Pith. sign in

REVIEW 3 major objections 5 minor 38 references

This paper claims that a learned orthogonal gauge on the KV cache coordinate basis is a practical post-training lever: at matched measured bits-per-value, it cuts zfp-cache KL divergence by 44.0%, logit MSE by 43.3%, top-1 flip rate by 24.5

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 07:29 UTC pith:NF7DQCOM

load-bearing objection A well-run matched-rate evaluation shows learned orthogonal gauges improve KV-cache compression at 2k prefixes; the long-context quality claim is not yet supported, but the paper deserves a serious referee. the 3 major comments →

arxiv 2607.20538 v1 pith:NF7DQCOM submitted 2026-07-10 cs.LG cs.AI

Codec-Gauge: Learning Compression-Friendly Gauges for Transformer KV Caches

Yitao Jiang , Yaoqing Yang , Luyang Zhao , Muhao Chen , Devin Balkcom This is my paper
classification cs.LG cs.AI
keywords KV cache compressionpost-training transformorthogonal gaugeDCT spectral centroidzfpquantizationcache-coordinate geometrylow-bit inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the channels of a Transformer's key/value cache are not just data—they are a coordinate system that a compression backend sees, and that coordinate system can be reshaped after training without touching model weights or backend code. It introduces Codec-Gauge, a one-time post-training step that learns a small orthogonal 'gauge' matrix per layer, head, and channel group, parameterized as exp(A − A^T), trained only on frozen-model KV tensors with a frequency-distribution objective. The objective pushes KV energy toward low frequencies in the codec's token–channel layout via a DCT spectral-centroid loss plus a smooth log-amplitude rate proxy. Evaluated through actual compression and decompression at matched measured bits per value, the learned gauge reduces zfp KL divergence by 44.0%, logit MSE by 43.3%, top-1 flip rate by 24.5%, and KV NRMSE by 18.3% on average across six models, and also improves block-uniform and KIVI-style quantization. If this holds, cache-coordinate geometry is a practical, model-agnostic lever for squeezing more fidelity out of any existing KV compression pipeline.

Core claim

The central claim is that the coordinate basis of each key/value vector—the channel layout a codec sees—is a compression variable in its own right, independent of model weights and backend rules. Concretely, the paper shows that learning a per-checkpoint orthogonal gauge G = exp(A − A^T), trained only on 64 frozen prefill windows with an objective that concentrates DCT energy toward low token-channel frequencies, improves reconstruction fidelity and output-distribution stability for zfp compression and two scalar quantizers at matched measured bits per value. Across six models and the 3, 4, and 6 bits-per-value operating points, the learned gauge reduces zfp KL divergence by 44.0%, logit MSE

What carries the argument

The mechanism is a learned orthogonal 'gauge'—a per-layer, per-head, per-channel-group rotation G = exp(A − A^T) placed between the raw KV tensor and the compression/quantization backend, with the inverse G^T applied after decompression. It is trained from frozen KV tensors by a frequency-distribution objective: a 2D DCT spectral-centroid term L_freq that pulls energy toward low radial frequencies in the codec's token–channel layout, plus a smooth log-amplitude rate proxy L_rate that encourages concentration into few significant coefficients. The gauge reshapes the numerical geometry the backend sees, so a fixed codec (zfp, block-uniform, or KIVI-style) preserves the recovered cache more fai

Load-bearing premise

The load-bearing premise is that the small-sample frequency-distribution objective—trained on 64 prefill windows with a DCT spectral-centroid plus rate proxy, and no reconstruction or task loss—is a transferable surrogate for what a fixed backend actually does to held-out KV streams; if that correlation breaks on other token distributions, longer contexts, or other codecs, the reported gains shrink even though the orthogonality and round-trip correctness claims remain true.

