The reviewed record of science sign in
Pith

arxiv: 2607.06523 · v1 · pith:3GB4ZE3A · submitted 2026-07-07 · cs.AI

DepthWeave-KV: Token-Adaptive Cross-Layer Residual Factorization for Long-Context KV Cache Compression

Anna Cordoba , Adam Puente Tercero , Nerea Angulo Hijo , Mar Linares Tercero , Julia Barrientos , Ainhoa Miranda , Jesus Olivera This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 02:50 UTCglm-5.2pith:3GB4ZE3Arecord.jsonopen to challenge →

classification cs.AI
keywords compressiondepthweave-kvmemoryacrossresidualtokensattentioncache
0
0 comments X

The pith

8.3x KV Cache Compression Holds Quality in Long-Context LLMs

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

DepthWeave-KV compresses the key-value cache in transformer language models by exploiting two forms of non-uniformity simultaneously. First, adjacent transformer layers share redundant structure, so instead of storing independent key-value tensors per layer, the method maintains shared low-rank channel bases across depth windows and reconstructs layer-specific states through lightweight mixing coefficients plus sparse, token-specific residuals. Second, not all tokens need the same reconstruction fidelity: a token-conditional depth router assigns higher residual rank to instruction-bearing and retrieval-critical tokens while aggressively compressing low-salience continuation tokens. An online attention-output probe mechanism periodically measures reconstruction error during generation and adjusts the router threshold without retraining the base model or requiring calibration data. A fused CUDA kernel ensures that the factorization itself does not introduce memory-traffic overhead at decode time. The paper claims this combination preserves near-full-cache quality (62.9% average task score vs. 63.8% for full cache, 96.1% Needle-in-a-Haystack retrieval accuracy) while reducing KV memory by 8.3x and improving decode throughput to 72.8 tokens/s at 64K context.

Core claim

The central object is cross-depth residual factorization: a decomposition that separates the key-value cache into shared low-rank channel bases across neighboring transformer layers (capturing inter-layer redundancy) and token-specific gated residuals whose rank is dynamically assigned by a depth router based on token salience and online attention-output error. The paper demonstrates that cache compression quality depends on treating fidelity as both depth-dependent and token-dependent, and that an online probe measuring attention-output discrepancy (rather than raw KV-vector error) provides a calibration-free signal sufficient to adapt compression ratios during generation. The ablations are

What carries the argument

Cross-depth residual factorization (shared low-rank bases + gated token-specific residuals), token-conditional depth router (Eq. 2), online attention-output probe error tracking (Eq. 3), fused CUDA kernel for basis lookup / residual dequantization / attention projection

If this is right

  • Long-context LLM serving systems could adopt depth-aware, token-adaptive cache compression as a default, potentially enabling larger batch sizes and longer context windows under fixed GPU memory budgets.
  • The online probe mechanism suggests that compression ratios need not be fixed at prefill but can track generation-phase-dependent sensitivity, which may generalize to other dynamic resource allocation problems in inference.
  • If the attention-output error signal is robust across architectures, it could become a standard compression-quality metric replacing perplexity or raw reconstruction error, which the paper shows can mask retrieval failures.
  • The factorization approach is presented as complementary to quantization and token eviction, suggesting it could be composed with those methods for multiplicative memory savings.

Load-bearing premise

The online attention-output probe periodically materializes a higher-fidelity cache for a small subset of attention heads and assumes that the reconstruction error measured on this subset is representative of the full attention behavior, which may not hold for rare retrieval-critical tokens whose importance only emerges late in generation.

What would settle it

A long-context benchmark containing rare, late-emerging retrieval targets that are not captured by the probe subset of heads would expose retrieval failures invisible to the online error tracker, causing Needle-in-a-Haystack accuracy to drop toward the levels of static-compression baselines.

Figures

Figures reproduced from arXiv: 2607.06523 by Adam Puente Tercero, Ainhoa Miranda, Anna Cordoba, Jesus Olivera, Julia Barrientos, Mar Linares Tercero, Nerea Angulo Hijo.

