Pith. sign in

REVIEW 4 major objections 6 minor 84 references

Prefill-stage KV selection fails in multimodal inference because decoding queries have much larger variance than prefill queries; MM-ShiftKV samples variance-expanded query proxies during prefilling and improves tight-budget accuracy by ret

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 11:52 UTC pith:35IMJO3M

load-bearing objection Solid, well-specified systems paper that likely works as claimed; the main risk is hyperparameter calibration on the evaluation benchmarks, so 'consistently outperforms' needs a held-out check. the 4 major comments →

arxiv 2607.22586 v1 pith:35IMJO3M submitted 2026-06-09 cs.AI cs.CL

MM-ShiftKV: Decode-Aware Prefill-Stage KV Selection for Multimodal Large Language Models

Jinsong Shu , Chenyang Wu , Zhongle Xie , Baokun Wang , Lidan Shou This is my paper
classification cs.AI cs.CL
keywords KV cache compressionmultimodal LLM inferenceprefill-stage selectionquery proxy samplingattention mass estimationvariance expansiontraining-free inferencedecode-aware KV selection
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.

This paper argues that existing prefill-stage KV cache selection methods rest on an assumption that breaks down in multimodal models: that the queries seen during prefilling are representative of those seen during decoding. In practice, decoding queries show much larger variance, so prefill-only rankings under-select semantically critical visual tokens under tight memory budgets. The paper proposes MM-ShiftKV, a training-free, strictly prefill-only selector that samples synthetic query proxies with inflated variance, aggregates attention mass across groups of proxies, and votes to retain the prompt keys most likely to be attended during decoding. On OCR, grounding, VQA, and captioning benchmarks, it reports consistent gains over prior prefill-only methods at the same cache budgets, with the largest gains at the most constrained budgets. If correct, this makes prefill-only KV compression substantially safer for multimodal workloads without introducing any decoding-time intervention.

Core claim

MM-ShiftKV's central claim is that the implicit assumption behind prefill-stage KV selection — that prefilling attention statistics represent decoding behavior — breaks down in multimodal inference. The paper measures a persistent scale mismatch on Qwen2.5-VL and LLaVA-v1.6: decoding hidden states have consistently larger variance than prefill hidden states, while means shift less. Because importance scores built from prefill queries are therefore under-scaled, small ranking errors can drop semantically critical visual tokens. MM-ShiftKV corrects this by sampling synthetic query proxies from a per-sample Gaussian whose standard deviation is inflated by a factor γ=10, computing proxy attentio

What carries the argument

The key mechanism is the variance-expanded query proxy: hidden states are summarized by per-sample, per-dimension mean and standard deviation; proxy hidden states are drawn from a Gaussian with the standard deviation multiplied by γ=10, projected through the model's query matrix, and assigned synthetic future positions via rotary positional embedding. These proxies are used to compute softmax attention over all prompt keys (proxy-induced attention mass), aggregated in groups; each group selects the smallest set of keys covering a τ=0.95 fraction of the mass, and selected keys receive votes. A last-query attention anchor acts as a tie-breaker, and the most recent token is always kept. This co

Load-bearing premise

The load-bearing premise is that decoding-time queries can be approximated by sampling from a per-sample Gaussian with the prefill mean and a variance inflated by a single global factor γ=10, without conditioning on the generated text or on layer/head structure; if that proxy distribution is wrong, the ranking it produces can drop the wrong keys.

What would settle it

Measure actual query vectors during decoding on OCRBench and compare their per-layer, per-head variance with prefill-stage queries: if the prefill-to-decode standard-deviation ratio is far from 10 or varies strongly across heads, layers, or inputs, the fixed variance expansion is misspecified. Also run KV selection at a 64-token budget using true decoding queries as an oracle; if the oracle's attention-mass coverage substantially exceeds MM-ShiftKV's, then proxy construction — not the selection objective — is the limiting factor.

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

If this is right

  • Under tight per-head budgets (64 tokens), MM-ShiftKV reports relative gains of roughly 20–30% over strong prefill-only baselines on document-understanding benchmarks, and larger gains over heuristic methods.
  • Because selection happens once at the end of prefilling, the compressed cache is reused unchanged during decoding, so the method adds no decoding-time latency and remains compatible with FlashAttention-style kernels.
  • The method is training-free and uses only prefill statistics, so it can be dropped into existing multimodal LLM inference pipelines without weight updates.
  • MM-ShiftKV is complementary to per-layer and per-head budget allocation strategies: applying its selection on top of such strategies improves results at the same budget.
  • Correcting the scale mismatch improves KV selection independently of how budgets are allocated, indicating that query-distribution calibration is a general design principle for prefill-only compression.

Where Pith is reading between the lines

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

  • If the variance mismatch is a general property of autoregressive decoding rather than a multimodal artifact, the same proxy-inflation idea could improve prefill-only KV selection in text-only long-context models; the paper does not test this.
  • The fixed global γ=10 leaves head- and layer-level variation on the table; a per-layer or per-head expansion factor calibrated on a small held-out set, or estimated from true query statistics, is a plausible refinement.
  • A direct oracle comparison — selecting keys with actual decoding queries at the same budget — would quantify how much of the remaining accuracy gap is proxy specification error versus the intrinsic difficulty of one-shot selection.
  • Because proxies are sampled once per input without conditioning on generated text, a draft-conditioned variant using a few sampled continuations could capture query drift that depends on what the model is about to say.

