{"id":"0006db49-51a0-4c1f-a47e-53d9cc966513","arxiv_id":"2507.20030","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FAEDKV compresses the KV cache by keeping the most important Fourier-frequency components of cached keys and values, reporting gains over eviction baselines on LongBench and position-flat retrieval on needle-in-a-haystack tests.","lead":"Large language models hold past context in a key-value cache that grows with text length. This paper proposes compressing that cache into Fourier-frequency coefficients and reports accuracy gains on long-context benchmarks, with fairness and theory caveats detailed below.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IWDFT recursion in Eq. 9 is not a Fourier transform of the growing history; the 'unbiased' claim is unsupported by the paper's own equations.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing issue: Eq. 9 is asserted, not derived, and the (N-1)/N -> 1 approximation changes per-token weights from equal to roughly 1/(arrival index) (or, more precisely, to a fixed-magnitude phase-rotated sum dominated by recent tokens). My analysis confirms this is the central correctness risk. The headline claim of 'unbiased information retention' is not supported by the paper's own equations; the method may still work empirically, but the theoretical justification is invalid as written. I do not see a stronger concern that would change the reader's CONDITIONAL verdict: the LongBench tables, if taken at face value, are internally consistent enough to warrant a conditional acceptance with requested revisions rather than outright rejection. The proposed test is the minimal check that would settle the issue: unroll Eq. 9 and compare per-token sensitivities to a true DFT. I also note secondary issues that do not change the central verdict: Table 1 labels accuracy as perplexity; the memory accounting omits complex coefficients and the S+R time-domain tokens; and the closest prior work FreqKV is cited but never benchmarked. These reinforce CONDITIONAL rather than ACCEPT, but the Eq. 9 gap is the load-bearing concern.","tokens_in":11806,"tokens_out":2175,"duration_ms":27642,"concrete_test":"Take a synthetic sequence x[1..T] (e.g., T=10,000, random Gaussian entries) and compute the IWDFT state S_T from Eq. 9 using the paper's stated approximation (N-1)/N -> 1 and also with the exact factor. Compare S_T to the direct DFT coefficient X[T][k] = sum_{n=1}^T x[n] e^{-j2πkn/T} for matching k, and, more directly, compute the sensitivity dS_T[k]/dx[n] for each n. If |dS_T/dx[n]| is not constant in n (and, in the approximate version, grows with T - n while |S_T| grows without bound), then Eq. 9 is not an unbiased Fourier representation. Repeating at T=100,000 will also show whether the claimed equal weighting and boundedness hold; they should not under either version of the recursion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IWDFT preserves all tokens with equal weight, making FAEDKV 'unbiased'. This rests on Eq. 9: S_{t+1}[k] = W_k ((N-1)/N S_t[k] + (1/N) x[t+1]). For this recursion to represent the DFT of the full KV-history prefix, W_k would have to be e^{-j2πk/N} with N changing at every step, and the new DFT bin frequencies k/N would change with N. The paper never defines W_k in Eq. 9 in terms of the current N, and a fixed twiddle factor cannot make a recursive accumulation equal the DFT of a sequence whose length grows. Moreover, the paper explicitly replaces (N-1)/N with 1 'since most context is larger than 1000'. With that replacement, the unrolled state is approximately S_t[k] = sum_{n=1}^{t} e^{-j2πk(t-n)/M} x[n] (for a fixed M), so the magnitude contributed by each past token is 1, not 1/t. Early tokens do not receive equal weight; recent tokens dominate the phase-rotated sum, and the state magnitude grows linearly with t unless some other normalization is applied. Even keeping the exact (N-1)/N factor, the recursion is a leaky integrator with a length-dependent normalization, not an equal-weight Fourier representation: the per-token weight is roughly (n/N) or an exponential-like decay, not constant. Thus the abstract's 'equalized contribution of all tokens' and the Section 5.3 attribution of position-agnostic retrieval to 'the DFT, which inherently processes all token information with equal weight' are contradicted by the method's own update rule. The benchmark numbers, if reproducible, would show an effective method, but the paper's core theoretical justification for 'unbiased information retention' does not follow from Eq. 9. This is a load-bearing gap because the method's novelty and its claimed advantage over eviction and learned compression rest on this property.