Figure 1
Figure 1. Figure 1: DepthWeave-KV overview: instead of compressing all layers and tokens uniformly, the method weaves shared depth bases across adjacent layers and spends residual capacity only on tokens whose attention behavior signals retrieval or instruction importance. Xiang et al., 2025; Wang et al., 2024; Akulov et al., 2025; Yu & Chai, 2025a;b). ChunkKV emphasizes semantic preser￾vation at segment granularity (Xiang et… view at source ↗
Figure 2
Figure 2. Figure 2: summarizes the full data path. During prefill, DepthWeave-KV forms depth-window bases, projects keys and values into shared coefficient space, stores quantized residuals according to the router, and records probe statistics. During decoding, the fused kernel loads the shared basis rows, dequantizes only the routed residual channels, recon￾structs the needed key/value fragments, and immediately feeds them i… view at source ↗
read the original abstract

Long-context language model inference is increasingly limited by the memory bandwidth and capacity required to store key-value caches, yet existing compression methods often apply uniform budgets across layers or tokens and degrade retrieval when lexical cues and semantic states require different preservation. We introduce DepthWeave-KV, a token-adaptive cache compression method that factorizes key and value states across neighboring transformer layers using shared low-rank channel bases while retaining lightweight token-specific residuals where attention behavior is sensitive. DepthWeave-KV combines cross-depth residual factorization with a token-conditional depth router that allocates higher reconstruction rank to instruction-bearing and retrieval-critical tokens, and uses calibration-free online error tracking from attention-output probes to adapt compression during generation without retraining the base model. A fused CUDA implementation jointly performs basis lookup, residual dequantization, and attention projection to reduce decode-time memory traffic. Across LongBench, Needle-in-a-Haystack, L-Eval, and long-form QA and summarization benchmarks, DepthWeave-KV achieves near-full-cache task quality with substantially lower memory use, improving average score and retrieval accuracy over prior compressed caches while reaching 8.3x KV memory reduction and 72.8 tokens per second at 64K context.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

4 major / 6 minor

Summary. The paper introduces DepthWeave-KV, a KV cache compression method that factorizes key/value states across neighboring transformer layers using shared low-rank channel bases, augmented by token-specific residuals whose ranks are assigned by a token-conditional depth router. An online attention-output probe mechanism adjusts compression aggressiveness during generation without retraining the base model. A fused CUDA kernel is provided to reduce decode-time memory traffic. The method is evaluated on LongBench, Needle-in-a-Haystack, L-Eval, and several QA/summarization benchmarks against nine compression baselines, reporting an average task score of 62.9% (vs. 63.8% for full cache) at 8.3× memory reduction and 96.1% needle retrieval accuracy.

Significance. The combination of cross-depth residual factorization with token-adaptive routing and calibration-free online error tracking is a reasonable and potentially useful contribution to the KV cache compression literature. The design is well-motivated by the observation that retrieval-critical tokens and depth-varying attention patterns require non-uniform treatment. The fused kernel design and the attention-output-based error metric (Eq. 3) are sensible engineering choices. However, the significance of the results is substantially undermined by reporting issues detailed below, most critically the unspecified base model and internally inconsistent ablation numbers.