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

4 major / 6 minor

Summary. MM-ShiftKV proposes a training-free, strictly prefill-only KV-cache selection method for multimodal LLMs. The paper first documents a prefill-decode hidden-state variance mismatch (Section 2, Fig. 1a), then constructs N=512 query proxies by sampling from N(mu_pre, diag((gamma*sigma_pre)^2)) with gamma=10, projects them through W_Q, applies future-position RoPE, computes proxy attention against prompt keys (Eqs. 5-7), and ranks keys by group-wise votes over minimal token sets covering tau=0.95 of aggregated attention mass, with a last-query anchor as tie-breaker. Experiments on Qwen2.5-VL-7B and LLaVA-v1.6-7B across six benchmarks and budgets C=64-512 report consistent gains over SnapKV, ExpectedAttn, KEYDIFF, StreamingLLM, plus integration with PyramidKV/AdaKV/SparseMM and latency/memory results. The code is provided. Section 8 candidly acknowledges prefill-stage overhead and long-context limitations.

Significance. If the reported gains are genuine, MM-ShiftKV is a useful contribution: it is training-free, prefill-only, compatible with FlashAttention-style decoding, and complementary to budget-allocation methods. The closed-form expected-attention calculation in Appendix A.2 is correct, and the ablations in Section 4.2 and Appendix D are informative and clearly specified. The main risks are that the central proxy assumption is validated in hidden-state space rather than query space, and that the default hyperparameters are selected on the same benchmarks and budget used for the headline claims. Both are fixable with additional experiments and reporting; the paper should not be rejected on the current evidence, but the central claim needs strengthening.

major comments (4)
  1. [Section 2, Fig. 1a; Appendix C/E.6; Eqs. (5)-(7)] The motivating prefill-decode scale mismatch is measured on hidden states, but MM-ShiftKV's ranking is computed after W_Q and RoPE in query space. A hidden-state std ratio does not directly imply the corresponding query covariance inflation, because W_Q can reweight or rotate dimensions and RoPE mixes neighboring positions. Please report, on a held-out set, the empirical norm/covariance of true decode-time query vectors versus prefill query vectors, and validate the Eq. (5) proxies by comparing their induced key rankings with actual decode attention (top-K overlap or attention-mass coverage). As written, the central proxy assumption is plausible but unvalidated at the level where selection decisions are made.
  2. [Appendix D, Tables 6-8; Tables 1-2] The default hyperparameters gamma=10, tau=0.95, and N=512 are selected via sensitivity runs on Qwen2.5-VL-7B at C=64 on OCRBench, TextVQA, and TextCaps -- the same tasks and the most restrictive budget used in the headline comparisons. The 'consistent' gains in Tables 1-2 may therefore partly reflect in-sample tuning. Please add a calibration split (choose defaults on a few held-out tasks/budgets and evaluate on the rest), or report the full sensitivity grid and show that the main conclusions are robust outside the selected points. Also, the paper says all results are averaged over 3 seeds, but Tables 1-2 report no per-seed values or standard deviations; please add them or explain why the spread is negligible.
  3. [Appendix A.2, Eq. (14); Eq. (7)] Eq. (14) justifies variance expansion via E[exp(q^T k / sqrt(d))] for a Gaussian query, but the actual selection score in Eq. (7) is softmax-normalized over all prompt keys. Inflating Sigma_q increases the numerator for every key, but it also increases the denominator, and the net effect can favor keys with large norm rather than keys with high true decoding attention. Since the method is described as decode-aware and the central claim is improved attention-mass coverage, the paper should either derive a softmax analogue of Eq. (14) or provide an empirical check that gamma=10 increases coverage against actual decode queries. As written, the theoretical motivation and the implemented ranking criterion are not the same quantity.
  4. [Appendix E.2, Table 10; Eqs. (5), (15)] Appendix E.2 states that measured prefill-to-decode variance expansion is 'typically 13x to 22x' and motivates a globally fixed gamma as a robust upper bound, yet the default gamma=10 is below that range. A single scalar gamma must be justified in terms of the query-space quantity it is meant to approximate; the layer-wise comparison in Table 10 is only on one benchmark and does not show per-layer/head variance estimates. Please report per-layer or per-head expansion factors derived from query vectors or hidden states, and clarify whether gamma=10 is a calibrated estimate or a tuned hyperparameter.
minor comments (6)
  1. [Appendix F.1; Table 9] The baseline name is written as 'ExceptAttn' in one place and 'ExcepAttn' in Table 9; it should be consistently 'ExpectedAttn'.
  2. [Figure 6] The caption does not define 'mean shift' and 'std ratio'; please add explicit definitions and state how they are computed layer-wise.
  3. [Section 3.2.1] Please specify the exact synthetic future position indices used for RoPE. Figure 2 suggests T+1,...,T+g, but the text only says 'future positions'.
  4. [Eq. (5)] Please clarify whether sigma_pre is the sample standard deviation across prompt tokens for each feature dimension or across dimensions; the notation 'diag((gamma sigma_pre)^2)' implies the former, but it should be stated explicitly.
  5. [Table 3] The abstract and Section 4.1 use 'consistently outperforms', but Table 3 contains several 0.0 gains and one negative value (TextCaps, budget 256, SparseMM). Please qualify the claim or quantify where gains are consistent.
  6. [Table 4] MMMU is used in Tables 1-2 but its average token count is omitted from Table 4; please add it or explain the omission.