What would settle it

Take a fixed checkpoint and backend, train the gauge exactly as described, then evaluate on a held-out corpus with a very different spectral profile (e.g., code, mathematics, or non-English text) or with a codec that preserves high-frequency coefficients better than zfp, and check whether the learned gauge's KL reduction relative to identity disappears or reverses. A simpler direct check: recompute the aggregate KL reduction after permuting the channel order before applying the gauge; if the gain persists, the effect is not about the codec-facing frequency layout.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • At the same measured bits per value, existing compression and quantization backends become more faithful when KV tensors pass through a learned gauge: zfp KL down 44.0%, block-uniform KL down 44.2%, KIVI-style KL down 27.2%.
  • Cache-coordinate geometry is an orthogonal lever to eviction, paging, and precision policies, so the gauge can be layered on top of them without changing their rules.
  • The gauge is fixed per checkpoint and does not scale with context length, so its parameter cost amortizes as retained context grows; serial runs confirm realized storage matches the nominal bit rate.
  • The learned frequency-shaped geometry beats generic rotations (random, Hadamard, DCT) and a data-derived PCA/KLT basis, indicating the effect is more than incoherence or outlier redistribution.
  • Because gauge training needs no language-modeling or task loss—only 64 prefill windows—the method is inexpensive to apply to a new frozen checkpoint.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Our inference: If the spectral-concentration surrogate transfers to other lossy tensor paths (activations, optimizer states, embeddings) that are consumed in a fixed memory layout, the same gauge idea could improve them without modifying the underlying codec.
  • Our inference: The gains are likely codec-dependent—a backend that preserves high-frequency coefficients better than zfp might see smaller or even reversed effects, so the technique should be re-validated per backend.
  • Our inference: The reported reductions are averages at moderate compression; the paper notes that at 8 bits per value transform-only numerical noise becomes comparable to remaining compression perturbation, suggesting a practical ceiling where gauge gains fade.
  • Our inference: Since gauge training uses only web-text prefill windows, a direct test is whether a gauge trained on one domain (e.g., code or mathematics) transfers to another (e.g., long-document reasoning); the paper does not report cross-domain transfer.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Codec-Gauge, a post-training method that learns small orthogonal channel transforms ('gauges') for Transformer KV caches. The gauge maps each key/value tensor into a codec-facing coordinate basis before fixed-rate compression (zfp) or quantization (block-uniform, KIVI-style), then applies the inverse gauge after decoding. The gauge is parameterized as exp(A - A^T), trained on frozen-model KV tensors with a frequency-distribution objective combining a DCT spectral-centroid loss and a log-amplitude rate proxy. The evaluation uses matched measured bits/value, actual backend compression, and rolling compressed-history scoring across six models at 3/4/6 bits, plus a 27B model check, task-prompt scoring, and serial storage/timing measurements. Headline results report zfp KL divergence reduced by 44.0%, logit MSE by 43.3%, top-1 flip rate by 24.5%, and KV NRMSE by 18.3% relative to identity coordinates, with 18/18 model-rate pairs improved; similar or weaker gains are reported for block-uniform and KIVI-style quantization.

Significance. If the reported results hold, the paper identifies a genuinely useful and previously under-explored variable: the coordinate basis presented to a fixed compression backend. The design has important strengths: the matched-rate protocol with actual compressed bytes rules out hidden rate differences; the no-compression sanity rows verify the G^{-1}G round-trip; the coordinate controls (random, Hadamard, DCT, PCA/KLT) isolate the effect of the learned geometry; and the 27B check plus released code and raw outputs support reproducibility. The 18/18 pair consistency for zfp across the main output metrics is a strong empirical signal. The main qualification is that the headline quality improvements are measured only on 2,049-token prefixes, while the practical motivation is long-context inference; the serial long-context run reports storage and timing but no quality metrics. The transferability of the small-sample frequency proxy to deployment-length contexts is therefore the key unverified assumption.

