acceptodds
Under review as a conference paper at ICLR 2027

CacheLink: Functional KV Transport for Zero-Prefill Cross-Model Context Handoff

13 views

Abstract

KV caches make long-context serving efficient by reusing past computation, but model switching breaks this reuse and forces the receiver to prefill the context again. We study zero-prefill cross-model context handoff: translating a sender cache into a receiver-compatible state so that the receiver can continue directly without prefilling same context. Existing affine transport methods show that receiver KV states can be substantially reconstructed from sender caches. However, through extensive analyses, we find that reconstruction fidelity is a poor proxy for receiver utility. Similar K/V reconstruction errors can lead to substantially different attention patterns, predictive distributions, and downstream accuracy, because conventional reconstruction objectives overlook the query-dependent functional importance of K and V in receiver attention. We thus introduce , which learns a coupled K/V correction through the receiver behavior directly. Across four independently trained sender–receiver pairs and seven evaluation panels, improves the equal-weight mean over all methods for every pair. In matched diagnostics, changes K/V squared error by less than 0.4% while substantially changing receiver attention and next-token behavior. At 32K context, it reduces handoff latency by up to relative to receiver re-prefill. Our findings verify that reconstruction fidelity alone is insufficient and that preserving the receiver computation induced by the transported cache is important for effective cross-model KV transport. The anonymized code is available at https://anonymous.4open.science/status/CacheLink-anonymous-2750the accompanying repository.

Primary area:
foundation or frontier models, including LLMs
Keywords:
KV Cache Resue

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_tcoyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cross-model KV cache reuse to avoid redundant prefilling during model switching. The authors propose CacheLink, which introduces a coupled K/V residual correction module on top of a frozen Ridge mapping, optimized through receiver-side CE and KL supervision. Experiments across four model pairs and seven benchmarks show improvements over Ridge while preserving zero-prefill efficiency.

Critical strengths

Cross-model KV cache reuse is relevant to multi-model collaboration, model routing, and long-context inference. This paper addresses the quality degradation associated with cache transfer and attempts to optimize the transferred cache based on the receiver's actual predictive behavior rather than solely minimizing numerical KV reconstruction errors. The reported improvements across multiple model pairs suggest potential practical value.

Critical weaknesses

W1. Limited methodological novelty and insufficient validation of the proposed architecture.

Cross-model KV transfer, nonlinear correction, and receiver-side functional supervision have already been explored in prior work. For instance, Heo et al. studied the mismatch between reconstruction fidelity and downstream performance, while Dery et al. employed receiver-side continuation losses.

The main novelty appears to be the coupled K/V late-fusion residual architecture. However, the paper lacks comparisons with simpler alternatives, such as independent K/V residuals or a joint MLP under identical supervision. Consequently, it remains unclear whether the improvements stem from the proposed architecture or from combining existing techniques.

W2. Insufficient baseline comparisons and concerns about experimental reliability.

Although Ridge, CacheBridge, and LatentAlign are included, important comparisons are missing, particularly the nonlinear MLP mapping from Heo et al. and simple functionally supervised residual baselines. Other closely related learning-based approaches are also insufficiently discussed.

Moreover, LatentAlign scores zero across all long-context evaluations, while CacheBridge performs unexpectedly poorly on several tasks. Without sufficient validation of baseline implementations, training convergence, and hyperparameter selection, the fairness and reliability of these comparisons remain uncertain. Some training configurations are also insufficiently documented.

Questions for authors

Q1. Core technical novelty

Given that cross-model KV transfer, nonlinear correction, and receiver-side functional supervision have already been explored in prior work, could the authors clarify the core technical contribution that distinguishes CacheLink from existing approaches? In particular, what new capabilities or substantive advantages does the proposed coupled K/V late-fusion residual architecture provide compared with existing residual correction methods?

Q2. Necessity of the proposed architecture

Under the same Ridge initialization, training data, approximately matched parameter budget, and CE+KL supervision, does the proposed coupled K/V residual architecture consistently and significantly outperform simpler alternatives, such as independent K/V residual modules or a straightforward joint MLP? How do the current experiments demonstrate that the improvements arise specifically from the proposed architecture rather than additional model capacity or functional supervision?