Circularity Check

1 steps flagged

Hyperparameters γ and τ are selected on the exact benchmarks and budget used for the headline comparisons, so part of the reported 'consistent outperformance' is a fitted result.

specific steps
  1. fitted input called prediction [Appendix D (Sensitivity Study, D.1) and Section 4.1 (Implementation and Hyperparameters); Tables 1-2]
    "Unless otherwise specified, all experiments in this section are performed on Qwen2.5-VL-7B-Instruct under a fixed per-head KV cache budget of C=64. We report results on three representative multimodal benchmarks: OCRBench and TextVQA (accuracy), and TextCaps (CIDEr). ... Performance peaks around γ=10, while further increasing γ yields diminishing returns. Based on this observation, we fix γ=10 as the default value in all experiments."

    The paper presents γ=10, τ=0.95, N=512, and λ=1 as fixed global hyperparameters, but Appendix D selects them by sensitivity sweeps on OCRBench, TextVQA, and TextCaps at C=64 — the same tasks, model, and budget reported in Tables 1-2. The headline claim that MM-ShiftKV 'consistently outperforms existing methods under strict KV-cache budgets' is therefore not an independent test at C=64 on those benchmarks: the constants were chosen to maximize exactly those numbers. The reported gains at the tightest budget are partly a function of fitting the evaluation target, rather than a prediction from a pre-specified method. This does not invalidate the internal ranking mechanism, but it makes the headline evidence partially constructed by the hyperparameter selection procedure.

full rationale

The central derivation is not definitionally circular: MM-ShiftKV ranks prompt keys by proxy-induced attention mass (Eq. 7) and is assessed against downstream benchmark accuracy, not against its own proxy scores. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in only via the authors' prior work. However, the strongest empirical claim — consistent gains under strict budgets — is materially weakened by the fact that the method's key hyperparameters (γ=10, τ=0.95, N=512, λ=1) are tuned in Appendix D on the same benchmarks, model, and C=64 budget used for the headline comparisons in Tables 1-2. The paper even states 'All hyperparameters are fixed globally and shared across models, datasets, and budgets,' obscuring that these defaults were selected on the evaluation set. This is a fitted-input-called-prediction issue: part of the reported advantage at C=64 on OCRBench/TextVQA/TextCaps reduces to choosing the constants that maximize those exact results. The remaining evidence (ablations, other budgets, other tasks, and outlier-coverage measurements in Appendix E) gives some independent support, so the circularity is partial rather than total. Correctness risks such as the unvalidated query-space Gaussian and the softmax-versus-unnormalized justification are separate from circularity and do not further raise this score.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The method's functional form rests on the Gaussian proxy assumption (Eq. 5) and on the empirical claim Sigma_dec > Sigma_pre (Eq. 13). The global constants gamma, tau, N, G are tuned on the evaluation suite, so part of the paper's contribution is parameter selection rather than derivation. No new physical entities are introduced; 'query proxies' are algorithmic constructions.

free parameters (5)
  • gamma = 10
    Variance expansion factor in Eq. 5; tuned via Table 6 on OCRBench/TextVQA/TextCaps at C=64, not derived from measured query variance.
  • tau = 0.95
    Cumulative attention mass threshold in Eq. 9; selected via Table 7 on the evaluation benchmarks.
  • N = 512
    Number of query proxies; selected via Table 8 on the evaluation benchmarks.
  • G = 32
    Number of proxy groups; listed as a fixed global hyperparameter, with group-wise voting central to the method.
  • lambda = 1
    Last-query anchor weight in Eq. 11; sensitivity analyzed in Appendix E.4.
axioms (4)
  • domain assumption Hidden states at sampling entrance are approximately Gaussian with diagonal covariance
    Appendix A.1 Eq. (12) and Appendix E.6 assume h_pre ~ N(mu, Sigma_pre) and h_dec ~ N(mu, Sigma_dec); this motivates Eq. (5).
  • domain assumption Decoding-time variance dominates prefill variance (Sigma_dec > Sigma_pre)
    Eq. (13) is asserted from hidden-state measurements; it is necessary to justify gamma > 1 and is not directly verified in query space.
  • domain assumption Proxy attention mass predicts decoding-time KV importance
    Eqs. (7)-(10) rank keys by synthetic query attention; the paper provides no direct evidence that these masses match the keys decoding actually attends.
  • ad hoc to paper A single global gamma transfers across layers, heads, and datasets
    Appendix E.2 reports actual variance ratios varying 13-22x across layers, yet one gamma=10 is applied everywhere; this is a robustness choice, not derived from the measurements.
invented entities (1)
  • Variance-expanded query proxies no independent evidence
    purpose: Synthetic Gaussian samples used to estimate decoding-time attention during prefill
    Algorithmic construct with no falsifiable handle outside the method's own benchmark performance; no direct query-space validation is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 20228 in / 14598 out tokens · 147048 ms · 2026-08-02T11:52:48.457034+00:00 · methodology