major comments (4)
  1. §4, Experimental setup: The base model is never named. The text states 'all methods use the same base model' but does not identify it. The model's layer count, head count, and head dimension directly determine what an 8.3× compression ratio means in absolute terms, whether the depth-window factorization (§3, Eq. 1) is meaningful, and how the results compare to published baselines. Without this information, the central quality and memory claims cannot be independently contextualized or reproduced. This is load-bearing for every result in Tables 1 and 2.
  2. §4, Ablations paragraph vs. Table 2: The ablation text in §4 contains at least three numerical mismatches with Table 2 that are not rounding errors. (1) The text states removing token-conditional routing reduces average score to 61.1% and Needle accuracy to 91.8%, but Table 2 reports 61.2% and 92.4%. (2) The text states disabling online probes reduces Needle accuracy to 89.9%, but Table 2 reports 93.5% — a 3.6-point discrepancy that changes the qualitative conclusion about how important online error tracking is for retrieval. (3) The text states replacing cross-depth factorization with direct layer sharing yields 6.9× memory reduction and reconstruction error of 0.052, but Table 2 reports 5.9× and 0.044. Notably, §5's ablation discussion uses the Table 2 numbers, suggesting the §4 text reflects a stale version of the results. These inconsistencies must be resolved before the paper can be
  3. References: The bibliography contains numerous citations to topics entirely unrelated to KV cache compression or language modeling, including sea turtle anatomy (A et al., 2026a), PLC code generation (MF et al., 2026), infrared imaging (Y et al., 2026), neuron bifurcation analysis (MO et al., 2026), and others. This raises serious concerns about the integrity of the reference list and whether cited related work has been accurately represented. The authors should verify that every cited work is genuinely relevant and that no citations are fabricated or misplaced.
  4. §4, Tables 1–2: No error bars, confidence intervals, or significance tests are reported for any benchmark. Given that the headline margin over the strongest baseline (TailorKV) is 1.5 average-score points and 3.5 needle-accuracy points, some indication of variance across seeds or benchmark subsets is needed to assess whether these gains are robust.
minor comments (6)
  1. §3, Eq. (1): The notation uses Z ∈ {K, V} as a subscript/superscript variable, but the hat on X̂ is introduced without explicit definition of what X̂ represents (presumably the reconstructed key or value). Clarify.
  2. §3, Eq. (2): The router parameter u is described as 'learned' but the training procedure for u and the residual gates is only briefly mentioned ('trained once on unlabeled long-context text using a frozen teacher cache objective'). More detail on this training objective, data, and hyperparameters would aid reproducibility.
  3. §3: The probe interval p, depth window size w, and rank levels (ρ ∈ {0, 2, 4, 8}) are mentioned but their specific values for the main experiments are not stated. A configuration table would help.
  4. §4, Table 1: The decode throughput of 72.8 tok/s for DepthWeave-KV is higher than all baselines including Full KV Cache (42.1 tok/s). While the fused kernel explanation is plausible, the magnitude of the throughput gain (1.73× over full cache) deserves more discussion, including hardware specifications.
  5. Figure 2 is referenced but not visible in the text provided; ensure it clearly shows the data path described.
  6. §7: The future work discussion of probe representativeness for late-emerging retrieval tokens is a genuine limitation. Consider moving a brief statement of this limitation into the main results section rather than only discussing it in future work.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee identifies four issues: (1) the base model is not named in the experimental setup, (2) ablation numbers in the §4 text are inconsistent with Table 2, (3) the bibliography contains citations to unrelated topics, and (4) no error bars or significance tests are reported. We agree that all four are legitimate concerns that must be addressed in revision. Below we respond point by point.

read point-by-point responses

  1. Referee: §4, Experimental setup: The base model is never named. The text states 'all methods use the same base model' but does not identify it. The model's layer count, head count, and head dimension directly determine what an 8.3× compression ratio means in absolute terms, whether the depth-window factorization (§3, Eq. 1) is meaningful, and how the results compare to published baselines. Without this information, the central quality and memory claims cannot be independently contextualized or reproduced. This is load-bearing for every result in Tables 1 and 2.

    Authors: The referee is correct. The base model identity, layer count, head count, head dimension, and hidden dimension are all omitted from the current manuscript, and these details are necessary to interpret the compression ratio, the depth-window size, and the comparability of our results to published baselines. This was an oversight in the experimental setup section. In the revised manuscript, we will explicitly name the base model and report its architecture (number of layers, attention heads, head dimension, hidden dimension, and context window). We will also clarify the depth-window size w used in our factorization and how it relates to the model's layer count, so that the 8.3× compression ratio and the ablation variants in Table 2 can be independently contextualized. We will additionally verify that all baseline numbers in Table 1 were obtained under the same model and configuration, or were taken from published results for the same model, and will state which is the case for each baseline. revision: yes

  2. Referee: §4, Ablations paragraph vs. Table 2: The ablation text in §4 contains at least three numerical mismatches with Table 2 that are not rounding errors. (1) The text states removing token-conditional routing reduces average score to 61.1% and Needle accuracy to 91.8%, but Table 2 reports 61.2% and 92.4%. (2) The text states disabling online probes reduces Needle accuracy to 89.9%, but Table 2 reports 93.5% — a 3.6-point discrepancy that changes the qualitative conclusion about how important online error tracking is for retrieval. (3) The text states replacing cross-depth residual factorization with direct layer sharing yields 6.9× memory reduction and reconstruction error of 0.052, but Table 2 reports 5.9× and 0.044. Notably, §5's ablation discussion uses the Table 2 numbers, suggesting the §4 text reflects a stale version of the results. These inconsistencies must be resolved before the纸can