Q3. Baseline selection and fairness

Why does the paper not include the nonlinear MLP mapping explored by Heo et al., or a simple Ridge-plus-residual mapper trained with the same functional objective, as direct baselines?

In addition, related approaches such as Dense Latent Communication, LatentPort, and SCD have explored cross-model state transfer, learned correction mechanisms, or functional optimization. Although these methods differ in architecture, available input states, and deployment constraints, could the authors further clarify the substantive differences between CacheLink and these approaches? How does the current baseline selection sufficiently support the claimed performance advantages? For methods that permit comparable experimental settings, could more direct comparisons be provided?

Q4. Unexpected baseline results

Why does LatentAlign obtain zero scores across all long-context evaluations? What explains the substantially degraded performance of CacheBridge on certain tasks? Have the authors verified the correctness of the baseline implementations, training convergence, and appropriate hyperparameter configurations? Could representative failure cases or reproduction results under the original experimental settings be provided to substantiate the validity of the comparisons?

Q5. Anonymous code accessibility and reproducibility

The anonymous code link provided in the paper could not be accessed from my environment. Could the authors verify the validity and access permissions of the link and provide an accessible anonymous repository with the necessary reproduction instructions, so that reviewers can inspect the implementation and verify the main experimental results?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tcoy (1/2)

We sincerely thank you for your evaluation of our work and constructive comments. Below we provide detailed clarifications regarding the raised concerns.

W1&Q1: What is CacheLink's core contribution? What substantive advantages does the coupled K/V late-fusion residual offer over simpler alternatives (MLP and independent K/V)?

Thank you for summarizing the related work and for your suggestions on the method design. Heo et al. studied Ridge and nonlinear MLP mappings based on KV reconstruction loss; Dery et al. translate KV caches through a shared latent space and train the translators using cross-entropy on the true suffix predicted by the receiver model. Through experimental analysis, CacheLink finds that:

  1. Reconstruction error does not fully reflect downstream performance.
  2. Both linear and nonlinear components are very important for cross-model KV-cache transfer.
  3. K and V are highly correlated, and joint prediction performs better.

These findings guide the subsequent design of CacheLink: we choose to learn a coupled K/V residual on top of a frozen Ridge mapping and train it using task CE together with forward KL to the native receiver's token output distribution.

To examine the effect of the coupled K/V late-fusion residual relative to simpler alternatives (MLP and independent K/V), we additionally evaluate independent K/V residual and joint MLP controls on Qwen3-8B→14B. Please refer to our responses to Q2 and Q3 for the detailed results.


Q2: Under the same Ridge initialization, training data, approximately matched parameter budget, and CE+KL supervision, does coupled K/V late fusion outperform independent K/V or a joint MLP? How do the experiments distinguish the effects of architecture, capacity, and functional supervision?

Thank you for the suggestion. To distinguish the effects of the training objective and residual architecture, we conducted additional ablation experiments on Qwen3-8B→14B:

MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Ridge43.0020.0012.5036.5032.7832.0010.0026.68
Independent K/V, MSE41.0018.0010.6734.0033.0135.6711.3326.24
Independent K/V, CE+KL76.1744.3327.6739.0035.0234.0025.6740.27
Joint MLP, MSE45.6717.5011.5035.0031.7433.6712.3326.77
Joint MLP, CE+KL84.5043.3332.1742.3325.0233.3321.0040.24
CacheLink, MSE46.1718.3313.0036.0030.7633.6712.3327.18
CacheLink81.9044.2830.0039.8043.0237.9335.4044.62

Functional supervision outperforms MSE: For independent K/V, CE+KL improves the mean score by 14.03 percentage points over MSE. The matched retrained Coupled K/V version scores 44.62, which is 17.44 percentage points higher than the MSE score of 27.18. Thus, functional supervision is a more reasonable optimization objective than MSE.

Coupled K/V late fusion outperforms the other architectures: Under the same CE+KL objective, CacheLink has a mean score of 44.62, independent K/V scores 40.27, and the joint MLP scores 40.24. CacheLink is higher by 4.35 and 4.38 points, respectively, with the three architectures having very similar parameter budgets. The coupled K/V approach used by CacheLink performs best.