major comments (3)
  1. [Evaluation Design; Results (Figure 5)] All quality metrics in the main matrix (KL, logit MSE, top-1 flip, KV NRMSE) are measured on 2,048-token prefixes (Appendix Table 1: seq_len=2,432, prefix_len=2,049). The serial context-length run extends to 32k tokens but reports only KV-allocation growth and decode time, not any output-quality or KV-error metric. Since the paper motivates the method for long-context inference and the gauge objective is trained on 4,096-token windows, the headline 44.0% zfp KL reduction has not been shown to persist at deployment-relevant context lengths. Please add quality metrics (at minimum zfp KL, logit MSE, and top-1 flip for learned vs. identity vs. random) at extended context lengths, or explicitly restrict the central claim to the evaluated prefix length.
  2. [Gauge Training Objective; Evaluation Design] The training objective L_train = λ_f L_freq + λ_r L_rate is a proxy that does not involve the codec, reconstruction, logits, or task loss. The 44% average reduction is an empirical correlation between this DCT spectral-centroid/log-amplitude proxy and what zfp actually does to held-out KV streams. The mechanism section reports spectral concentration (0.570 to 0.709) but does not quantify the correlation between the proxy and backend perturbation, nor does it ablate the two loss terms. Because the entire practical claim rests on this transfer, the paper should report (a) the held-out correlation between L_freq/L_rate and zfp output perturbation across models and rates, and (b) an ablation removing L_rate or L_freq, to establish which part of the objective is load-bearing. Without this, the reader cannot judge how much of the gain is due to the frequency objective versus generic smoothing
  3. [Scale, Task, and Storage Checks (Figure 5)] The task-prompt scoring is described as providing 'long-context task formats,' but the context length of these prompts is not reported, and the main evaluation protocol uses 2,049-token prefixes. The LongBench-v2 and RULER prompts may be longer than 2k, but no length-specific breakdown is given. Since the central practical claim is about long-context compression, the paper should state the actual context lengths used in the task-prompt runs and, if they exceed 2k, report quality metrics separately by context-length bucket. If they do not exceed 2k, the claim that the effect is validated in long-context task formats is overstated.
minor comments (5)
  1. [Evaluation Design; Appendix Table 1] The main text says each window has 2,048 prefix tokens, while Appendix Table 1 reports prefix_len=2,049. Please make this consistent.
  2. [Figure 2] The annotations such as '+37%', '+46%' are described as 'reductions' relative to identity; a plus sign is misleading for a reduction. Use '↓37%' or '-37%' with an explicit note that lower is better.
  3. [Appendix Table 4] PCA/KLT improves KV NRMSE on 18/18 pairs but worsens output metrics on average. This is an interesting finding that deserves one sentence in the main text to explain why reconstruction-oriented bases do not align with output fidelity; currently it is only in the appendix.
  4. [Gauge Training Objective] The hyperparameters λ_f=1.0, λ_r=0.02, τ=0.04, T_b=16 are fixed across all checkpoints, which is good, but no sensitivity analysis for λ_r is provided. A small grid would help show that the result is not sharply tuned to this one setting.
  5. [Discussion and Limitations] The 'Discussion and Limitations' section does not explicitly state the context-length limitation of the quality evaluation. Please add a sentence acknowledging that quality metrics are reported only up to 2k prefixes and that longer-context quality is an open question.

Circularity Check

1 steps flagged

No load-bearing circularity; minor internal tautology in the mechanism figure only.

specific steps
  1. self definitional [Section 'Gauge Training Objective' (spectral concentration definition) and Section 'Mechanism' / Figure 4]
    "We track spectral concentration as 1−L_freq. ... The training objective directly targets frequency concentration in the codec-facing token–channel layout. Figure 4 confirms that the objective transfers to held-out evaluation windows: identity and random gauges average approximately 0.570 spectral concentration, while learned gauges reach 0.709."

    The 'mechanism' evidence reports the very statistic the loss minimizes: L_freq is the spectral-centroid objective, and concentration is defined as 1−L_freq. Showing that learned gauges raise concentration on held-out windows verifies that the optimizer moved its own objective, so it is an internal sanity check rather than independent confirmation that frequency concentration causes the quality gain. The headline zfp/quantizer improvements, however, are measured from actual compression/decompression with matched measured bits/value and rolling-cache scoring, and do not reduce to L_freq by construction. This is therefore a minor, non-load-bearing tautology, not a circular derivation of the central claim.