    Authors: The referee is correct, and we appreciate the careful cross-checking. The §4 ablation paragraph does reflect a stale version of the results that was not updated to match the final Table 2 numbers. Specifically: (1) The correct numbers for the no-token-conditional-router variant are 61.2% average score and 92.4% Needle accuracy (as in Table 2), not 61.1% and 91.8%. The 91.8% figure in the §4 text actually corresponds to the no-residual-gates variant, which was conflated during revision. (2) The correct Needle accuracy for the no-online-probes variant is 93.5% (Table 2), not 89.9%. The 89.9% figure came from an earlier experimental run with a different probe interval; the final run used p=32 decode steps and showed a smaller but still meaningful degradation of 2.6 Needle points. We agree that this changes the qualitative emphasis: online error tracking provides a moderate improvement rather than a dramatic one, and the revised text will reflect this accurately. (3) The correct numbers for the cross-depth-factorization-removed variant are 5.9× memory reduction and 0.044 reconstruction error (Table 2), not 6.9× and 0.052. The §5 discussion already uses the correct Table 2 numbers, confirming that the §4 text is the stale version. In the revised manuscript, we will ensure that all ablation numbers in §4 are consistent with Table 2 and will remove any references to superseded experimental runs. We will also add a note clarifying which ablation variant each number corresponds to. revision: yes

  3. Referee: References: The bibliography contains numerous citations to topics entirely unrelated to KV cache compression or language modeling, including sea turtle anatomy (A et al., 2026a), PLC code generation (MF et al., 2026), infrared imaging (Y et al., 2026), neuron bifurcation analysis (MO et al., 2026), and others. This raises serious concerns about the integrity of the reference list and whether cited related work has been accurately represented. The authors should verify that every cited work is genuinely relevant and that no citations are fabricated or misplaced.

    Authors: The referee is correct. The bibliography contains multiple citations to works entirely unrelated to KV cache compression, language modeling, or any topic discussed in the paper. Examples include citations on sea turtle anatomy, PLC code generation, infrared imaging, neuron bifurcation analysis, and others. These appear to have been introduced through an automated reference management error during manuscript preparation and were not caught before submission. We will conduct a complete audit of the reference list, remove all irrelevant citations, and verify that every remaining citation is genuinely relevant and accurately represented in the text. We will also verify that no cited related work has been mischaracterized. We take this concern seriously and acknowledge that it should have been caught before submission. revision: yes

  4. Referee: §4, Tables 1–2: No error bars, confidence intervals, or significance tests are reported for any benchmark. Given that the headline margin over the strongest baseline (TailorKV) is 1.5 average-score points and 3.5 needle-accuracy points, some indication of variance across seeds or benchmark subsets is needed to assess whether these gains are robust.

    Authors: The referee is correct that no variance estimates are reported, and given the relatively small margins over the strongest baseline, this is a legitimate concern. In the revised manuscript, we will report results averaged over at least three random seeds for the main benchmarks (LongBench average, Needle-in-a-Haystack, and L-Eval) and will include standard deviations or confidence intervals in Tables 1 and 2. For Needle-in-a-Haystack specifically, we will also report variance across needle positions and context lengths, since the 3.5-point margin over TailorKV is the most important retrieval claim and should be supported by variance estimates. We note that for deterministic decoding (greedy, as used in our experiments), seed variance arises from the compression method's stochastic components (basis initialization, router threshold adaptation) rather than from the decoding itself, so the variance may be modest; but this should be demonstrated empirically rather than assumed. revision: yes

Circularity Check

0 steps flagged

No circularity found: the derivation chain is self-contained against external benchmarks and external ground-truth references.

full rationale

The paper's core claims are empirical results measured against external benchmarks (LongBench, Needle-in-a-Haystack, L-Eval, etc.) and external baselines (StreamingLLM, H2O, SnapKV, MiniCache, etc.). The method's three components do not exhibit circular dependencies: (1) Cross-depth residual factorization (Eq. 1) is an architectural design choice — shared low-rank bases plus gated residuals — not a claimed derivation from first principles. (2) The token-conditional router (Eq. 2) is trained on a frozen-teacher reconstruction objective and then evaluated on downstream task quality; the training objective (cache reconstruction) differs from the evaluation metric (task scores), so the results are not forced by the training setup. (3) Online error tracking (Eq. 3) compares the compressed cache against a temporarily materialized full-fidelity cache — an external ground truth, not a quantity defined by the method itself. The error signal feeds back into the router threshold, but this is a feedback control loop, not a circular definition: the reference (full cache) is independent of the compressed representation. No self-citations are load-bearing: the paper cites MiniCache (Liu et al., 2024a), KVSharer (Yang et al., 2024a), and other works by different author groups as motivation and baselines. The reconstruction error reported in Table 1 (0.027) is measured against the full cache via Eq. 3, which is a legitimate measurement against an external reference. The skeptic's concerns about ablation-text/Table-2 numerical mismatches and the unnamed base model are correctness and reproducibility risks, not circularity issues. No step in the derivation chain reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 2 invented entities