AuthorsOpenReview ↗

Response to Reviewer tcoy (2/2)

W2&Q3: Baseline selection and fairness

Thank you for the suggestion. We provide additional comparisons with MLP (Heo) [1], Ridge+independent K/V residuals with functional supervision, and Dense Latent Communication [2]. LatentPort [3] targets hybrid architectures with GDN and requires transferring recurrent and convolutional states, while SCD [4] requires sender hidden-state encodings in addition to KV. These requirements go beyond our current handoff setting, which uses only the existing KV cache, so we have not performed direct comparisons. The results by dataset for Qwen3-8B→14B are as follows:

MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Ridge43.0020.0012.5036.5032.7832.0010.0026.68
MLP (Heo), 20 epochs51.0012.508.0037.0032.2131.006.0025.39
MLP (Heo), 40 epochs49.0021.5020.0037.0031.4030.0011.0028.56
Independent K/V residual, CE+KL76.1744.3327.6739.0035.0234.0025.6740.27
Dense Latent Communication40.0016.007.0028.000.0015.000.0015.14
CacheLink81.9044.2830.0039.8043.0237.9335.4044.62

MLP (Heo) obtains mean scores of 25.39 and 28.56 after 20 and 40 epochs, respectively, while Dense Latent Communication collapses on the three long-context datasets. Under the same Ridge, CE+KL, data, and approximately matched parameter budget, the independent K/V residual scores 40.27, while CacheLink scores 44.62.

We will also expand the related-work discussion: LatentPort additionally transfers the recurrent and convolutional states of hybrid models, while the full SCD method also uses Patch information generated from sender hidden states. Our method uses only the sender's KV cache and does not depend on other inputs.


Q4: Unexpected baseline results

We reproduced CacheBridge using the official code and conducted an experiment on HellaSwag for Qwen3-14B→32B to verify the correctness of the CacheBridge reproduction:

SettingNative 32BCacheBridgeRetention relative to Native
Original paper v180.277.496.51%
Our reproduction82.479.696.60%

However, on the long-context datasets we use, such as BABILong, we observe that the receiver under CacheBridge often claims that it cannot find the information and therefore fails to provide the correct answer.

LatentAlign does not provide official code, and its authors used their own pretrained Gemma-2 architecture models of approximately 100M–400M parameters to study a shared latent space. We therefore reproduced LatentAlign according to the original paper's description, implementing LatentAlign for our current model pairs, and have added relevant reproduction details to the paper.


Q5: Anonymous code accessibility and reproducibility

Thank you for pointing out that the anonymous code link could not be accessed. We have corrected the anonymous repository. The repository provides CacheLink's core implementation, Ridge fitting, training and evaluation scripts, as well as dependency and usage instructions.

Thank you again for your valuable feedback.

[1] Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse, arXiv:2608.03893.

[2] See What I See, Know What I Think: Dense Latent Communication Across Heterogeneous Agents, arXiv:2606.13594.

[3] LatentPort: Beyond KV Cache—Cross-Model Transfer of Recurrent Memory in Hybrid Language Models, arXiv:2609.25053.

[4] Semantic Cache Distillation: Efficient State Transfer via Reuse and Selective Patching, ICML 2026.

thanks for your clarification.

Dear Reviewer tcoy

Thank you for your reconsideration and for raising the score. We sincerely appreciate your thoughtful evaluation and are glad that our clarifications helped address your concerns. Your recognition and support are truly encouraging to us.

Best regards,

Authors

Reviewer_HJB5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Switching between language models typically requires the receiver to prefill the same context again because the sender’s KV cache cannot be directly reused. This paper proposes CacheLink for zero-prefill cross-model KV-cache transfer. The authors argue that KV reconstruction fidelity alone does not reliably determine downstream utility. Building on Ridge, a KV-cache transfer baseline that minimizes KV reconstruction loss, CacheLink introduces coupled residual corrections for the key and value states and trains the residual model using cross-entropy and KL-divergence objectives. Across four model pairs and seven evaluation panels, CacheLink outperforms Ridge in 26 of 28 comparisons and achieves up to 7.27× lower handoff latency than receiver re-prefill at 32K context.