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FAEDKV, a training-free KV-cache compression method that stores a pruned frequency-domain representation of the middle segment of the context, keeps small sink and recent token windows, and updates the frequency representation with a recursive \"Infinite-Window Fourier Transform\" (IWDFT). A one-time layer-wise frequency ablation on WikiText-103 selects which frequency chunks are retained, and the representation is pruned and reconstructed via sparse IDFT during decoding. Experiments on LongBench with Llama3-8B report gains over H2O and SnapKV at small cache sizes, and a Needle-in-a-Haystack study with Qwen2.5-7B reports more position-consistent retrieval than LoCoCo. The central stated contribution is that the method is unbiased because the IWDFT gives equal weight to all tokens.","tokens_in":12065,"tokens_out":7167,"duration_ms":84255,"significance":"If the method genuinely provided an equal-weight, fixed-memory spectral summary of the entire KV history, it would be a valuable training-free alternative to eviction-based and learned compression. The paper has several strengths: the ablation hyperparameters are tuned on WikiText-103/PG-19 while the headline LongBench and NIAH benchmarks are held out; the authors compare with official baseline code; and the Limitations section is candid about single-GPU resource constraints. However, the central theoretical claim is not supported by the paper's own equations: the IWDFT recursion in Eq. (9) is not a Fourier transform of the growing sequence, and the approximation of (N-1)/N to 1 changes the per-token weighting. The empirical numbers may still indicate a useful heuristic, but the paper's novelty and its 'unbiased' characterization rest on an incorrect mathematical premise.","major_comments":[{"comment":"The IWDFT recursion is not a DFT of the growing KV history. In the DFT definition of Eq. (4), the twiddle factor W_k = e^{-j2πk/N} depends on the current sequence length N, so the bin frequencies k/N change as N grows; a fixed W_k in Eq. (9) cannot reproduce the DFT coefficients of the full sequence. Moreover, the paper immediately approximates (N-1)/N to 1, which changes the per-token weight from 1/N (equal magnitude) to roughly 1 per token, so early tokens do not contribute with equal weight and the state magnitude grows linearly with t. This contradicts the claim in §4.1 that the normalization prevents overflow and the Abstract's 'equalized contribution of all tokens'. The authors must either use the exact normalization and provide a rigorous statement of what equal weighting means for the phase-rotated accumulation, or remove the unbiased/equal-weight claims and rename the update accordingly.","section":"Section 4.1, Eq. (9)"},{"comment":"The caption of Table 1 describes the scores as 'perplexity', but LongBench tasks are scored with accuracy/F1 metrics and the reported values (e.g., FullKV 22.53 on NtrvQA) are not perplexity numbers. Since Table 1 is the main quantitative evidence for the paper's performance claims, the metric must be corrected and all prose statements should be checked against the correct metric.","section":"Table 1, Section 5.2"},{"comment":"The Needle-in-a-Haystack setup is described inconsistently: the text says the two methods were evaluated on 1024 cache size with r=0.05 of 24K, while the Figure 3 caption says LoCoCo uses cache size 2048 and FAEDKV uses a 5% compression rate; the text says contexts are 8K-30K tokens, while the caption says the x-axis runs from 8K to 300K tokens. This experiment is used to support the position-agnostic retrieval claim, so the actual configuration must be stated precisely and consistently.","section":"Section 5.3 and Figure 3"}],"minor_comments":[{"comment":"In the assembly of V_t, the second block should be the reconstructed value cache eVt, not eKt as written.","section":"Eq. (13)"},{"comment":"The model is named 'Qwen2-7B-Instruct' in Section 5.1 and 'Qwen2.5-7B-Instruct' in Section 5.3; please use one consistent name.","section":"Section 5.1 vs 5.3"},{"comment":"The heatmap's x-axis is labeled 1-20, while the text states C=22 chunks; please align the figure labels with the stated hyperparameter.","section":"Figure 6"},{"comment":"There are several typos and style inconsistencies, including 'mathmatical', 'convulution', and the typography of (N-1)/N; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The theoretical flaw in Eq. (9) is foundational: the method is presented as an unbiased Fourier transform, but the update is neither an exact DFT of the growing history nor an equal-weight accumulation after the paper's own approximation. If the authors cannot supply a correct derivation or a precise redefinition of 'unbiased', I would not recommend acceptance even after revision. The empirical results may still be salvageable as a heuristic, but the paper would need a substantial rewrite of its claims, not just local edits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me save you some time: this paper has a plausible engineering recipe and a very shaky theoretical justification. The empirical numbers hold together, but the 'unbiased' claim does not survive its own equations, and the memory comparison understates the true cost. I don't think it's a desk reject, but it needs major revisions before I'd trust the claims.