The method introduces several hand-tuned parameters (rank levels, window size, probe interval) and learned components (router parameter u) without full specification. The axioms rely on domain assumptions about layer redundancy and attention-output error as a quality proxy, which are plausible but not independently verified.

free parameters (6)
  • u (router parameter)
    Learned router parameter shared across layers in the window, trained on unlabeled long-context text.
  • tau_m (adaptive threshold)
    Window-specific threshold chosen to satisfy current memory budget, adjusted online.
  • rb (basis rank)
    Rank of shared basis channels, not explicitly specified.
  • w (depth window size)
    Size of overlapping depth windows, not explicitly specified.
  • p (probe interval)
    Number of decode steps between probes, not explicitly specified.
  • rho rank levels = 0, 2, 4, 8
    Residual rank levels for different token types, chosen by hand.
axioms (3)
  • domain assumption Adjacent transformer layers contain correlated cache structure that can be captured by shared low-rank bases.
    Section 3: The method partitions layers into overlapping depth windows and assumes redundancy is mostly local in depth.
  • domain assumption Attention-output reconstruction error is a reliable proxy for downstream task quality and retrieval fidelity.
    Section 3: The online error tracking uses attention-output probes to adjust compression, assuming this metric captures behaviorally relevant distortion.
  • domain assumption Token salience can be inferred from accumulated attention mass, attention spikes, and delimiter indicators.
    Section 3: The router computes a score from these four online features to allocate residual rank.
invented entities (2)
  • Shared low-rank channel bases (B_K, B_V) no independent evidence
    purpose: Capture common component among adjacent layers to reduce per-layer storage.
    Introduced by this paper; no external evidence provided for their effectiveness beyond the paper's own benchmarks.
  • Token-conditional depth router no independent evidence
    purpose: Allocate reconstruction rank based on token salience and attention behavior.
    A novel routing mechanism; its effectiveness is only validated within the paper's experiments.

pith-pipeline@v1.1.0-glm · 17068 in / 2221 out tokens · 319371 ms · 2026-07-08T02:50:32.238589+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

146 extracted references · 146 canonical work pages · 13 internal anchors

  1. [1]

    2026 , journal =

    PolyKV: A Shared Asymmetrically-Compressed KV Cache Pool for Multi-Agent LLM Inference , author =. 2026 , journal =

    work page 2026

  2. [2]

    2024 , journal =

    MiniCache: KV Cache Compression in Depth Dimension for Large Language Models , author =. 2024 , journal =

    work page 2024

  3. [3]

    2026 , journal =

    Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction , author =. 2026 , journal =

    work page 2026

  4. [4]

    2026 , journal =

    ReST-KV: Robust KV Cache Eviction with Layer-wise Output Reconstruction and Spatial-Temporal Smoothing , author =. 2026 , journal =

    work page 2026

  5. [5]

    2026 , journal =

    The risk of KV cache compression , author =. 2026 , journal =

    work page 2026

  6. [6]

    2025 , journal =

    Attention Is All You Need for KV Cache in Diffusion LLMs , author =. 2025 , journal =

    work page 2025

  7. [7]

    2025 , journal =

    R-KV: Redundancy-aware KV Cache Compression for Reasoning Models , author =. 2025 , journal =

    work page 2025

  8. [8]

    2024 , journal =

    Lossless KV Cache Compression to 2\ author =. 2024 , journal =

    work page 2024

  9. [9]

    2026 , journal =

    Knowledge Packs: Zero-Token Knowledge Delivery via KV Cache Injection , author =. 2026 , journal =

    work page 2026

  10. [10]

    2025 , journal =

    The Pitfalls of KV Cache Compression , author =. 2025 , journal =

    work page 2025

  11. [11]

    2025 , journal =

    ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference , author =. 2025 , journal =

    work page 2025

  12. [12]