Critical strengths
  1. The paper presents insightful observations about cross-model KV-cache transfer. Lower reconstruction error does not necessarily lead to better downstream performance, while small cache corrections can substantially change receiver behavior without materially changing reconstruction error.
  2. CacheLink is effective at improving downstream performance. As shown in Table 1, it outperforms Ridge in 26 of 28 model–task comparisons and improves the equal-weight mean score by 2.96–16.01 points across the four model pairs.
Critical weaknesses
  1. The motivation for coupled K/V correction is not sufficiently established. Specifically, (1) Section 4.3 does not clearly define the claimed shared latent structure between K and V, nor does it provide direct empirical evidence that such a shared structure exists; (2) Figure 5 only shows that jointly replacing Ridge-based K/V with native K/V is beneficial, but does not demonstrate that joint correction remains advantageous when Ridge-based K/V is replaced by the proposed functional-residual K/V; and (3) the rationale for adopting feature fusion as the mechanism for joint K/V correction is insufficiently explained.
  2. The paper lacks sufficient ablation studies to justify its design choices. In particular, the individual contributions of Ridge-based initialization, feature fusion, and coupled K/V correction are not separately evaluated. These ablations would strengthen the evidence supporting the proposed design.
Questions for authors
  1. Can CacheLink support KV-cache transfer between models from different families?
  2. Why does CacheLink show larger JS divergence from the native receiver than Ridge on several panels in Figure 3(d)(e)?
6 comments
AuthorsOpenReview ↗

Response to Reviewer HJB5 (1/3)

We sincerely thank you for your constructive comments. Below we provide detailed clarifications regarding the raised concerns.


W1: Are the motivation and empirical evidence for joint K/V correction sufficient?

  1. Shared Structure of K and V

First, within a layer of the receiver model, K and V are both obtained by projecting the same hidden representation, which can be written as K=hWKK = hW_K and V=hWVV = hW_V. There are also precedents for shared K/V representations: MLA [1] jointly represents K and V through a shared low-dimensional latent representation, while AQUA-KV [2] empirically shows strong dependencies between K and V in the same layer. We also show the dependency between K and V in Figure 5. We have added these references to the paper.

We further provide two sets of experiments to demonstrate the strong relationship between K and V. We select 8 layers in the Qwen3-14B receiver and test on 100 FineWeb documents, 20 MultiDoc2Dial documents, and 18 ContractNLI documents. We use a rank-64 linear mapping. The table below reports the test-set R2R^2 averaged across the eight layers:

Prediction targetDirectionHeld-out FineWebMultiDoc2DialContractNLI
Native cacheK → V0.4060.3890.389
Native cacheV → K0.6600.6420.643
Ridge transfer errorK error → V error0.1700.1710.166
Ridge transfer errorV error → K error0.2160.2270.216

The R2R^2 values for directly predicting V from K and K from V reach approximately 0.4 and 0.6, respectively, indicating a high degree of linearity. In contrast, when Ridge's K and V transfer errors are used to predict each other, R2R^2 is only approximately 0.2, indicating a low degree of linearity. This also motivates adding a nonlinear residual module.

Therefore, the true K–V pairings carry information that can be used to predict the other side, exhibiting substantial shared structure.

  1. Does Learned Joint K/V Correction Help?

We compare independent K/V prediction with the coupled K/V prediction used by CacheLink:

MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Independent K/V (MSE)41.0018.0010.6734.0033.0135.6711.3326.24
CacheLink (MSE)+5.17+0.33+2.33+2.00−2.25−2.00+1.00+0.94
Independent K/V (CE+KL)76.1744.3327.6739.0035.0234.0025.6740.27
CacheLink (CE+KL)+5.73−0.05+2.33+0.80+8.00+3.93+9.73+4.35

The mean score difference between coupled K/V and independent K/V is +0.94 percentage points under MSE and +4.35 percentage points under CE+KL. Joint K/V correction achieves better transfer performance. K and V are computed from the same vector, and CacheLink uses this prior in its coupled K/V fusion prediction design.