0 comments
read the original abstract

Key-Value (KV) caching is essential for efficient inference in multimodal large language models (MLLMs), yet its memory footprint grows linearly with context length and becomes a major bottleneck due to the large number of visual tokens. Recent prefill-stage KV selection methods estimate KV importance from prefilling statistics, implicitly assuming that prefilling-time queries are representative of those encountered during decoding. We show that this assumption breaks down in multimodal inference, where decoding-time queries exhibit substantially larger variance than prefilling-stage representations, leading to unstable KV importance estimation under tight cache budgets. As a result, small ranking errors can disproportionately discard semantically critical visual tokens and degrade grounding and reasoning performance. We propose MM-ShiftKV, a training-free, decode-aware and strictly prefill-only KV selection method. MM-ShiftKV approximates decoding-time query behavior during prefilling by constructing variance-expanded query proxies and estimates prompt KV importance based on their aggregated attention mass. Experiments on multimodal benchmarks demonstrate that MM-ShiftKV consistently outperforms existing methods under strict KV-cache budgets. Our code is available at https://github.com/zjuDBxAI/MM-ShiftKV.

Figures

Figures reproduced from arXiv: 2607.22586 by Baokun Wang, Chenyang Wu, Jinsong Shu, Lidan Shou, Zhongle Xie.

Figure 1
Figure 1. Figure 1: Prefill–decode statistics and decode-time attention coverage. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MM-ShiftKV. The method computes statistics over the full prefill sequence of the current [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: End-to-end decoding latency and peak GPU [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The figure shows the overall numerical dis [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional visualizations of the prefill–decode [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: In the left example, the image contains two people wearing [PITH_FULL_IMAGE:figures/full_fig_p019_7.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

84 extracted references · 28 linked inside Pith

  1. [1]

    Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebr \'o n, and Sumit Sanghai. 2023. https://arxiv.org/abs/2305.13245 GQA : Training generalized multi-query transformer models from multi-head checkpoints . Preprint, arXiv:2305.13245

    Pith/arXiv arXiv 2023

  2. [2]

    Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji

    Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. 2025. HiRED : Attention-guided token dropping for efficient inference of high-resolution vision-language models. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence

    2025

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and 1 others. 2025. https://arxiv.org/abs/2502.13923 Qwen2.5-vl technical report . Preprint, arXiv:2502.13923

    Pith/arXiv arXiv 2025

  4. [6]

    Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. 2025. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In Computer Vision -- ECCV 2024 , pages 19--35, Cham. Springer Nature Switzerland

    2025

  5. [8]

    Tri Dao. 2023. https://arxiv.org/abs/2307.08691 Flashattention-2: Faster attention with better parallelism and work partitioning . Preprint, arXiv:2307.08691

    Pith/arXiv arXiv 2023

  6. [9]

    Alessio Devoto, Maximilian Jeblick, and Simon J \'e gou. 2025. https://arxiv.org/abs/2510.00636 Expected attention: Kv cache compression by estimating attention from future query distributions . Preprint, arXiv:2510.00636

    arXiv 2025

  7. [13]

    Yicheng Ji, Jun Zhang, Jinpeng Chen, Cong Wang, Lidan Shou, Gang Chen, and Huan Li. 2026. https://arxiv.org/abs/2604.05650 See the forest for the trees: Loosely speculative decoding via visual-semantic guidance for efficient inference of video llms . Preprint, arXiv:2604.05650

    Pith/arXiv arXiv 2026

  8. [14]

    Abdi, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu

    Huiqiang Jiang, Yucheng Li, Chengruidong Zhang, Qianhui Wu, Xufang Luo, Surin Ahn, Zhenhua Han, Amir H. Abdi, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. MInference 1.0: Accelerating pre-filling for long-context LLMs via dynamic sparse attention. In Advances in Neural Information Processing Systems ( NeurIPS 2024) , volume 37, pages 52481--52515

    2024

  9. [16]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and 1 others. 2024 a . https://arxiv.org/abs/2408.03326 Llava-onevision: Easy visual task transfer . Preprint, arXiv:2408.03326

    Pith/arXiv arXiv 2024

  10. [17]

    Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. 2024 b . Snapkv: Llm knows what you are looking for before generation. In Advances in Neural Information Processing Systems, volume 37

    2024

  11. [18]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. AWQ : Activation-aware weight quantization for on-device LLM compression and acceleration. In Proceedings of Machine Learning and Systems ( MLSys ) , volume 6, pages 87--100

    2024

  12. [19]

    Junyan Lin, Haoran Chen, Yue Fan, Yingqi Fan, Xin Jin, Hui Su, Jinlan Fu, and Xiaoyu Shen. 2025. Multi-layer visual feature fusion in multimodal LLMs : Methods, analysis, and best practices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 4156--4166

    2025

  13. [20]

    Jiahui Liu, Praveen Ponnusamy, Tianle Cai, Hanlin Guo, Yoon Kim, and Ben Athiwaratkun. 2024 a . https://arxiv.org/abs/2408.14690 Training-free activation sparsity in large language models . Preprint, arXiv:2408.14690

    Pith/arXiv arXiv 2024

  14. [23]

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. 2022. ChartQA : A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022 , pages 2263--2279, Dublin, Ireland. Association for Computational Linguistics

    2022

  15. [24]

    Minesh Mathew, Dimosthenis Karatzas, and C.V. Jawahar. 2021. DocVQA : A dataset for VQA on document images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision ( WACV ) , pages 2200--2209

    2021

  16. [25]

