{"id":"c888e559-a6ae-4c17-bec9-4d667178c90a","arxiv_id":"2608.03893","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Cross-model KV cache transfer within LLM families is feasible with a per-head ridge regression, retaining 73-98% of accuracy on four matched-KV pairs.","lead":"A method that reuses one LLM's KV cache when switching to a different-sized model in the same family, skipping the expensive prefill pass. A closed-form linear mapping trained on 500 sequences retains 73-98% of target accuracy on four of six pairs, with two pairs failing badly.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Calibration-to-eval distribution shift is the load-bearing risk: Table 4 shows deeply negative eval-domain R2_K on two Ministral pairs, so the four-pair 73–98% retention claim is not yet shown to generalize beyond FineWeb-Edu-like prompts.","rationale":"The paper is careful and transparent: it reports two Tier 2 failures, includes leave-one-benchmark-out k-selection analysis (Appendix H), and proposes attention-output cosine as a post-hoc diagnostic. I nonetheless think the calibration-representativeness assumption is the hinge for the central 'practical' claim. The four successful pairs are all evaluated on five short-context benchmarks that resemble the FineWeb-Edu register; the two failures show the same linear fit can catastrophically mispredict on another distribution (Table 4). Because no a priori screening exists to separate Tier 1 from Tier 2 before fitting, the operational claim that a deployment can 'skip prefill' with a closed-form mapper is not yet established for arbitrary prompts. The proposed concrete test would settle whether the four successes are robust to distribution shift or artifacts of benchmark overlap. If it passes, the conditional verdict can be upgraded; if it fails, the claim needs a calibration-selection safeguard. This does not change the reader's conditional verdict.","tokens_in":18641,"tokens_out":8708,"duration_ms":82505,"concrete_test":"Using the exact FineWeb-Edu calibration and each pair's selected k from §4.1, evaluate the four Tier 1 pairs on a held-out suite not used for selection or analysis—e.g., MMLU-Pro, GPQA, PubMedQA, HumanEval, plus an 8k-token long-context QA set—and compute raw and floor-normalized retention per benchmark. If any Tier 1 pair falls below 73% raw retention on the aggregate or below 60% floor-normalized on a majority of the new benchmarks, the calibration-representativeness assumption fails and the 'practical' claim must be narrowed. This directly tests the condition that the calibration set's linear fit transfers to the evaluation distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that cross-model KV transfer is practical rests on the 500-sequence FineWeb-Edu calibration being representative of the prompts a deployed mapper will see (§3.1). The paper's own Table 4 contradicts that premise for the two failure pairs: on HellaSwag tokens, ridge's eval-domain R2_K is -7.81 (3B→14B) and -3.22 (8B→14B), meaning the calibration-fit linear map is worse than predicting the mean target key. Even the Tier 1 Ministral 3B→8B pair shows R2_K = -0.10 on eval tokens, so negative eval R2 does not by itself destroy accuracy (attention-output cosine is the better diagnostic). The problem is that this diagnostic is post-hoc: it requires the fitted mapper and the target model on the eval distribution. There is no pre-fit screening to tell a deployment which matched-KV pairs will land in Tier 1. Appendix C measures calibration-domain sensitivity only on Qwen3 14B→32B; it does not bound the domain shift for the other pairs. Therefore the 73–98% success may be partly a lucky draw on FineWeb-Edu-like benchmarks, and the 'practical' claim is conditional on the deployment distribution resembling the calibration domain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes cross-model KV cache transfer: for matched-KV pairs within an LLM family, the source model's KV cache is mapped to the target via per-head ridge regression. The mapper selects the top-k most predictive source layers per target layer, strips RoPE from keys before fitting and reapplies target RoPE at inference, and is calibrated on 500 FineWeb-Edu sequences of 1,024 tokens. The method is evaluated on six matched-KV pairs across Qwen3, Llama 3.1, and Ministral 3. Four pairs retain 73–98% of the target's standalone accuracy across five benchmarks, while two Ministral pairs degrade sharply (42–44% retention). An MLP recovers +24.3 to +36.8 pp HellaSwag retention on the failures. The mapper runs 2.7–25x faster than re-prefill and shows small multi-turn CoQA drift. The paper also reports ablations, calibration sensitivity, leave-one-out k re-selection, and held-out benchmarks.","tokens_in":19001,"tokens_out":6823,"duration_ms":63243,"significance":"If the result holds, cross-model KV cache transfer would be a meaningful prefill-skipping technique for multi-model serving, and the finding that substantial linear structure exists across matched-KV pairs is valuable. The paper is honest about the two failing pairs and provides useful ablations, calibration sensitivity analysis, held-out benchmarks, and a bounded analysis of k selection. These are strengths. However, the practical claim is only partially supported: success is demonstrated on a subset of pairs, the calibration distribution is a