{"id":"2d8afcb7-fc62-4e4b-9548-e1531a09e963","arxiv_id":"2508.16134","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CommonKV uses SVD-based cross-layer parameter sharing and cosine-similarity-guided budget allocation to compress the KV cache, reportedly reaching 98% compression without major loss.","lead":"This paper introduces CommonKV, a training-free way to compress the KV cache of large language models by sharing parameters across adjacent layers via SVD, plus an adaptive budget based on cosine similarity. The authors claim up to 98% compression without significant performance loss on long-context benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cosine-similarity budget allocation may not track low-rank spectral compatibility; high-similarity layer pairs could still be over-compressed, threatening the 98% claim.","rationale":"The reader's weakest assumption is the similarity/compressibility premise. I partially agree. The abstract stresses 'high similarity observed in cross-layer hidden states' as motivation, and the adaptive budget uses cosine similarity. However, cosine similarity is a direction-only measure that does not reflect the distribution of singular values. Low-rank approximation error depends on the singular-value spectrum, especially the tail energy. Merging two caches via shared SVD creates a shared subspace; the reconstruction error for each cache is the projection loss onto that subspace. This projection loss can be large even when the two caches point in similar directions, if one cache has significant energy in directions orthogonal to the shared subspace. In fact, two vectors with high cosine similarity can have very different ranks: one could be rank-1 and the other full rank with a large component orthogonal to the first. Thus cosine similarity fails as a compressibility metric. This is not an external consensus issue; it is an internal mismatch between the proposed metric and the objective. The lack of experimental detail in the abstract leaves open the possibility that the method only works when the heuristic aligns. A concrete spectral test can settle it. Since the paper is abstract-only, this keeps the verdict unverified.","tokens_in":650,"tokens_out":7124,"duration_ms":82879,"concrete_test":"Run CommonKV-style SVD sharing on a standard LLM (e.g., LLaMA-2-7B) with real inputs from LongBench. For each adjacent layer pair, compute (i) cosine similarity of the key/value caches and (ii) the shared spectral energy: the sum of squared singular values of the concatenated cache captured by the top-k shared singular vectors used by CommonKV. Then check: do high-cosine pairs always have high shared energy? If any pair has cosine > 0.9 but shared-energy fraction < 0.5, the budget allocation is misleading. As a stronger check, replace the cosine-based budget with a spectral-energy-based budget; if performance on LongBench improves, the concern lands.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"CommonKV's adaptive budget allocation (abstract: 'dynamically assigns compression budgets based on cosine similarity') treats pairwise cosine similarity between adjacent caches as the signal for how much compression is safe. But low-rank compressibility is governed by the singular-value spectrum and the energy captured by shared subspaces, not by cosine similarity alone. Two activation sequences can have high cosine similarity yet occupy almost orthogonal low-rank subspaces (e.g., one vector is a scaled version of another plus a large orthogonal component). In that case, merging via shared SVD parameters will discard a large fraction of the second cache's energy even though cosine similarity suggests safe compression. Conversely, low cosine similarity can coexist with high spectral overlap after centering. If the budget allocation is based on the wrong statistic, the method will systematically over-compress certain layer pairs; the 98% compression result, which likely includes quantization/eviction, would then reflect favorable cases rather than a general mechanism. The abstract provides no spectral analysis of the caches, so this cannot be dismissed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CommonKV, a training-free cross-layer KV cache compression method. It uses SVD-based parameter sharing across adjacent layers to create a mergeable latent cache, and an adaptive budget allocation strategy based on cosine similarity between caches. The abstract claims consistent improvement over existing low-rank and cross-layer approaches on LongBench and Ruler, and reports that combining with quantization and eviction can reach a 98% compression ratio without significant performance loss. This review is based solely on the abstract; the full text was not available.","tokens_in":963,"tokens_out":1783,"duration_ms":21283,"significance":"If the claims are substantiated, the contribution is practically meaningful: reducing KV cache memory by an order of magnitude without retraining, while remaining compatible with quantization and eviction, would directly benefit long-context serving. The adaptive budget idea is a sensible response to the known fragility of fixed cross-layer sharing. The claimed training-free property is also attractive. However, the significance is conditional; the abstract provides no experimental details, and the central numerical claim (98% compression) is not supported by evidence in the manuscript available for review.","major_comments":[{"comment":"The central claim '98% compression ratio without significant performance loss' is stated without supporting numbers, baselines, error bars, or a definition of 'significant performance loss'. At such extreme