    Morse, Raghavv Goel, Mingu Lee, and Chris Lott

    Junyoung Park, Dalton Jones, Matthew J. Morse, Raghavv Goel, Mingu Lee, and Chris Lott. 2025. https://arxiv.org/abs/2504.15364 Keydiff: Key similarity-based kv cache eviction for long-context llm inference in resource-constrained environments . Preprint, arXiv:2504.15364

    arXiv 2025

  17. [26]

    David Peer, Philemon Sch \"o pf, Volckmar Nebendahl, Alexander Rietzler, and Sebastian Stabinger. 2024. https://arxiv.org/abs/2402.03848 ANLS* -- a universal document processing metric for generative large language models . Preprint, arXiv:2402.03848

    Pith/arXiv arXiv 2024

  18. [28]

    Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. 2020. TextCaps : A dataset for image captioning with reading comprehension. In Computer Vision -- ECCV 2020 , pages 742--758, Cham. Springer International Publishing

    2020

  19. [29]

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. Towards VQA models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR )

    2019

  20. [31]

    Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. 2025. DyCoke : Dynamic compression of tokens for fast video large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 18992--19001

    2025

  21. [34]

    Jiahui Wang, Zuyan Liu, Yongming Rao, and Jiwen Lu. 2025. https://arxiv.org/abs/2506.05344 Sparsemm: Head sparsity emerges from visual concept responses in mllms . Preprint, arXiv:2506.05344

    Pith/arXiv arXiv 2025

  22. [35]

    Chaojun Xiao, Pengle Zhang, Xu Han, Guangxuan Xiao, Yankai Lin, Zhengyan Zhang, Zhiyuan Liu, and Maosong Sun. 2024. InfLLM : Training-free long-context extrapolation for LLMs with an efficient context memory. In Proceedings of the 38th International Conference on Neural Information Processing Systems ( NeurIPS 2024) , Red Hook, NY, USA. Curran Associates Inc

    2024

  23. [36]

    Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2023. https://arxiv.org/abs/2309.17453 Efficient streaming language models with attention sinks . Preprint, arXiv:2309.17453

    Pith/arXiv arXiv 2023

  24. [37]

    Jiaming Xu, Jiayi Pan, Yongkang Zhou, Siming Chen, Jinhao Li, Yaoxiu Lian, Junyi Wu, and Guohao Dai. 2025. SpecEE : Accelerating large language model inference with speculative early exiting. In Proceedings of the 52nd Annual International Symposium on Computer Architecture ( ISCA '25) , pages 467--481, New York, NY, USA. Association for Computing Machinery

    2025

  25. [38]

    Cheng Yang, Yang Sui, Jinqi Xiao, Lingyi Huang, Yu Gong, Chendi Li, Jinghua Yan, Yu Bai, Ponnuswamy Sadayappan, Xia Hu, and Bo Yuan. 2025 a . TopV : Compatible token pruning with inference time optimization for fast and low-memory multimodal vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR ...

    2025

  26. [39]

    Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. 2025 b . VisionZip : Longer is better but not necessary in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 19792--19802

    2025

  27. [40]

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, and 3 others. 2024. MMMU : A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI . In...

    2024

  28. [41]

    Bowen Zeng, Feiyang Ren, Jun Zhang, Xiaoling Gu, Ke Chen, Lidan Shou, and Huan Li. 2026. https://arxiv.org/abs/2604.05887 Hybridkv: Hybrid kv cache compression for efficient multimodal large language model inference . Preprint, arXiv:2604.05887

    Pith/arXiv arXiv 2026

  29. [42]

    Jun Zhang, Yicheng Ji, Feiyang Ren, Yihang Li, Bowen Zeng, Zonghao Chen, Ke Chen, Lidan Shou, Gang Chen, and Huan Li. 2026. https://arxiv.org/abs/2604.05546 Efficient inference for large vision-language models: Bottlenecks, techniques, and prospects . Preprint, arXiv:2604.05546

    Pith/arXiv arXiv 2026

  30. [44]

    Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis A Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang. 2025. https://openreview.net/forum?id=1xG3MN1RRW SparseVLM : Visual token sparsification for efficient vision-language models inference

    2025

  31. [45]

    Simeng Zheng, Chih-Hui Ho, Wenyu Peng, and Paul H. Siegel. 2025. https://doi.org/10.1109/TCOMM.2024.3442691 Flash-Gen : Spatio-temporal generator for flash memory systems . IEEE Transactions on Communications, 73(2):1100--1113

    arXiv 2025

  32. [46]

    Yuxin Zhou, Zheng Li, Jun Zhang, Jue Wang, Yiping Wang, Zhongle Xie, Ke Chen, and Lidan Shou. 2025. https://proceedings.mlr.press/v267/ Floe: On-the-fly moe inference on memory-constrained gpu . In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research. PMLR

    2025

  33. [47]

    2025 , eprint =

    Expected Attention: KV Cache Compression by Estimating Attention from Future Query Distributions , author =. 2025 , eprint =

    2025

  34. [48]

    2025 , eprint =

    SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs , author =. 2025 , eprint =

    2025

  35. [49]

    Advances in Neural Information Processing Systems , year =

    SnapKV: LLM Knows What You Are Looking For Before Generation , author =. Advances in Neural Information Processing Systems , year =

  36. [50]

    2025 , eprint =

    KeyDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments , author =. 2025 , eprint =