single domain, and no pre-fit screening identifies which pairs will succeed. The paper is not internally inconsistent, but the abstract and conclusion claim practicality more strongly than the evidence supports.","major_comments":[{"comment":"The abstract and conclusion claim the mapper is practical, but this rests on the 500-sequence FineWeb-Edu calibration set being representative of deployment prompts. Table 4 shows eval-domain ridge R2_K = -7.81 (3B→14B) and -3.22 (8B→14B), and even the Tier 1 Ministral 3B→8B pair has R2_K = -0.10 on HellaSwag tokens; the paper's own text says the calibration-fit linear mapper 'does not extrapolate to HellaSwag tokens.' Appendix C measures calibration-domain sensitivity only on Qwen3 14B→32B (CodeAlpaca costs 5.24 pp on HellaSwag), so it does not bound the domain shift for the failure pairs. Because attention-output cosine is computed post hoc using the fitted mapper and target on the eval distribution, it cannot serve as a deployment-time screening signal. The practical claim needs either a pre-fit transferability criterion, multi-domain calibration with coverage of failing pairs, or a s","section":"§4.5, Table 4; §3.1; Appendix C"},{"comment":"The headline retention numbers are partially in-sample: k is selected per pair by maximizing accuracy on ARC-C, HellaSwag, WinoGrande, and MMLU, the same benchmarks averaged in Table 1. Appendix H estimates the maximum inflation at 2.49 pp via leave-one-out re-selection, which is useful and transparent, but the reported Avg retention should be presented as 'selected-k' rather than as an out-of-sample estimate, or supplemented with a fixed-k analysis. This matters most for Ministral 8B→14B, where leave-one-out changes k from 12 to 20 and moves the held-out benchmark by 2.49 pp. The central 73–98% claim is not invalidated, but its precision is weaker than the text suggests.","section":"§4.1; Table 1; Appendix H"},{"comment":"The paper's central success story is that four of six pairs work, but no architectural or calibration-time property is shown to separate the two tiers before fitting; matched KV correlates with success but does not guarantee it. The proposed predictor, attention-output cosine, is measured post hoc on the fitted mapper and target using eval-distribution tokens. Its reported r=+0.57 is computed over 12 evaluations drawn from only six pairs in both directions, so the effective sample size is small and the observations are not independent. The claim that this diagnostic 'predicts cross-pair retention' is therefore not established at the level the practical story requires. I would ask for confidence intervals, an independence-aware analysis, and a clear statement that cosine is a post-hoc diagnostic rather than a pre-fit predictor.","section":"§4.2; §4.5"}],"minor_comments":[{"comment":"The term 'closed-form' may mislead. Equation (4) is the ridge normal equations solution fit from calibration data, not a parameter-free derivation from model weights. Suggest 'closed-form linear-solve mapper' or 'gradient-free mapper' and state explicitly that calibration requires supervised access to target KV.","section":"Title; §3.1"},{"comment":"No uncertainty or confidence intervals are reported. Many retention differences are within a few points (e.g., 97.6% vs 98.5% Avg on Qwen3 14B→32B). Report error bars or state that evaluation is deterministic under fixed prompts.","section":"Table 1; §4.1"},{"comment":"The meaning of 'eval-domain R2_K' and how deeply negative values arise should be defined precisely. R2 is not bounded below once evaluated on a shifted distribution, and the reader must infer the token subset, the baseline mean, and the exact computation.","section":"Table 4"},{"comment":"Domain sensitivity is measured only on Qwen3 14B→32B. A sentence explaining whether this pair is representative of the other families, or why it is not, would help the reader calibrate the limitation.","section":"Appendix C"},{"comment":"The correlation r=+0.57 should be accompanied by the number of points, a p-value or confidence interval, and a caveat about non-independence across directions and families.","section":"§4.5"},{"comment":"The manual override of Ministral 3B→8B to k=all is noted in appendix, but a brief justification in the main text would avoid confusion, especially because Table 1 reports k=all while the selection criterion would have chosen k=20.","section":"Appendix D; §4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is worth a revision rather than rejection. The empirical findings, honest reporting of failures, and thorough ablations are valuable. The main risk is overclaiming practicality: the abstract and conclusion state that cross-model KV transfer is practical, but the evidence is conditional on a single-domain calibration set and on post-hoc identification of successful pairs. If the authors re-scope the claims and add failure-pair domain analysis or a screening mechanism, the paper could be acceptable. The lack of code release was not treated as a blocking issue but would strengthen reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is a solid, honestly-written empirical contribution to the LLM-serving literature. The headline claim — closed-form ridge mapping of cross-model KV caches works on four of six matched-KV pairs and preserves 73–98% of standalone accuracy — holds up as far as I can tell from the reported numbers. It doesn't open a new paradigm; it makes a specific optimization plausible, and the paper is candid that two of six pairs fail.