    2024 , journal =

    A Survey on Large Language Model Acceleration based on KV Cache Management , author =. 2024 , journal =

    work page 2024

  13. [13]

    2024 , journal =

    KVSharer: Efficient Inference via Layer-Wise Dissimilar KV Cache Sharing , author =. 2024 , journal =

    work page 2024

  14. [14]

    2025 , journal =

    Towards Efficient Large Language Model Serving: A Survey on System-Aware KV Cache Optimization , author =. 2025 , journal =

    work page 2025

  15. [15]

    2025 , journal =

    TailorKV: A Hybrid Framework for Long-Context Inference via Tailored KV Cache Optimization , author =. 2025 , journal =

    work page 2025

  16. [16]

    2026 , journal =

    Compressing Kv Cache for Long-Context LLM Inference with Inter-Layer Attention Similarity , author =. 2026 , journal =

    work page 2026

  17. [17]

    2025 , journal =

    SpindleKV: A Novel KV Cache Reduction Method Balancing Both Shallow and Deep Layers , author =. 2025 , journal =

    work page 2025

  18. [18]

    2024 , journal =

    Model Tells You Where to Merge: Adaptive KV Cache Merging for LLMs on Long-Context Tasks , author =. 2024 , journal =

    work page 2024

  19. [19]

    2025 , journal =

    Hcattention: Extreme Kv Cache Compression Via Heterogeneous Attention Computing for Llms , author =. 2025 , journal =

    work page 2025

  20. [20]

    2025 , journal =

    EvolKV: Evolutionary KV Cache Compression for LLM Inference , author =. 2025 , journal =

    work page 2025

  21. [21]

    2025 , journal =

    Key, Value, Compress: A Systematic Exploration of KV Cache Compression Techniques , author =. 2025 , journal =

    work page 2025

  22. [22]

    2026 , journal =

    The kv4 Trade-off Is Workload-Dependent: A Depth-and Workload-Resolved Study of 4-bit KV-Cache Quantization on a 4 GB Turing GPU , author =. 2026 , journal =

    work page 2026

  23. [23]

    2025 , journal =

    TableKV: KV Cache Compression for In-Context Table Processing , author =. 2025 , journal =

    work page 2025

  24. [24]

    2026 , journal =

    Q-KVComm: Efficient Multi-Agent Communication via Adaptive KV Cache Compression , author =. 2026 , journal =

    work page 2026

  25. [25]

    2026 , journal =

    HCAttention: Extreme KV cache compression via heterogeneous attention computing for LLMs , author =. 2026 , journal =

    work page 2026

  26. [26]

    2025 , journal =

    Enhancing Large Multimodal Models with Adaptive Sparsity and KV Cache Compression , author =. 2025 , journal =

    work page 2025

  27. [27]

    2024 , journal =

    Eigen Attention: Attention in Low-Rank Space for KV Cache Compression , author =. 2024 , journal =

    work page 2024

  28. [28]

    2025 , journal =

    KVCompose: Efficient Structured KV Cache Compression with Composite Tokens , author =. 2025 , journal =

    work page 2025

  29. [29]

    2024 , journal =

    Residual vector quantization for KV cache compression in large language model , author =. 2024 , journal =

    work page 2024

  30. [30]

    2024 , journal =

    More Tokens, Lower Precision: Towards the Optimal Token-Precision Trade-off in KV Cache Compression , author =. 2024 , journal =

    work page 2024

  31. [31]

    2026 , journal =

    How Much Cache Does Reasoning Need? Depth-Cache Tradeoffs in KV-Compressed Transformers , author =. 2026 , journal =

    work page 2026

  32. [32]

    2026 , journal =

    Stochastic KV Routing: Enabling Adaptive Depth-Wise Cache Sharing , author =. 2026 , journal =

    work page 2026

  33. [33]

    2025 , journal =

    Reimagining Memory Access for LLM Inference: Compression-Aware Memory Controller Design , author =. 2025 , journal =

    work page 2025

  34. [34]

    2026 , journal =

    HACK++: Towards More Effective Head-Aware Key-Value Compression for Efficient Visual Autoregressive Modeling , author =. 2026 , journal =

    work page 2026

  35. [35]

    2026 , journal =

    From Layers to Submodules: Rethinking Granularity in Replacement-Based LLM Compression , author =. 2026 , journal =

    work page 2026

  36. [36]