    2025

  37. [51]

    2023 , eprint =

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , author =. 2023 , eprint =

    2023

  38. [52]

    2023 , eprint =

    Efficient Streaming Language Models with Attention Sinks , author =. 2023 , eprint =

    2023

  39. [53]

    2025 , eprint =

    Qwen2.5-VL Technical Report , author =. 2025 , eprint =

    2025

  40. [54]

    2024 , eprint =

    LLaVA-OneVision: Easy Visual Task Transfer , author =. 2024 , eprint =

    2024

  41. [55]

    2024 , eprint =

    Training-Free Activation Sparsity in Large Language Models , author =. 2024 , eprint =

    2024

  42. [56]

    OCRBench : On the Hidden Mystery of OCR in Large Multimodal Models

    Liu, Yuliang and Li, Zhang and Huang, Mingxin and Yang, Biao and Yu, Wenwen and Li, Chunyuan and Yin, Xucheng and Liu, Cheng-lin and Jin, Lianwen and Bai, Xiang , year =. OCRBench : On the Hidden Mystery of OCR in Large Multimodal Models. 2305.07895 , archivePrefix =

    Pith/arXiv arXiv

  43. [57]

    TextCaps : A Dataset for Image Captioning with Reading Comprehension

    Sidorov, Oleksii and Hu, Ronghang and Rohrbach, Marcus and Singh, Amanpreet , booktitle = "Computer Vision --. TextCaps : A Dataset for Image Captioning with Reading Comprehension. 2020

    2020

  44. [58]

    MMMU : A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI

    Yue, Xiang and Ni, Yuansheng and Zhang, Kai and Zheng, Tianyu and Liu, Ruoqi and Zhang, Ge and Stevens, Samuel and Jiang, Dongfu and Ren, Weiming and Sun, Yuxuan and Wei, Cong and Yu, Botao and Yuan, Ruibin and Sun, Renliang and Yin, Ming and Zheng, Boyuan and Yang, Zhenzhu and Liu, Yibo and Huang, Wenhao and Sun, Huan and Su, Yu and Chen, Wenhu , booktit...

    2024

  45. [59]

    Findings of the Association for Computational Linguistics:. ChartQA : A Benchmark for Question Answering about Charts with Visual and Logical Reasoning. 2022 , address =

    Masry, Ahmed and Long, Do Xuan and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , booktitle = "Findings of the Association for Computational Linguistics:. ChartQA : A Benchmark for Question Answering about Charts with Visual and Logical Reasoning. 2022 , address = "Dublin, Ireland", publisher = "Association for Computational Linguistics", pages =

    2022

  46. [60]

    Towards VQA Models That Can Read

    Singh, Amanpreet and Natarajan, Vivek and Shah, Meet and Jiang, Yu and Chen, Xinlei and Batra, Dhruv and Parikh, Devi and Rohrbach, Marcus , booktitle = "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (. Towards VQA Models That Can Read

  47. [61]

    , booktitle = "Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C.V. , booktitle = "Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (. DocVQA : A Dataset for VQA on Document Images. 2021 , pages =

    2021

  48. [62]

    KIVI : A Tuning-Free Asymmetric 2-bit Quantization for KV Cache

    Liu, Zirui and Yuan, Jiayi and Jin, Hongye and Zhong, Shaochen and Xu, Zhaozhuo and Braverman, Vladimir and Chen, Beidi and Hu, Xia , year =. KIVI : A Tuning-Free Asymmetric 2-bit Quantization for KV Cache. 2402.02750 , archivePrefix =

    Pith/arXiv arXiv

  49. [63]

    AWQ : Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration

    Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and Chen, Wei-Ming and Wang, Wei-Chen and Xiao, Guangxuan and Dang, Xingyu and Gan, Chuang and Han, Song , booktitle = "Proceedings of Machine Learning and Systems (. AWQ : Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration. 2024 , volume =

    2024

  50. [64]

    Computer Vision --. An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models. 2025 , publisher =

    Chen, Liang and Zhao, Haozhe and Liu, Tianyu and Bai, Shuai and Lin, Junyang and Zhou, Chang and Chang, Baobao , booktitle = "Computer Vision --. An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models. 2025 , publisher = "Springer Nature Switzerland", address = "Cham", pages =

    2025

  51. [65]

    and Chen, Deming and Dao, Tri , year =

    Cai, Tianle and Li, Yuhong and Geng, Zhengyang and Peng, Hongwu and Lee, Jason D. and Chen, Deming and Dao, Tri , year =. Medusa : Simple LLM Inference Acceleration Framework with Multiple Decoding Heads. 2401.10774 , archivePrefix =

    Pith/arXiv arXiv

  52. [66]

    Association for Computing Machinery

    Xu, Jiaming and Pan, Jiayi and Zhou, Yongkang and Chen, Siming and Li, Jinhao and Lian, Yaoxiu and Wu, Junyi and Dai, Guohao , booktitle = "Proceedings of the 52nd Annual International Symposium on Computer Architecture (. SpecEE : Accelerating Large Language Model Inference with Speculative Early Exiting. 2025 , publisher = "Association for Computing Mac...