\n\nWhat is actually new: the observation of substantial linear structure in cross-model KV after RoPE stripping, the per-head ridge mapper with top-k source-layer selection, and the empirical result that attention-output cosine predicts per-pair retention better than R2. The ablations are real: component removal, calibration sensitivity to λ, N and domain, leave-one-benchmark-out k re-selection, held-out benchmarks, multi-turn drift, and latency across seven pairs. That is more reproducible scaffolding than most papers at this level, and the limitations section names the main gaps rather than hiding them.\n\nThe soft spots are proportionate. The stress-test concern about calibration-to-eval shift is real: Table 4 shows deeply negative eval-domain R2 on both Ministral failures, and even one Tier 1 pair has negative K R2 on HellaSwag tokens. So the “practical” claim is conditional on the deployment distribution resembling FineWeb-Edu-like prompts, and the paper offers no pre-fit screening to tell an operator which matched-KV pair will land in Tier 1. That is a genuine gap. It is not fatal to the paper’s core claim, because the paper already scopes success to the four pairs and labels the rest as failures. But a deployment-minded reader should not assume the method will transfer to arbitrary prompt distributions without checking the mapped cache’s attention-output cosine on representative data.\n\nOther weaknesses: k is selected on eval benchmarks, though the leave-one-out bound of 2.49 pp is a reasonable mitigation; there are no error bars over calibration draws or eval seeds; large-to-small transfer accuracy is only shown on HellaSwag. These are minor in context. The citation pattern looks appropriate for a systems/ML paper.\n\nWho is it for: anyone working on multi-model serving, model routing, or KV-cache reuse. It is not a theory paper and not a general claim about all LLM families. It deserves a serious referee: the right reviewer will ask for more pairs, a real screening signal, and variance estimates, but this is not a desk reject. I’d bring it to a reading group and would likely cite it when writing about prefill reuse.","headline":"A solid, honestly-scoped empirical paper on cross-model KV transfer; the central claim holds for its four successful pairs, but the calibration-to-eval distribution shift is an unresolved deployment risk.","tokens_in":19430,"tokens_out":3141,"would_cite":true,"duration_ms":30605,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cross-model KV cache transfer via closed-form per-head ridge mapping can skip receiver prefill in model-family swaps, retaining 73-98% of standalone accuracy on four of six matched-KV pairs.","keywords":["KV cache reuse","cross-model transfer","prefill acceleration","linear mapping","ridge regression","rotary position embedding","LLM serving","model families"],"falsifier":"Apply the paper's exact protocol to a matched-KV pair from an unseen fourth family, fitting on 500 FineWeb-Edu sequences and measuring floor-normalized HellaSwag retention; if the linear map consistently shows high calibration R2 yet near-zero retention, the claim that matched KV supports a transferable linear fit is falsified.","tokens_in":18636,"feed_emoji":"⚡","tokens_out":9563,"duration_ms":86086,"temperature":0.7,"pith_summary":"Production settings often swap between different-sized models in the same family, and each swap forces the receiving model to recompute prefill from scratch, which is costly for long contexts. This paper tries to establish that the first model's KV cache can be linearly transformed into the second model's format, allowing the receiver to skip prefill entirely. The central claim is that, when source and target share KV head count and per-head dimension, a closed-form per-head ridge regression fitted on just 500 calibration sequences retains 73-98% of standalone accuracy on four of six tested pairs and runs 2.7-25x faster than re-prefill. Two Ministral pairs degrade sharply, and a nonlinear MLP recovers most of the gap. If this holds, multi-model serving can reuse prefill work across model swaps without gradient-based adapters.","feed_headline":"Linear map lets a sibling model reuse the KV cache, no re-prefill","feed_subtitle":"A 500-sequence fit keeps 73-98% accuracy on four family pairs and cuts swap latency up to 25x.","key_machinery":"The central object is a per-head closed-form ridge mapper. For each target layer and head, it concatenates key and value features from the top-k most predictive source layers and solves a Tikhonov-regularized least-squares problem, W* = (X^T X + lambda I)^-1 X^T Y, to project them into the target's KV space. Keys are mapped after stripping RoPE and re-encoded with target RoPE, making the fit position-free and reusable across context lengths. Cross-layer source selection, not reconstruction quality alone, carries most of the transfer performance.","core_discovery":"For matched-KV pairs, the target model's keys and values are largely linear functions of a small set of source layers: on Qwen3 14B to 32B, one source layer explains 56% of key variance and multiple source layers reach 79%. The authors build a per-head ridge mapper that strips rotary