  1. Rationale for Using Feature Fusion as the Joint K/V Correction Mechanism

We do not use a more complex joint correction mechanism for two reasons:

  1. This design is motivated by the fact that K and V are projected from the same hidden representation. We therefore use feature fusion to model the hidden representation, with linear mappings used separately before and after fusion.
  2. We aim to keep the method as simple as possible and avoid introducing too many variables, so that the effect of our insight can be directly demonstrated.
AuthorsOpenReview ↗

Response to Reviewer HJB5 (2/3)

W2: More detailed ablation studies are needed: for example, the individual contributions of Ridge-based initialization, feature fusion, and coupled K/V correction have not been separately evaluated.

Thank you for the suggestion. We find that there is a clear linear relationship between cross-model K and cross-model V, as shown in Figure 6 of the paper. We therefore use Ridge to first fit the linear relationship as a good starting point, making use of this prior, whereas randomly initialized CacheLink has greater difficulty fitting the linear relationship because of the randomness of gradient descent. We further provide ablation experiments on the Ridge base, functional supervision, and nonlinear residuals, all of which improve transfer performance. For the joint K/V correction ablation, please refer to our response to W1.

  1. Ridge-Based Initialization
MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Ridge47.5020.7815.5134.4041.2031.909.0028.61
CacheLink (random initialization)33.0010.505.5029.001.0012.002.0013.29
CacheLink (Ridge-Based Initialization)81.9044.2830.0039.8043.0237.9335.4044.62

The Ridge base brings clear benefits. Randomly initialized CacheLink directly predicts the full receiver cache and achieves an average of 13.29, even below Ridge; CacheLink with the Ridge base scores 44.62. According to Figure 6 of our paper, cross-model K/V contains substantial linear components (with R2R^2 of approximately 0.8 and 0.6, respectively). Therefore, Ridge-based initialization can provide a good starting point for gradient-based optimization and significantly improve fitting.

  1. Functional Supervision vs. MSE
MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Independent K/V+MSE41.0018.0010.6734.0033.0135.6711.3326.24
Independent K/V+CE+KL+35.17+26.33+17.00+5.00+2.01−1.67+14.33+14.03
CacheLink (MSE)46.1718.3313.0036.0030.7633.6712.3327.18
CacheLink (functional supervision)+35.73+25.95+17.00+3.80+12.26+4.26+23.07+17.44

For independent K/V, CE+KL improves the mean score by 14.03 percentage points over MSE. The matched retrained Coupled K/V version scores 44.62, which is 17.44 percentage points higher than the MSE score of 27.18. Thus, functional supervision is a more reasonable optimization objective than MSE.

  1. Linear vs. Nonlinear Residuals
MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
CacheLink (linear)70.5036.0013.5035.500.321.000.0022.40
CacheLink (nonlinear)+11.40+8.28+16.50+4.30+42.70+36.93+35.40+22.22

The nonlinear version has a mean score of 44.62, while the linear version scores 22.40, a difference of 22.22 percentage points. The nonlinear residual performs much better than the linear residual, particularly on the three long-context datasets, where the model based on the linear residual performs close to zero. As noted in Figure 6 of our paper, cross-model K/V translation contains a non-negligible nonlinear component (1−R2(K)≈0.21-R^2(K)\approx 0.2, 1−R2(V)≈0.41-R^2(V)\approx 0.4); we therefore choose a nonlinear residual for further fitting.

AuthorsOpenReview ↗

Response to Reviewer HJB5 (3/3)

Q1: Can CacheLink support KV-cache transfer between models from different families?

We conducted an additional Qwen3-8B→Llama-3.1-8B-Instruct cross-family experiment, as shown below:

MethodOolong-Synth 32KLongBench-v2BABILong 32KMean
Number of examples100100100—
Native Qwen 8B35.9927.0035.0032.66
Native Llama 8B13.4227.0041.0027.14
Ridge5.0011.002.006.00
CacheLink9.0032.0022.0021.00

A key difficulty in cross-family handoff is that the token boundaries of the two tokenizers do not match. We use a causal alignment heuristic based on character-end positions, drawing on HeteroFold [3]. CacheLink scores 15.00 points higher than Ridge on average, exceeding native Llama on LongBench-v2 but remaining below native Llama on Oolong and BABILong.