    2025

  53. [67]

    InfLLM : Training-free Long-context Extrapolation for LLMs with an Efficient Context Memory

    Xiao, Chaojun and Zhang, Pengle and Han, Xu and Xiao, Guangxuan and Lin, Yankai and Zhang, Zhengyan and Liu, Zhiyuan and Sun, Maosong , booktitle = "Proceedings of the 38th International Conference on Neural Information Processing Systems (. InfLLM : Training-free Long-context Extrapolation for LLMs with an Efficient Context Memory

  54. [68]

    H2O : Heavy-hitter Oracle for Efficient Generative Inference of Large Language Models

    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. H2O : Heavy-hitter Oracle for Efficient Generative Inference of Large Language Models. Proceedings of the 37th International Conference on Neural Information Processing Systems ( NeurIPS 2023)

    2023

  55. [69]

    RotateKV : Accurate and Robust 2-bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations

    Su, Zunhai and Chen, Zhe and Shen, Wang and Wei, Hanyu and Li, Linge and Yu, Huangqi and Yuan, Kehong , year =. RotateKV : Accurate and Robust 2-bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations. 2501.16383 , archivePrefix =

    Pith/arXiv arXiv

  56. [70]

    Generating Long Sequences with Sparse Transformers

    Child, Rewon and Gray, Scott and Radford, Alec and Sutskever, Ilya , year =. Generating Long Sequences with Sparse Transformers. 1904.10509 , archivePrefix =

    Pith/arXiv arXiv 1904

  57. [71]

    Training-free and Adaptive Sparse Attention for Efficient Long Video Generation

    Xia, Yifei and Ling, Suhan and Fu, Fangcheng and Wang, Yujie and Li, Huixia and Xiao, Xuefeng and Cui, Bin , booktitle = "Proceedings of the IEEE/CVF International Conference on Computer Vision (. Training-free and Adaptive Sparse Attention for Efficient Long Video Generation. 2025 , pages =

    2025

  58. [72]

    A Simple and Effective L_2 Norm-Based Strategy for KV Cache Compression

    Devoto, Alessio and Zhao, Yu and Scardapane, Simone and Minervini, Pasquale , year =. A Simple and Effective L_2 Norm-Based Strategy for KV Cache Compression. 2406.11430 , archivePrefix =

    Pith/arXiv arXiv

  59. [73]

    , journal =

    Zheng, Simeng and Ho, Chih-Hui and Peng, Wenyu and Siegel, Paul H. , journal =. Flash-Gen : Spatio-Temporal Generator for Flash Memory Systems. 2025 , volume =

    2025

  60. [74]

    PyramidKV : Dynamic KV Cache Compression based on Pyramidal Information Funneling

    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 , year =. PyramidKV : Dynamic KV Cache Compression based on Pyramidal Information Funneling. 2406.02069 , archivePrefix =

    Pith/arXiv arXiv

  61. [75]

    VisionZip : Longer is Better but Not Necessary in Vision-Language Models

    Yang, Senqiao and Chen, Yukang and Tian, Zhuotao and Wang, Chengyao and Li, Jingyao and Yu, Bei and Jia, Jiaya , booktitle = "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (. VisionZip : Longer is Better but Not Necessary in Vision-Language Models. 2025 , pages =

    2025

  62. [76]

    DyCoke : Dynamic Compression of Tokens for Fast Video Large Language Models

    Tao, Keda and Qin, Can and You, Haoxuan and Sui, Yang and Wang, Huan , booktitle = "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (. DyCoke : Dynamic Compression of Tokens for Fast Video Large Language Models. 2025 , pages =

    2025

  63. [77]

    and Soran, Bilge and Krishnamoorthi, Raghuraman and Elhoseiny, Mohamed and Chandra, Vikas , year =

    Shen, Xiaoqian and Xiong, Yunyang and Zhao, Changsheng and Wu, Lemeng and Chen, Jun and Zhu, Chenchen and Liu, Zechun and Xiao, Fanyi and Varadarajan, Balakrishnan and Bordes, Florian and Liu, Zhuang and Xu, Hu and Kim, Hyunwoo J. and Soran, Bilge and Krishnamoorthi, Raghuraman and Elhoseiny, Mohamed and Chandra, Vikas , year =. LongVU : Spatiotemporal Ad...

    Pith/arXiv arXiv

  64. [78]

    TopV : Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision-Language Model

    Yang, Cheng and Sui, Yang and Xiao, Jinqi and Huang, Lingyi and Gong, Yu and Li, Chendi and Yan, Jinghua and Bai, Yu and Sadayappan, Ponnuswamy and Hu, Xia and Yuan, Bo , booktitle = "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (. TopV : Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory...

    2025

  65. [79]

    LOOK-M : Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference

    Wan, Zhongwei and Wu, Ziang and Liu, Che and Huang, Jinfa and Zhu, Zhihong and Jin, Peng and Wang, Longyue and Yuan, Li , year =. LOOK-M : Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference. 2406.18139 , archivePrefix =

    Pith/arXiv arXiv

  66. [80]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Li, Kunxi and Jiang, Zhonghua and Shen, Zhouzhou and ZhaodeWang, ZhaodeWang and Lv, Chengfei and Zhang, Shengyu and Wu, Fan and Wu, Fei , booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", year =. MadaKV : Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Lon...