compression, low-rank methods typically exhibit sharp degradation; without experimental tables, the claim is unverifiable. This is load-bearing and must be supported with full results, including per-task accuracy and variance.","section":"Abstract"},{"comment":"The adaptive budget allocation uses cosine similarity between adjacent caches, but low-rank compressibility depends on spectral properties (singular-value decay and subspace overlap), not solely on cosine similarity. Two caches can have high cosine similarity while occupying nearly orthogonal low-rank subspaces, causing the shared SVD to discard energy from one cache. The authors should provide evidence that cosine similarity tracks spectral compatibility, e.g., a correlation analysis or an ablation comparing cosine-based budgets against spectral-based budgets on a few tasks.","section":"Abstract"},{"comment":"The method is described as 'training-free', but SVD-based parameter sharing across adjacent layers requires specifying what data the SVD is computed on. If the SVD is computed on hidden states from a calibration set, the method is data-dependent and may suffer from distribution shift; if computed only on pretrained weights, the claim of 'more easily mergeable latent cache' needs a derivation. The abstract is ambiguous about this, and the ambiguity affects the generality of the 98% claim.","section":"Abstract"}],"minor_comments":[{"comment":"The phrase '98% compression ratio' is ambiguous: it is unclear whether this refers to CommonKV alone or to the combined effect of CommonKV plus quantization and eviction. Please clarify the decomposition.","section":"Abstract"},{"comment":"Please standardize benchmark names (e.g., 'Ruler' should be cited consistently with the original benchmark's capitalization) and add references for LongBench and Ruler.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"This is an abstract-only review. The manuscript could not be technically evaluated because the full text was not provided. I recommend that the editor obtain the full manuscript before making a decision; the current abstract is not sufficient to establish soundness, and the 98% compression claim needs careful scrutiny."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"We only have the abstract, so this is a judgment of the framing, not the substance. The core idea—training-free cross-layer KV cache compression by SVD-based parameter sharing across adjacent layers, with an adaptive budget allocated by cosine similarity—is a fresh combination that directly targets a known limitation: existing methods either require retraining or degrade at high compression. That is a real gap, and the authors are honest that this is an alternative to quantization and eviction, with orthogonality claimed as a bonus.\n\nWhat I can credit: the motivation is concrete, the method is training-free, and the experiments apparently span multiple backbones and LongBench/Ruler, which are solid benchmarks for this area. If the full paper delivers on that, it is a useful contribution.\n\nThe soft spots are exactly what you can't see from an abstract. The 98% compression claim, achieved by combining with quantization and eviction, has no error bars, no baseline variance, and no definition of \"significant performance loss.\" That phrase can hide a lot. The stress-test note is also legitimate: cosine similarity between layer caches is not, by itself, a reliable guide to low-rank spectral compatibility. Two high-similarity vectors can live in nearly orthogonal subspaces after centering, so merging them via shared SVD parameters could destroy exactly the energy you wanted to preserve. I can't dismiss that concern from the abstract; it needs to be addressed with actual spectral statistics. If the paper already does this, great—but the abstract doesn't say so.\n\nAlso, the claim of \"consistently outperforming\" existing methods is strong for an abstract with no numerical details. That is normal for an abstract, but it means we cannot assess soundness yet.\n\nNet: this is a plausible, potentially important idea that deserves a serious referee. If the full paper has experiments with standard deviation or significance tests, and if it examines whether cosine similarity tracks compressibility in the relevant spectral sense, it could be a real step forward. For now, I'd send it to review but expect the reviewers to push on the budget-allocation metric and the 98% claim. I wouldn't cite it until I've seen the full text.\n\nRecommendation: accept for peer review, with the understanding that the cosine-similarity issue is the key technical question.","headline":"Plausible training-free KV cache compression via cross-layer SVD sharing, but the 98% claim is unverifiable from the abstract and the cosine-similarity budget rule needs spectral justification.","tokens_in":1314,"tokens_out":1942,"would_cite":false,"duration_ms":22526,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CommonKV compresses LLM KV caches to 98% by sharing SVD parameters across adjacent layers.","keywords":["KV cache compression","cross-layer parameter sharing","singular value decomposition","training-free compression","large language models","cosine similarity","adaptive budget allocation","long-context inference"],"falsifier":"Take a transformer where adjacent-layer hidden-state cosine similarity is known to be low in some regions (e.g., early layers or layers with abrupt specialization), apply CommonKV at 98% total compression, and compare task accuracy against the full-cache baseline; a sharp drop on those regions would show the cosine-similarity assumption is the failure