frozen0.301 ± 0.0030.301 ± 0.0010.296 ± 0.0070.298 ± 0.006
Mistral-7BHyperBridge0.386 ± 0.0040.438 ± 0.0040.457 ± 0.0090.458 ± 0.005
Mistral-7Bcopied head, learnable0.378 ± 0.0090.441 ± 0.0130.456 ± 0.0060.459 ± 0.005
Mistral-7Bcopied head, frozen0.304 ± 0.0030.304 ± 0.0030.299 ± 0.0060.296 ± 0.005

(Q2) HyperBridge details

Following our response to (Q2) and Tables 3 in paper, we show that preserving the Perceiver is the most critical part.

(Q4) Probing details

Yes, probe documents are in a separate validation split.

(Q5) Stability of the HyperBridge gains

We replicate our main experiment with two more random seeds. The average over three datasets is reported. We show that HyperBridge is still well above Native in all cells.

TargetMethod5K10K15K20K
Gemma2-2BNative0.314 ± 0.0230.354 ± 0.0250.380 ± 0.0340.390 ± 0.032
Gemma2-2BQwen3-4B →0.380 ± 0.0110.415 ± 0.0170.416 ± 0.0020.421 ± 0.010
Gemma2-2BMistral-7B →0.386 ± 0.0030.421 ± 0.0060.419 ± 0.0100.423 ± 0.010
TargetMethod5K10K15K20K
Qwen3-4BNative0.358 ± 0.0060.395 ± 0.0080.410 ± 0.0120.420 ± 0.011
Qwen3-4BGemma2-2B →0.409 ± 0.0060.436 ± 0.0070.445 ± 0.0070.451 ± 0.006
Qwen3-4BMistral-7B →0.386 ± 0.0040.438 ± 0.0040.457 ± 0.0090.458 ± 0.005
TargetMethod5K10K15K20K
Mistral-7BNative0.366 ± 0.0070.388 ± 0.0080.424 ± 0.0130.437 ± 0.011
Mistral-7BGemma2-2B →0.465 ± 0.0120.517 ± 0.0060.537 ± 0.0050.541 ± 0.006
Mistral-7BQwen3-4B →0.445 ± 0.0040.499 ± 0.0270.512 ± 0.0240.519 ± 0.016

Dear Reviewer,

As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed your concerns. We would be happy to answer any further question you might have. Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.