full rationale

The central derivation is not circular. The gauge is trained on a frequency-distribution proxy (L_freq and L_rate) that is explicitly kept outside the codec and output loops, and the paper's headline reductions are observed outcomes from actual zfp, block-uniform, and KIVI-style compression on held-out windows with matched measured rates. The training loss contains no reconstruction, logit, or task loss, so the reported KL/logit/flip/NRMSE improvements cannot be implied by the objective by construction. The evaluation is self-contained: disjoint FineWeb-Edu windows, rolling compressed-history scoring, measured bytes, and sanity rows for the G^{-1}G round trip. The paper does not rely on load-bearing self-citation or an imported uniqueness theorem; it compares against identity, random, Hadamard, DCT, and PCA/KLT controls under the same backend. The only mild concern is that the mechanism figure reports spectral concentration, which is definitionally 1−L_freq and thus confirms the training objective was optimized rather than independently validating the causal mechanism; this is not load-bearing for the main quantitative claims. The long-context quality-transfer question is a correctness and external-validity risk, not circularity, since the quality metrics are all measured at 2k-prefix settings while the serial length run reports only storage and timing. Overall: no significant circularity beyond the minor internal tautology.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The method introduces a small per-checkpoint trainable transform and several hand-set hyperparameters. No new physical entities or unobservable forces are postulated; the gauge's behavior is directly measured.

free parameters (3)
  • Objective hyperparameters λ_f, λ_r, τ, T_b, group size g = λ_f=1.0; λ_r=0.02; τ=0.04; T_b=16; g=16
    Hand-chosen constants in the training loss; the central results depend on them, though the paper reports stability across g=8/16/32/full.
  • Learned skew-symmetric generator matrices A_{l,s,h,r} = not listed (checkpoints released)
    The gauge matrices themselves are fit to 64 FineWeb-Edu KV windows per checkpoint. Their values determine the reported reductions; they are trained, not derived from first principles.
  • Gauge training corpus budget = 64 windows; 25 epochs
    Training set size and duration chosen by hand; if too small or unrepresentative, generalization could fail.
axioms (5)
  • domain assumption DCT spectral concentration in the token-channel layout is a valid proxy for fixed-rate zfp reconstruction fidelity.
    The central objective L_freq assumes concentrating low-frequency energy helps the codec; supported empirically by Figure 4 and zfp results, but not proven analytically.
  • domain assumption A gauge trained on FineWeb-Edu prefill windows transfers to held-out contexts, tasks, and other checkpoints.
    64 windows per checkpoint are used to fit gauges; eval on disjoint docs, RULER/LongBench, and 27B supports but does not guarantee generalization.
  • domain assumption The same gauge geometry improves scalar quantizers even though the objective is not quantization-aware.
    Empirical transfer is observed, but the mechanism (range redistribution/outlier smoothing) is not formalized.
  • standard math Matrix exponential of a skew-symmetric matrix gives an exactly orthogonal gauge with inverse equal to transpose.
    Uses exp(A−Aᵀ) parameterization without an orthogonality penalty.
  • domain assumption Measured compressed bytes from the CUDA zfp bridge equal the bytes a deployment backend would produce.
    Implementation assumption; sanity rows verify zero-error gauge round-trip, but not byte-level equivalence to all production zfp paths.
invented entities (1)
  • Orthogonal cache gauge G_{l,s,h,r} = exp(A_{l,s,h,r}−Aᵀ_{l,s,h,r}) independent evidence
    purpose: Changes KV channel coordinates fed to compression/quantization backends before compression and inverts after decompression.
    A deterministic trained transform, tested through held-out compression, task prompts, a 27B model, and released artifacts; not an unobservable postulate.

pith-pipeline@v1.3.0-alltime-deepseek · 15701 in / 13224 out tokens · 148645 ms · 2026-08-02T07:29:01.292136+00:00 · methodology