    2026 , journal =

    StructKV: Preserving the Structural Skeleton for Scalable Long-Context Inference , author =. 2026 , journal =

    work page 2026

  37. [37]

    , author =

    Structural optimization principles for edge AI in motorsport telemetry. , author =. 2026 , journal =

    work page 2026

  38. [38]

    , author =

    Building large-scale English-Romanian literary translation resources with open models. , author =. 2026 , journal =

    work page 2026

  39. [39]

    , author =

    Limit-Cycle Proliferation Under Parametric Delayed Feedback in a Conductance-Based Neuron: Bifurcation Landscape, Orbit Catalog, and Capacity Analysis. , author =. 2026 , journal =

    work page 2026

  40. [40]

    , author =

    Towards sustainable AI knowledge-base assistants in computer science education: on-premise deployment and optimization with open educational resources. , author =. 2026 , journal =

    work page 2026

  41. [41]

    , author =

    Multi-Hardware Benchmarking of Open-Source Large Language Models with Retrieval-Augmented Generation for Mitsubishi FX-Series PLC Instruction List Code Generation. , author =. 2026 , journal =

    work page 2026

  42. [42]

    , author =

    Infrared Imaging for Autonomous Power Inspection: A Review from Detector to System Integration. , author =. 2026 , journal =

    work page 2026

  43. [43]

    , author =

    A survey on LLM-as-a-judge. , author =. 2026 , journal =

    work page 2026

  44. [44]

    , author =

    The Rise of Small Language Models in Healthcare: A Comprehensive Survey. , author =. 2026 , journal =

    work page 2026

  45. [45]

    , author =

    Development of a Rapid Deep-Learning-Assisted Multiphase Multicomponent Numerical Simulation Protocol. , author =. 2026 , journal =

    work page 2026

  46. [46]

    , author =

    F-Transformer: a federated transformer for efficient and privacy-preserving sequence generation. , author =. 2026 , journal =

    work page 2026

  47. [47]

    2021 , journal =

    The Modern Mathematics of Deep Learning , author =. 2021 , journal =

    work page 2021

  48. [48]

    2023 , journal =

    Learn to Accumulate Evidence from All Training Samples: Theory and Practice , author =. 2023 , journal =

    work page 2023

  49. [49]

    2023 , journal =

    Deep Learning and Computational Physics (Lecture Notes) , author =. 2023 , journal =

    work page 2023

  50. [50]

    2026 , journal =

    Monodense Deep Neural Model for Determining Item Price Elasticity , author =. 2026 , journal =

    work page 2026

  51. [51]

    2017 , journal =

    A multitask deep learning model for real-time deployment in embedded systems , author =. 2017 , journal =

    work page 2017

  52. [52]

    2026 , journal =

    Training Transformers for KV Cache Compressibility , author =. 2026 , journal =

    work page 2026

  53. [53]

    2019 , journal =

    Deep learning observables in computational fluid dynamics , author =. 2019 , journal =

    work page 2019

  54. [54]

    2024 , journal =

    Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving , author =. 2024 , journal =

    work page 2024

  55. [55]

    2025 , journal =

    Efficient Many-Shot In-Context Learning with Dynamic Block-Sparse Attention , author =. 2025 , journal =

    work page 2025

  56. [56]

    2025 , journal =

    SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation , author =. 2025 , journal =

    work page 2025

  57. [57]

    2026 , journal =

    Sub-Token Routing for KV Cache Compression , author =. 2026 , journal =

    work page 2026

  58. [58]

    2024 , journal =

    Hymba: A Hybrid-head Architecture for Small Language Models , author =. 2024 , journal =

    work page 2024

  59. [59]

    2025 , journal =

    KV-RAPTOR: Scalable Tree-Structured Retrieval with KV Cache Compression for Question-Answering Systems , author =. 2025 , journal =

    work page 2025

  60. [60]

    1970 , journal =

    Deep Learning-Based Point Cloud Compression: An In-Depth Survey and Benchmark\_supp1-3594355.pdf , author =. 1970 , journal =

    work page 1970

  61. [61]

    2026 , journal =

    KV Pareto: Systems-Level Optimization of KV Cache and Model Compression for Long Context Inference , author =. 2026 , journal =

    work page 2026

  62. [62]

    2025 , journal =

    FAEDKV: Infinite-Window Fourier Transform for Unbiased KV Cache Compression , author =. 2025 , journal =

    work page 2025

  63. [63]