position embeddings from keys, selects the top-k most predictive source layers for each target layer, fits a closed-form linear regression in position-free content space, and re-applies the target's rotation. This mapper retains 73-98% of standalone accuracy on four of six matched-KV pairs, remains stable over multi-turn handoff, and is faster than re-prefill at every measured sequence length, while two harder p","pith_inferences":["The RoPE-stripped, position-free design suggests the same fitted weights should apply at context lengths far beyond the 1,024-token calibration window; the paper measures latency at 32K but does not report whether accuracy holds there, so a long-context accuracy test is the natural next step.","Nothing in the ridge solver requires matching KV head counts or per-head dimensions, so the same closed-form machinery could be tried on mismatched-KV pairs; the paper explicitly leaves this untested.","Attention-output cosine is measured only after fitting; a pre-fit proxy built from static source-target attention similarity on a few shared prompts could turn the diagnostic into a screening tool, though the paper does not construct one.","The two Ministral failures suggest that training-recipe or data-recipe differences, not just architecture, decide how linear the cross-model KV relationship is; comparing matched-architecture pairs with deliberately different fine-tuning data would separate those factors."],"forward_implications":["A serving stack can skip the receiver's prefill during model-family swaps by applying one batched matrix multiply per target layer, cutting prefill latency by 2.7-25x at long context.","The fit is gradient-free and small-scale: 500 calibration sequences and one node fit a pair in under an hour, with no backpropagation through either model.","Transfer quality depends on where residual error lands relative to attention-sensitive directions, not on average reconstruction error, so attention-output cosine predicts cross-pair retention better than R2 does.","A nonlinear MLP drop-in recovers much of the accuracy that linear ridge loses on harder pairs, with HellaSwag retention gains up to 36.8 percentage points.","Multi-turn handoff between two Qwen3 sizes accumulates only small drift over ten turns, indicating that mid-conversation model switching does not cascade immediately."],"supporting_citations":[{"why":"Supplies the RoPE formulation whose rotation is stripped before regression and re-applied after mapping, making the fit position-free.","marker":"[Su et al., 2024]"},{"why":"The neural-fuser baseline that the closed-form ridge mapper is designed to replace, showing prior cross-model KV transfer required gradient training.","marker":"[Fu et al., 2026]"},{"why":"Learned latent-space adapter baseline that the paper contrasts with a closed-form, small-sample fit.","marker":"[Dery et al., 2026]"},{"why":"Attention-pattern mapping alternative; the paper contrasts it with direct KV value transfer.","marker":"[Zhao et al., 2025]"},{"why":"Prior KV sharing restricted to identical architectures, motivating the matched-KV relaxation studied here.","marker":"[Liu et al., 2026]"},{"why":"Provides the linear cross-model representation alignment background that motivates probing cross-model KV for linear structure.","marker":"[Huh et al., 2024]"},{"why":"Defines HellaSwag, the benchmark on which retention, failure, and MLP recovery are measured.","marker":"[Zellers et al., 2019]"},{"why":"Defines ARC-Challenge, one of the benchmarks used for k selection and retention evaluation.","marker":"[Clark et al., 2018]"}],"fun_headline_variants":["Linear map transfers KV caches between sibling LLMs, skipping prefill","Cross-model KV cache transfer via linear mapping: 73-98% accuracy, no re-prefill","Sibling LLM KV cache reuse: linear mapper, 2.7-25x faster than re-prefill","KV cache transfer works for 4/6 LLM family pairs using linear mapping"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The mapper is fitted on 500 FineWeb-Edu sequences of 1,024 tokens, and the entire method assumes that this small calibration sample is representative of the prompts the receiver will actually serve; the paper's own two Ministral failures show the premise can break.","fun_headline_variants_meta":{"raw":{"variants":["Linear map transfers KV caches between sibling LLMs, skipping prefill","Cross-model KV cache transfer via linear mapping: 73-98% accuracy, no re-prefill","Sibling LLM KV cache reuse: linear mapper, 2.7-25x faster than re-prefill","KV cache transfer works for 4/6 LLM family pairs using linear mapping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000368,"raw_usage":{"total_tokens":1862,"prompt_tokens":848,"completion_tokens":1014,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":916}},"tokens_in":592,"tokens_out":1014,"duration_ms":9860,"temperature":1.0,"reasoning_tokens":916,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:11:50.530601+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the paper's exact protocol to a matched-KV pair from an unseen fourth family, fitting on 500 FineWeb-Edu sequences and measuring floor-normalized HellaSwag retention; if the linear map consistently shows high calibration R2 yet near-zero retention, the claim that matched KV supports a transferable linear fit is falsified.","supporting_citations":[],"review_version":1}