Our current character-boundary alignment is relatively simple and does not explicitly model the semantic correspondence between the two models. In the future, if direct cross-tokenizer latent-space alignment can be learned while preserving causality, it may reduce alignment errors and further improve transfer performance. We consider this a direction for future work.


Q2: Why does CacheLink show larger JS divergence from the native receiver than Ridge on several panels in Figure 3(d)(e)?

Thank you for pointing out this phenomenon. The purpose of Figure 3 is to show that a small cache correction can substantially change the receiver's attention and predictive distributions. In Figure 4, we further analyze the relationship between changes in KV reconstruction error and changes in JS divergence. The results show that the receiver's behavior can still change substantially even when KV reconstruction error does not improve. Therefore, cache reconstruction error alone is insufficient to fully reflect the effect of the translated cache on receiver computation.

Thank you again for your valuable feedback.

[1] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model. arXiv:2405.04434.

[2] Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models. ICML, 2025.

[3] Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs. arXiv:2609.32259.


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

Switching between LLMs requires the receiving model to process the same context again because KV caches cannot be directly reused across models. Existing cache-transfer methods address this by mapping the sender’s KV cache into the receiver’s cache space, but accurate reconstruction does not preserve downstream accuracy. To address this, the paper introduces CacheLink. It adds learned, coupled K/V corrections to a linear mapping and trains them through a frozen receiver using continuation and distillation losses. CacheLink is evaluated across four model pairs and shows that it works better than reconstruction-based cache transfer while giving up to 7.27× faster handoff than receiver re-prefill at 32K context.

Critical strengths
  • The analysis that shows that nearly identical KV reconstruction errors can produce substantially different receiver behavior and task accuracy is very well done.
  • CacheLink outperforms Ridge on almost all the benchmarks across four model pairs.
Critical weaknesses
  • The quality gaps on some long-context tasks remain substantial. This limits the practical benefit of switching models, despite the faster handoff.
Questions for authors
  • CacheLink introduces 3 components: functional supervision, additional nonlinear capacity, and joint K/V correction. Ablations isolating these 3 components would help understand the contribution of each component.
4 comments
AuthorsOpenReview ↗

Response to Reviewer qAcS (1/2)

We sincerely thank you for your positive evaluation of our work and constructive comments. Below we provide detailed clarifications regarding the raised concerns.

W1: The quality gaps on some long-context tasks remain substantial, limiting the practical benefit of switching models.

We would like to emphasize that, on long-context tasks, CacheLink is only slightly behind the native receiver in most scenarios while clearly outperforming the other cache translation methods. To quantify this gap, the following table summarizes the change in each method's average score on the three long-context tasks in the paper's main table, relative to the native receiver.

MethodQwen3-14B to 32BMinistral 3-8B to 14BQwen3-8B to 14BYi-1.5-9B to 34BAverage
Ridge Δ-9.03-4.63-19.71-16.64-12.50
CacheBridge Δ-26.13-9.65-34.92-26.22-24.23
LatentAlign Δ-46.49-46.45-47.08-31.19-42.80
CacheLink Δ-0.38-6.18-8.30-4.97-4.96

For three of the four model pairs, CacheLink's performance is closer to that of the native receiver than the other translation methods in the table. For Qwen3-14B→32B, the average score across the three long-context panels is only 0.38 percentage points lower. Overall, the results show a trend in which larger model pairs incur smaller transfer performance losses; investigating this phenomenon will be a direction for future work.

To quantify this trade-off, we compare handoff costs in the paper's latency experiments when the sender cache is already available. With a 32K prefix, receiver re-prefill takes 7.039 seconds for Qwen3-14B→32B, while CacheLink handoff takes 0.969 seconds, a 7.27× speedup. For Ministral 3-8B→14B, the corresponding times are 2.507 seconds and 0.406 seconds, a 6.18× speedup.


AuthorsOpenReview ↗

Response to Reviewer qAcS (2/2)

Q1: Ablation experiments that separately isolate functional supervision, nonlinear capacity, and joint K/V correction.

Table 2 in the paper's appendix reports ablations of training objectives such as CE and KL. To further distinguish the three components raised by the reviewer, we provide three sets of ablation comparisons for the Qwen3-8B→14B model pair. The results show that functional supervision, nonlinear capacity, and joint K/V correction all improve transfer performance.