  67. [81]

    DuoAttention : Efficient Long-Context LLM Inference with Retrieval and Streaming Heads

    Xiao, Guangxuan and Tang, Jiaming and Zuo, Jingwei and Guo, Junxian and Yang, Shang and Tang, Haotian and Fu, Yao and Han, Song , year =. DuoAttention : Efficient Long-Context LLM Inference with Retrieval and Streaming Heads. 2410.10819 , archivePrefix =

    Pith/arXiv arXiv

  68. [82]

    Kevin , year =

    Feng, Yuan and Lv, Junlin and Cao, Yukun and Xie, Xike and Zhou, S. Kevin , year =. Ada-KV : Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference. 2407.11550 , archivePrefix =

    Pith/arXiv arXiv

  69. [83]

    and Li, Dongsheng and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , booktitle = "Advances in Neural Information Processing Systems (

    Jiang, Huiqiang and Li, Yucheng and Zhang, Chengruidong and Wu, Qianhui and Luo, Xufang and Ahn, Surin and Han, Zhenhua and Abdi, Amir H. and Li, Dongsheng and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , booktitle = "Advances in Neural Information Processing Systems (. MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse At...

    2024

  70. [84]

    SparseVLM : Visual Token Sparsification for Efficient Vision-Language Models Inference

    Zhang, Yuan and Fan, Chun-Kai and Ma, Junpeng and Zheng, Wenzhao and Huang, Tao and Cheng, Kuan and Gudovskiy, Denis A and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , year =. SparseVLM : Visual Token Sparsification for Efficient Vision-Language Models Inference

  71. [85]

    and Chen, Beidi and Wu, Carole-Jean , year =

    Elhoushi, Mostafa and Shrivastava, Akshat and Liskovich, Diana and Hosmer, Basil and Wasti, Bram and Lai, Liangzhen and Mahmoud, Anas and Acun, Bilge and Agarwal, Saurabh and Roman, Ahmed and Aly, Ahmed A. and Chen, Beidi and Wu, Carole-Jean , year =. LayerSkip : Enabling Early Exit Inference and Self-Speculative Decoding. 2404.16710 , archivePrefix =

    Pith/arXiv arXiv

  72. [86]

    GQA : Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

    Ainslie, Joshua and Lee-Thorp, James and de Jong, Michiel and Zemlyanskiy, Yury and Lebr. GQA : Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints. 2023 , eprint =

    2023

  73. [87]

    Sparse Upcycling : Training Mixture-of-Experts from Dense Checkpoints

    Komatsuzaki, Aran and Puigcerver, Joan and Lee-Thorp, James and Riquelme Ruiz, Carlos and Mustafa, Basil and Ainslie, Joshua and Tay, Yi and Dehghani, Mostafa and Houlsby, Neil , year =. Sparse Upcycling : Training Mixture-of-Experts from Dense Checkpoints. 2212.05055 , archivePrefix =

    Pith/arXiv arXiv

  74. [88]

    Sparse VideoGen : Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity

    Xi, Haocheng and Yang, Shuo and Zhao, Yilong and Xu, Chenfeng and Li, Muyang and Li, Xiuyu and Lin, Yujun and Cai, Han and Zhang, Jintao and Li, Dacheng and Chen, Jianfei and Stoica, Ion and Keutzer, Kurt and Han, Song , year =. Sparse VideoGen : Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity. 2502.01776 , archivePrefix =

    Pith/arXiv arXiv

  75. [89]

    Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence

    Arif, Kazi Hasan Ibn and Yoon, JinYi and Nikolopoulos, Dimitrios S. and Vandierendonck, Hans and John, Deepu and Ji, Bo , booktitle = "Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence", year =. HiRED : Attention-guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models

  76. [90]

    Multi-Layer Visual Feature Fusion in Multimodal LLMs : Methods, Analysis, and Best Practices

    Lin, Junyan and Chen, Haoran and Fan, Yue and Fan, Yingqi and Jin, Xin and Su, Hui and Fu, Jinlan and Shen, Xiaoyu , booktitle = "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (. Multi-Layer Visual Feature Fusion in Multimodal LLMs : Methods, Analysis, and Best Practices. 2025 , pages =

    2025

  77. [91]

    LMMs-Eval : Reality Check on the Evaluation of Large Multimodal Models

    Zhang, Kaichen and Li, Bo and Zhang, Peiyuan and Pu, Fanyi and Cahyono, Joshua Adrian and Hu, Kairui and Liu, Shuai and Zhang, Yuanhan and Yang, Jingkang and Li, Chunyuan and Liu, Ziwei , year =. LMMs-Eval : Reality Check on the Evaluation of Large Multimodal Models. 2407.12772 , archivePrefix =

    Pith/arXiv arXiv

  78. [92]

    Exact Matching: Algorithms and Related Problems

    El Maalouly, Nicolas , year =. Exact Matching: Algorithms and Related Problems. 2203.13899 , archivePrefix =

    Pith/arXiv arXiv

  79. [93]

    Lawrence and Parikh, Devi , year =

    Vedantam, Ramakrishna and Zitnick, C. Lawrence and Parikh, Devi , year =. CIDEr : Consensus-based Image Description Evaluation. 1411.5726 , archivePrefix =

    Pith/arXiv arXiv

  80. [94]

    ANLS* -- A Universal Document Processing Metric for Generative Large Language Models

    Peer, David and Sch. ANLS* -- A Universal Document Processing Metric for Generative Large Language Models. 2024 , eprint =

    2024

Showing first 80 references.