0 comments
read the original abstract

Long-context Transformer inference increasingly relies on KV-cache compression or quantization. Prior rotation and transform-coding results suggest that the channel basis of each key/value vector affects how faithfully a fixed backend preserves model behavior. We introduce Codec-Gauge, a post-training cache-coordinate layer that learns small orthogonal channel transforms around existing compression and quantization backends. Its frequency-distribution objective combines a token-channel DCT spectral-centroid loss with a smooth rate proxy to concentrate KV energy in low-frequency codec-facing layouts. We evaluate actual compression and decompression using measured bytes and rolling compressed-history scoring. Across six models at $3$, $4$, and $6$ bits/value, learned gauges reduce zfp KL divergence by $44.0\%$ on average relative to raw coordinates and outperform random, Hadamard, DCT, and PCA/KLT controls. The same gauges improve quality preservation for block-uniform and KIVI-style quantization. Experiments on a 27B model and long-context task prompts reproduce the quality trend, while serial storage and timing measurements validate the implemented compressed-cache paths. These results establish cache-coordinate geometry as a practical post-training variable for improving compression fidelity without changing model weights, attention semantics, or backend coding rules.

Figures

Figures reproduced from arXiv: 2607.20538 by Devin Balkcom, Luyang Zhao, Muhao Chen, Yaoqing Yang, Yitao Jiang.

Figure 1
Figure 1. Figure 1: Overview of Codec-Gauge. A one-time post [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Aggregate output-distribution perturbation under [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coordinate-control checks for zfp compression at [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Gauge training dynamics and spectral cache geometry. Left: mean objective terms across six models with standard-error [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Validation beyond the six-model matrix. Top left: [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

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

38 extracted references · 2 linked inside Pith

  1. [1]

    2019 , eprint =

    Fast Transformer Decoding: One Write-Head is All You Need , author =. 2019 , eprint =

    2019

  2. [2]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =

    Ainslie, Joshua and Lee-Thorp, James and de Jong, Michiel and Zemlyanskiy, Yury and Lebr. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =. 2023 , publisher =. doi:10.18653/v1/2023.emnlp-main.298 , url =

    doi:10.18653/v1/2023.emnlp-main.298 2023

  3. [3]

    and Ermon, Stefano and Rudra, Atri and R

    Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R. Advances in Neural Information Processing Systems , volume =

  4. [4]

    and Zhang, Hao and Stoica, Ion , booktitle =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph E. and Zhang, Hao and Stoica, Ion , booktitle =. Efficient Memory Management for Large Language Model Serving with. 2023 , doi =

    2023

  5. [5]

    2024 , volume =

    Liu, Zirui and Yuan, Jiayi and Jin, Hongye and Zhong, Shaochen and Xu, Zhaozhuo and Braverman, Vladimir and Chen, Beidi and Hu, Xia , booktitle =. 2024 , volume =

    2024

  6. [6]

    and Shao, Yakun Sophia and Keutzer, Kurt and Gholami, Amir , booktitle =

    Hooper, Coleman Richard Charles and Kim, Sehoon and Mohammadzadeh, Hiva and Mahoney, Michael W. and Shao, Yakun Sophia and Keutzer, Kurt and Gholami, Amir , booktitle =. 2024 , url =

    2024

  7. [7]

    2024 , volume =

    Kang, Hao and Zhang, Qingru and Kundu, Souvik and Jeong, Geonhwa and Liu, Zaoxing and Krishna, Tushar and Zhao, Tuo , booktitle =. 2024 , volume =

    2024

  8. [8]

    No Token Left Behind: Reliable

    Yang, June Yong and Kim, Byeongwook and Bae, Jeongin and Kwon, Beomseok and Park, Gunho and Yang, Eunho and Kwon, Se Jung and Lee, Dongsoo , year =. No Token Left Behind: Reliable. 2402.18096 , archivePrefix =

    Pith/arXiv arXiv

  9. [9]

    2025 , eprint =

    Lin, Yujun and Tang, Haotian and Yang, Shang and Zhang, Zhekai and Xiao, Guangxuan and Gan, Chuang and Han, Song , booktitle =. 2025 , eprint =

    2025

  10. [10]

    2023 , url =

    Chee, Jerry and Cai, Yaohui and Kuleshov, Volodymyr and De Sa, Christopher , booktitle =. 2023 , url =

    2023

  11. [11]