\n\nWhat's actually new is the combination of a recursive, fixed-memory frequency-domain update (Eq. 9) with layer-wise spectral chunk selection based on a perplexity ablation. That combination is absent from the cited prior work. The ablation study is a sensible way to pick which frequency bands to keep, and the LongBench numbers are internally consistent with the prose. The Needle-in-a-Haystack result showing position-agnostic retrieval is the most interesting empirical finding.\n\nThe soft spots are significant. First, Eq. 9 is not a DFT of a growing sequence. The twiddle factor W_k is never defined with respect to the current length N, and the approximation (N-1)/N ≈ 1 turns the update into an unnormalized sum, not an equal-weight running mean. The abstract's 'equalized contribution of all tokens' and the Section 5.3 claim that the DFT 'inherently processes all token information with equal weight' are therefore unsupported by the paper's own math. You could rescue this by presenting IWDFT as a deliberately chosen recursive spectral filter, not as a Fourier transform, and by quantifying its actual weighting behavior. Second, the budget comparison is unfair as stated: the text says baselines get 512, 1024, and 2048 cache entries, but the first block of Table 1 shows 768 for both baselines and FAEDKV. If FAEDKV receives 768 while baselines receive 512, the headline 'up to 22%' improvement is confounded. Third, the nominal '9% cache size' ignores that each complex coefficient stores two floats, plus the S+R time-domain tokens, so real memory is closer to double the nominal value. Fourth, FreqKV is the closest prior method and is cited but never benchmarked. That is a serious omission. Finally, there is no code, no error bars, and Table 1 mislabels accuracy as perplexity.\n\nNone of these is necessarily fatal. The core idea is worth pursuing, and the fixes are concrete: re-derive or reframe IWDFT, run equal-budget comparisons, provide byte-level memory accounting, add FreqKV as a baseline, and release code. As it stands, I would not cite these numbers, but I would still send the paper to a referee. The reading group might enjoy pulling apart the theory.","headline":"Plausible frequency-domain KV compression recipe, but the 'unbiased' claim is contradicted by its own equations and the memory comparison is misleading.","tokens_in":12823,"tokens_out":6278,"would_cite":false,"duration_ms":64059,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FAEDKV claims a training-free Fourier transform of the KV cache compresses long-context memory to 9% while preserving all-token information.","keywords":["KV cache compression","Fourier transform","infinite window","frequency domain","long-context","training-free","attention sinks","LLM inference"],"falsifier":"Construct a deterministic sequence of length N, compute its exact DFT, and run the IWDFT update over the same sequence. If the resulting state differs from the DFT coefficients—which the (N-1)/N-to-1 substitution predicts—the equal-weight property is false. A simpler check is to measure the state magnitude after N steps: an exact normalized DFT stays bounded, while the paper's approximation grows logarithmically.","tokens_in":11450,"feed_emoji":"🧠","tokens_out":4057,"duration_ms":40061,"temperature":0.7,"pith_summary":"The paper proposes FAEDKV, a KV cache compression method that keeps a running frequency-domain summary of the entire token history instead of evicting or repeatedly recompressing old tokens. The central claim is that an 'Infinite-Window Fourier Transform' can update this summary in O(M) per step with all tokens contributing equally, so early and recent context are both preserved. On LongBench with a 9% cache, FAEDKV beats eviction baselines by up to 22% in accuracy, and it shows position-agnostic retrieval in needle-in-a-haystack tests. If the claims hold, long-context inference can keep most of the original attention quality at a fraction of the memory cost and without any fine-tuning.","feed_headline":"Fourier-based KV cache keeps long-context QA at 9% memory","feed_subtitle":"FAEDKV claims unbiased token retention via an infinite-window Fourier transform, beating eviction by up to 22%.","key_machinery":"The Infinite-Window Fourier Transform (IWDFT) is a recursive update to a frequency-domain KV cache: each new token is rotated by a fixed twiddle factor and blended into the stored spectral state, as in Equation 9. It is what lets the cache grow without storing time-domain history; the paper also uses a frequency ablation to pick, per layer, which spectral chunks survive compression, and a sparse IDFT to reconstruct only those components. The key claim is that this preserves all tokens with equal weight, in contrast to eviction's recency bias and learned compression's repeated degradation of old tokens.","core_discovery":"FAEDKV's central discovery is that the KV