    2026 , journal =

    FastKV: Decoupling of Context Reduction and KV Cache Compression for Prefill-Decoding Acceleration , author =. 2026 , journal =

    work page 2026

  64. [64]

    , author =

    Large-Scale Model-Enhanced Vision-Language Navigation: Recent Advances, Practical Applications, and Future Challenges. , author =. 2026 , journal =

    work page 2026

  65. [65]

    , author =

    SS-AdaMoE: Spatio-Spectral Adaptive Mixture of Experts with Global Structural Priors for Graph Node Classification. , author =. 2026 , journal =

    work page 2026

  66. [66]

    , author =

    AI Agents as Universal Task Solvers. , author =. 2026 , journal =

    work page 2026

  67. [67]

    , author =

    Green prompt engineering for sustainable generative AI. , author =. 2026 , journal =

    work page 2026

  68. [68]

    , author =

    Anatomic Interactive Atlas of the Loggerhead Sea Turtle (Caretta caretta) Coelomic Cavity. , author =. 2026 , journal =

    work page 2026

  69. [69]

    , author =

    On Demand Secure Scalable Video Streaming for Both Human and Machine Applications. , author =. 2026 , journal =

    work page 2026

  70. [70]

    , author =

    Compute-in-memory implementation of state space models for event sequence processing. , author =. 2026 , journal =

    work page 2026

  71. [71]

    , author =

    Kilohertz volumetric imaging of in vivo dynamics using squeezed light field microscopy. , author =. 2025 , journal =

    work page 2025

  72. [72]

    and S, S

    A, A. and S, S. Ai agents as universal task solvers. PMC, 2026. doi:10.3390/e28030332

    work page doi:10.3390/e28030332 2026

  73. [73]

    Anatomic interactive atlas of the loggerhead sea turtle (caretta caretta) coelomic cavity

    A, A., A, M., and J, O. Anatomic interactive atlas of the loggerhead sea turtle (caretta caretta) coelomic cavity. PMC, 2026 a . doi:10.3390/ani16050754

    work page doi:10.3390/ani16050754 2026

  74. [74]

    On demand secure scalable video streaming for both human and machine applications

    A, Z., Y, F., and J, Z. On demand secure scalable video streaming for both human and machine applications. PMC, 2026 b . doi:10.3390/s26041285

    work page doi:10.3390/s26041285 2026

  75. [75]

    D., Salem, T

    Akulov, D., Sana, M., Domenico, A. D., Salem, T. S., Piovesan, N., and Ayed, F. Kvcompose: Efficient structured kv cache compression with composite tokens. Preprint, 2025. doi:10.48550/arXiv.2509.05165

    work page doi:10.48550/arxiv.2509.05165 2025

  76. [76]

    Rest-kv: Robust kv cache eviction with layer-wise output reconstruction and spatial-temporal smoothing

    An, Y., Lu, C., Zhu, K., Yu, T., Zhao, C., Wu, H., Tang, M., and Wang, J. Rest-kv: Robust kv cache eviction with layer-wise output reconstruction and spatial-temporal smoothing. arXiv preprint, 2026

    work page 2026

  77. [77]

    The modern mathematics of deep learning

    Berner, J., Grohs, P., Kutyniok, G., and Petersen, P. The modern mathematics of deep learning. arXiv preprint, 2021. doi:10.1017/9781009025096.002

    work page doi:10.1017/9781009025096.002 2021

  78. [78]

    T., Cohan, A., and Ying, R

    Bui, N., Nguyen, H. T., Cohan, A., and Ying, R. Make each token count: Towards improving long-context performance with kv cache eviction. arXiv preprint, 2026

    work page 2026

  79. [79]

    R-kv: Redundancy-aware kv cache compression for reasoning models

    Cai, Z., Xiao, W., Sun, H., Luo, C., Zhang, Y., Wan, K., Li, Y., Zhou, Y., Chang, L.-W., Gu, J., Dong, Z., Anandkumar, A., Asi, A., and Hu, J. R-kv: Redundancy-aware kv cache compression for reasoning models. arXiv preprint, 2025

    work page 2025

  80. [80]

    V., and Israel, D

    Chen, A., Geh, R., Grover, A., den Broeck, G. V., and Israel, D. The pitfalls of kv cache compression. arXiv preprint, 2025

    work page 2025

Showing first 80 references.