    2024 , volume =

    Tseng, Albert and Chee, Jerry and Sun, Qingyao and Kuleshov, Volodymyr and De Sa, Christopher , booktitle =. 2024 , volume =

    2024

  12. [12]

    and Li, Bo and Cameron, Pashmina and Jaggi, Martin and Alistarh, Dan and Hoefler, Torsten and Hensman, James , booktitle =

    Ashkboos, Saleh and Mohtashami, Amirkeivan and Croci, Maximilian L. and Li, Bo and Cameron, Pashmina and Jaggi, Martin and Alistarh, Dan and Hoefler, Torsten and Hensman, James , booktitle =. 2024 , url =

    2024

  13. [13]

    2025 , url =

    Liu, Zechun and Zhao, Changsheng and Fedorov, Igor and Soran, Bilge and Choudhary, Dhruv and Krishnamoorthi, Raghuraman and Chandra, Vikas and Tian, Yuandong and Blankevoort, Tijmen , booktitle =. 2025 , url =

    2025

  14. [14]

    2025 , volume =

    Sun, Yuxuan and Liu, Ruikang and Bai, Haoli and Bao, Han and Zhao, Kang and Li, Yuening and Hu, Jiaxin and Yu, Xianzhi and Hou, Lu and Yuan, Chun and Jiang, Xin and Liu, Wulong and Yao, Jun , booktitle =. 2025 , volume =

    2025

  15. [15]

    2025 , doi =

    Su, Zunhai and Wei, Hanyu and Chen, Zhe and Shen, Wang and Li, Linge and Yu, Huangqi and Yuan, Kehong , booktitle =. 2025 , doi =

    2025

  16. [16]

    doi:10.48550/arXiv.2510.05373 , url =

    Saxena, Utkarsh and Roy, Kaushik , year =. doi:10.48550/arXiv.2510.05373 , url =. 2510.05373 , archivePrefix =

    doi:10.48550/arxiv.2510.05373

  17. [17]

    Advances in Neural Information Processing Systems , volume =

    Zhang, Zhenyu and Sheng, Ying and Zhou, Tianyi and Chen, Tianlong and Zheng, Lianmin and Cai, Ruisi and Song, Zhao and Tian, Yuandong and R. Advances in Neural Information Processing Systems , volume =

  18. [18]

    International Conference on Learning Representations , year =

    Efficient Streaming Language Models with Attention Sinks , author =. International Conference on Learning Representations , year =

  19. [19]

    Model Tells You What to Discard: Adaptive

    Ge, Suyu and Zhang, Yunan and Liu, Liyuan and Zhang, Minjia and Han, Jiawei and Gao, Jianfeng , booktitle =. Model Tells You What to Discard: Adaptive. 2024 , url =

    2024

  20. [20]

    2024 , url =

    Li, Yuhong and Huang, Yingbing and Yang, Bowen and Venkitesh, Bharath and Locatelli, Acyr and Ye, Hanchen and Cai, Tianle and Lewis, Patrick and Chen, Deming , booktitle =. 2024 , url =

    2024

  21. [21]

    2025 , url =

    Cai, Zefan and Zhang, Yichi and Gao, Bofei and Liu, Yuliang and Li, Yucheng and Liu, Tianyu and Lu, Keming and Xiong, Wayne and Dong, Yue and Hu, Junjie and Xiao, Wen , booktitle =. 2025 , url =

    2025

  22. [22]

    2024 , volume =

    Tang, Jiaming and Zhao, Yilong and Zhu, Kan and Xiao, Guangxuan and Kasikci, Baris and Han, Song , booktitle =. 2024 , volume =

    2024

  23. [23]

    2024 , publisher =

    Lee, Wonbeom and Lee, Jungi and Seo, Junghwan and Sim, Jaewoong , booktitle =. 2024 , publisher =

    2024

  24. [24]

    2025 , volume =

    Behnam, Payman and Fu, Yaosheng and Zhao, Ritchie and Tsai, Po-An and Yu, Zhiding and Tumanov, Alexey , booktitle =. 2025 , volume =