cache can be represented in the frequency domain and updated recursively as new tokens arrive, so that a fixed-size spectral state summarizes the unbounded history. The paper introduces the Infinite-Window Fourier Transform (IWDFT), given by S_{t+1}[k] = W_k ((N-1)/N S_t[k] + (1/N) x[t+1]), to fold each new token into the state while avoiding the overflow of a naive recursive DFT. A layer-wise frequency ablation on WikiText identifies which spectral chunks are critical per layer, and only those are retained at the chosen compression ratio. FAEDKV keeps the first S and last R tokens as time-domain anchors, reconstructs the compressed middle segment with a sparse IDFT, and combines them for attention. The paper reports average LongBench gains of 2.91 points over H2O and 2.12 over SnapKV, with the largest gains at the tightest cache budgets.","pith_inferences":["If the equal-weight property is what drives the gains, then restoring the exact (N-1)/N factor instead of approximating it to 1 could change the story at very long contexts; the paper does not report that ablation.","The frequency-ablation principle suggests a testable extension where the selected spectral bands adapt per input rather than being fixed per layer, which could combine the robustness of frequency retention with per-query flexibility.","Since FAEDKV explicitly does not extend the model's context limit, its real-world value is in compressing existing long-context deployments; pairing it with an eviction-based recent-token cache could yield further memory savings without changing the frequency-side claim."],"forward_implications":["At a 9% cache the method reports average LongBench accuracy gains of about 2.9 points over H2O and 2.1 over SnapKV, and up to 22% at the tightest budgets.","Needle-in-a-haystack accuracy stays roughly flat across needle positions, unlike compression methods that favor recent tokens, supporting the position-agnostic claim.","FAEDKV requires no fine-tuning and works within the model's existing context length, so it can be dropped into a deployed LLM with a one-time per-layer ablation.","Memory drops from O(N) to O(N·r) for the transformed segment, and the sparse IDFT keeps reconstruction cost proportional to the retained components."],"supporting_citations":[{"why":"Supplies the LongBench benchmark and the full-KV baseline used for all accuracy comparisons.","marker":"(Bai et al., 2024b)"},{"why":"Provides the H2O eviction baseline that FAEDKV compares against and claims to beat by up to 22%.","marker":"(Zhang et al., 2023b)"},{"why":"Provides the SnapKV baseline, the state-of-the-art long-context compression method that FAEDKV claims to outperform on LongBench.","marker":"(Li et al., 2024c)"},{"why":"Provides the LoCoCo learned-compression baseline used in the needle-in-a-haystack comparison.","marker":"(Cai et al., 2024a)"},{"why":"Supplies WikiText-103, the dataset used for the layer-wise frequency ablation study that determines which spectral chunks are retained.","marker":"(Merity et al., 2016)"},{"why":"Supplies the PG19 dataset used for the chunk-size ablation and latency evaluation.","marker":"(Rae et al., 2019)"},{"why":"Motivates the retention of the first S tokens as attention sinks in the pre-filling stage.","marker":"(Han et al., 2024)"}],"fun_headline_variants":["Training-free KV cache in frequency domain cuts memory","Infinite-window Fourier transform for unbiased KV cache","FAEDKV: Fourier-based KV compression without retraining","Frequency-domain KV cache improves LongBench up to 22%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire 'unbiased' claim rests on the IWDFT recursion being an exact Fourier transform of the full token history with equal per-token weights; the paper asserts this and then replaces (N-1)/N with 1, which changes the effective weights so early tokens dominate.","fun_headline_variants_meta":{"raw":{"variants":["Training-free KV cache in frequency domain cuts memory","Infinite-window Fourier transform for unbiased KV cache","FAEDKV: Fourier-based KV compression without retraining","Frequency-domain KV cache improves LongBench up to 22%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000906,"raw_usage":{"total_tokens":3901,"prompt_tokens":955,"completion_tokens":2946,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2881}},"tokens_in":571,"tokens_out":2946,"duration_ms":22660,"temperature":1.0,"reasoning_tokens":2881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:55:14.832622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a deterministic sequence of length N, compute its exact DFT, and run the IWDFT update over the same sequence. If the resulting state differs from the DFT coefficients—which the (N-1)/N-to-1 substitution predicts—the equal-weight property is false. A simpler check is to measure the state magnitude after N steps: an exact normalized DFT stays bounded, while the paper's approximation grows logarithmically.","supporting_citations":[],"review_version":1}