Table 1: Functional Supervision vs. MSE

MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Independent K/V (MSE)41.0018.0010.6734.0033.0135.6711.3326.24
Independent K/V (functional supervision)+35.17+26.33+17.00+5.00+2.01−1.67+14.33+14.03
CacheLink (MSE)46.1718.3313.0036.0030.7633.6712.3327.18
CacheLink (functional supervision)+35.73+25.95+17.00+3.80+12.26+4.26+23.07+17.44

For independent K/V, CE+KL improves the mean score by 14.03 percentage points over MSE. The matched retrained Coupled K/V version scores 44.62, which is 17.44 percentage points higher than the MSE score of 27.18. Thus, functional supervision is a more reasonable optimization objective than MSE.

Table 2: Linear vs. Nonlinear Residuals

MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
CacheLink (linear)70.5036.0013.5035.500.321.000.0022.40
CacheLink (nonlinear)+11.40+8.28+16.50+4.30+42.70+36.93+35.40+22.22

The nonlinear version has a mean score of 44.62, while the linear version scores 22.40, a difference of 22.22. The nonlinear residual performs much better than the linear residual, particularly on the three long-context datasets, where the model based on the linear residual performs close to zero. As noted in Figure 6 of our paper, cross-model K/V translation contains a non-negligible nonlinear component (1−R2(K)≈0.21-R^2(K)\approx 0.2, 1−R2(V)≈0.41-R^2(V)\approx 0.4); we therefore choose a nonlinear residual for further fitting.

Table 3: Independent K/V vs. Coupled K/V Correction

MethodSciQMMLU-ProMMLU-Pro+OBQAOolongLB-v2BABILongMean
Independent K/V (MSE)41.0018.0010.6734.0033.0135.6711.3326.24
CacheLink (MSE)+5.17+0.33+2.33+2.00−2.25−2.00+1.00+0.94
Independent K/V (CE+KL)76.1744.3327.6739.0035.0234.0025.6740.27
CacheLink (CE+KL)+5.73−0.05+2.33+0.80+8.00+3.93+9.73+4.35

The mean score difference between coupled K/V and independent K/V is +0.94 percentage points under MSE and +4.35 percentage points under CE+KL. K and V are computed from the same vector, and CacheLink uses this prior in its coupled K/V fusion prediction design.

Thank you again for your valuable feedback.


Thank you for your response and the additional results. I will retain my score of 3.

Dear Reviewer qAcS

Your feedback has greatly helped us improve the clarity and completeness of the paper. Thank you very much!

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal Discussion

Dear ICLR 2027 Area Chair and Reviewers,

Thank you for your careful reading and constructive feedback. We would like to summarize the additional evidence we provided and the reviewers’ follow-up comments.

For Reviewer qAcS, we quantified the trade-off between long-context task quality and handoff latency. With the sender’s cache already available, CacheLink was up to 7.27× faster than receiver re-prefill at 32K context. We also added ablations on functional supervision, residual nonlinearity, and joint K/V correction. Reviewer qAcS thanked us for the additional results and retained their weak-accept recommendation.

For Reviewer HJB5, we measured shared structure in both native K/V and Ridge transfer errors, tested the contribution of Ridge initialization, and clarified the attention and output distribution analyses. In a new Qwen3-to-Llama cross-family experiment, CacheLink scored 15.0 points above Ridge on average across three long-context tasks. Reviewer HJB5 stated that the new results addressed their main concerns and indicated that they would raise their score.

For Reviewer tcoy, we clarified CacheLink’s contribution relative to prior work and added comparisons with simpler residual architectures, Heo’s nonlinear MLP, and Dense Latent Communication. And we corrected the anonymous repository link. Reviewer tcoy raised their score.

CacheLink enables zero-prefill cross-model handoff by learning coupled K/V residuals. Training through the frozen receiver uses continuation supervision and forward KL to improve how the receiver uses the transferred cache. Across all four evaluated model pairs, CacheLink improved mean downstream performance over Ridge while avoiding receiver prefill on the reusable prefix.

Thank you for considering the paper and the additional evidence.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.