    2025

  25. [25]

    2026 , url =

    Kai, Jushi and Wang, Yixuan and Zeng, Boyi and Bai, Haoli and Jiang, Bo and He, Ziwei and Lin, Zhouhan , booktitle =. 2026 , url =

    2026

  26. [26]

    2025 , publisher =

    Li, Runchao and Fu, Yao and Sheng, Mu and Long, Xianxuan and Yu, Haotian and Li, Pan , booktitle =. 2025 , publisher =. doi:10.18653/v1/2025.findings-emnlp.914 , url =

    doi:10.18653/v1/2025.findings-emnlp.914 2025

  27. [27]

    International Conference on Learning Representations , year =

    Staniszewski, Konrad and. International Conference on Learning Representations , year =. 2511.01815 , archivePrefix =

    arXiv

  28. [28]

    2024 , doi =

    Liu, Yuhan and Li, Hanchen and Cheng, Yihua and Ray, Siddhant and Huang, Yuyang and Zhang, Qizheng and Du, Kuntai and Yao, Jiayi and Lu, Shan and Ananthanarayanan, Ganesh and Maire, Michael and Hoffmann, Henry and Holtzman, Ari and Jiang, Junchen , booktitle =. 2024 , doi =

    2024

  29. [29]

    and Wu, Kai-Chiang , booktitle =

    Chang, Chi-Chih and Lin, Wei-Cheng and Lin, Chien-Yu and Chen, Chong-Yan and Hu, Yu-Fang and Wang, Pei-Shuo and Huang, Ning-Chi and Ceze, Luis and Abdelfattah, Mohamed S. and Wu, Kai-Chiang , booktitle =. 2025 , url =

    2025

  30. [30]

    2025 , url =

    Lin, Bokai and Zeng, Zihao and Xiao, Zipeng and Kou, Siqi and Hou, Tianqi and Gao, Xiaofeng and Zhang, Hao and Deng, Zhijie , booktitle =. 2025 , url =

    2025

  31. [31]

    2024 , doi =

    Liu, Akide and Liu, Jing and Pan, Zizheng and He, Yefei and Haffari, Gholamreza and Zhuang, Bohan , booktitle =. 2024 , doi =

    2024

  32. [32]

    Dynamic Memory Compression: Retrofitting

    Nawrot, Piotr and. Dynamic Memory Compression: Retrofitting. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , volume =

    2024

  33. [33]

    Training Transformers for

    Gelberg, Yoav and Eitan, Yam and Bronstein, Michael and Gal, Yarin and Maron, Haggai , year =. Training Transformers for. 2605.05971 , archivePrefix =

    Pith/arXiv arXiv

  34. [34]

    2025 , publisher =

    Bai, Yushi and Tu, Shangqing and Zhang, Jiajie and Peng, Hao and Wang, Xiaozhi and Lv, Xin and Cao, Shulin and Xu, Jiazheng and Hou, Lei and Dong, Yuxiao and Tang, Jie and Li, Juanzi , booktitle =. 2025 , publisher =. doi:10.18653/v1/2025.acl-long.183 , url =

    doi:10.18653/v1/2025.acl-long.183 2025

  35. [35]

    2024 , url =

    Hsieh, Cheng-Ping and Sun, Simeng and Kriman, Samuel and Acharya, Shantanu and Rekesh, Dima and Jia, Fei and Ginsburg, Boris , booktitle =. 2024 , url =

    2024

  36. [36]

    Penedo, Guilherme and Kydl. The. Advances in Neural Information Processing Systems , pages =. 2024 , volume =. doi:10.52202/079017-0970 , url =

    doi:10.52202/079017-0970 2024

  37. [37]

    IEEE Transactions on Visualization and Computer Graphics , volume =

    Fixed-Rate Compressed Floating-Point Arrays , author =. IEEE Transactions on Visualization and Computer Graphics , volume =. 2014 , doi =

    2014

  38. [38]

    IEEE Transactions on Computers , volume =

    Discrete Cosine Transform , author =. IEEE Transactions on Computers , volume =. 1974 , doi =

    1974