point.","tokens_in":639,"feed_emoji":"🧠","tokens_out":1577,"duration_ms":19591,"temperature":0.7,"pith_summary":"This paper tries to establish that cross-layer KV cache compression can be done without any training by sharing singular-value-decomposition parameters across adjacent transformer layers, and that an adaptive budget based on cosine similarity prevents dissimilar layers from being over-compressed. If true, it would let very long contexts be served with a small fraction of the memory, directly attacking the main scaling bottleneck of large language models. The authors claim their method consistently beats existing low-rank and cross-layer compression techniques across multiple models and benchmarks, and that combining it with quantization and eviction reaches a 98% compression ratio without significant accuracy loss.","feed_headline":"Compress KV cache 98% by sharing parameters across layers","feed_subtitle":"Training-free SVD method splits compression budgets by layer similarity, beating low-rank baselines on LongBench and Ruler.","key_machinery":"The central mechanism is cross-layer SVD parameter sharing: singular value decomposition of adjacent layers' KV parameters creates a shared low-rank latent space that makes per-layer caches mergeable, combined with an adaptive budget allocation that sets each layer pair's compression ratio based on the cosine similarity between their caches. The SVD sharing supplies the compression structure; the cosine similarity decides where compression is safe.","core_discovery":"The central claim is that adjacent layers' hidden states are similar enough that a training-free SVD-based weight-sharing scheme can produce a latent KV cache that is easily mergeable across layers, and that a cosine-similarity-driven budget allocation correctly decides how aggressively each layer pair can be compressed. The paper proposes CommonKV, which applies SVD to adjacent parameters to create shared low-rank bases, then dynamically assigns compression budgets based on the cosine similarity of the caches so that dissimilar layers retain more capacity. Across LongBench and Ruler benchmarks on several backbone models, the method is reported to outperform existing low-rank and cross-layer","pith_inferences":["A direct extension would test whether cosine-similarity budgeting can be replaced or supplemented by a cheap learned predictor of per-layer compressibility, potentially improving robustness when layer similarities vary with input domain.","The SVD sharing principle may extend beyond K and V caches to other weight matrices that exhibit cross-layer similarity, such as MLP projections, though the paper does not claim this.","The method's success depends on the stability of adjacent-layer similarity across sequence lengths; a stress test on very long contexts or off-distribution inputs would reveal if the budget allocation remains safe when representations shift.","One could combine CommonKV with retrieval-based eviction policies that keep the most useful tokens, since the two mechanisms operate on different axes: parameter-level redundancy vs. token-level importance."],"forward_implications":["If CommonKV holds, long-context serving memory can drop by roughly two orders of magnitude when combined with quantization and eviction, making much longer sequences feasible on existing hardware.","The method could be applied post-hoc to already-trained models without fine-tuning, lowering the barrier for adopting compression in production systems.","Adaptive per-layer budgets based on similarity might generalize to other cache compression schemes, suggesting a principled way to choose where to spend bits.","Because CommonKV is orthogonal to quantization and eviction, it can be stacked with existing memory-saving techniques rather than replacing them.","Consistent gains across multiple backbones suggest the cross-layer similarity is a structural property of trained transformers, not an artifact of one architecture."],"supporting_citations":[],"fun_headline_variants":["Compress KV cache 98% by sharing SVD bases across layers","Adaptive SVD sharing shrinks KV cache up to 98% without training","Cross-layer parameter sharing compresses KV cache 98% training-free","Budget-aware SVD sharing compresses KV cache 98% without training"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes that adjacent layers' representations are similar enough that SVD weight sharing and a cosine-similarity budget can over-compress some layers without hurting accuracy; if certain adjacent layers diverge, those layers get over-compressed and the reported performance cannot hold.","fun_headline_variants_meta":{"raw":{"variants":["Compress KV cache 98% by sharing SVD bases across layers","Adaptive SVD sharing shrinks KV cache up to 98% without training","Cross-layer parameter sharing compresses KV cache 98% training-free","Budget-aware SVD sharing compresses KV cache 98% without training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2757,"prompt_tokens":707,"completion_tokens":2050,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":1970}},"tokens_in":451,"tokens_out":2050,"duration_ms":14663,"temperature":1.0,"reasoning_tokens":1970,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:28:34.349161+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a transformer where adjacent-layer hidden-state cosine similarity is known to be low in some regions (e.g., early layers or layers with abrupt specialization), apply CommonKV at 98% total compression, and compare task accuracy against the full-cache baseline; a sharp drop on those regions would show the cosine-similarity assumption is the failure point.","supporting_citations":[],"review_version":1}