Paper deep dive
VPRune: Efficient Training-free Pre-LLM Visual Token Pruning
Guangchuan Lv, Dianxing Shi, Dingjie FU
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Visual token pruning is a promising approach to reducing the inference cost of large vision-language models (LVLMs), yet aggressive token reduction often causes substantial performance degradation. We identify three key factors behind this degradation: text-guided selection bias, information loss from discarded tokens, and positional distortion caused by sequence compaction. Based on these observations, we propose \textbf{VPRune}, a training-free pre-LLM pruning framework consisting of visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. Experiments on FastVLM-1.5B across multiple vision-language benchmarks demonstrate that VPRune achieves a favorable accuracy--compression trade-off, with particularly pronounced advantages under aggressive compression. Furthermore, evaluations on edge-device show that VPRune effectively reduces end-to-end inference latency while maintaining superior task performance, demonstrating its practicality for resource-constrained LVLM deployment.
Tags
Links
- Source: https://arxiv.org/abs/2609.24485v1
- Canonical: https://arxiv.org/abs/2609.24485v1
Trouble viewing inline? Open PDF directly →
Full Text
23,349 characters extracted from source content.
Expand or collapse full text
VPRUNE: EFFICIENT TRAINING-FREE PRE-LLM VISUAL TOKEN PRUNING Guangchuan Lv 1 , Dianxing Shi 2∗ , Dingjie Fu 3 1 Northeastern University, 2 Beihang University, 3 Huazhong University of Science and Technology ABSTRACT Visual token pruning is a promising approach to reducing the inference cost of large vision-language models (LVLMs), yet aggressive token reduction often causes substantial per- formance degradation.We identify three key factors be- hind this degradation: text-guided selection bias, informa- tion loss from discarded tokens, and positional distortion caused by sequence compaction. Based on these observa- tions, we propose VPRune, a training-free pre-LLM prun- ing framework consisting of visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. Experiments on FastVLM-1.5B across multi- ple vision-language benchmarks demonstrate that VPRune achieves a favorable accuracy–compression trade-off, with particularly pronounced advantages under aggressive com- pression. Furthermore, evaluations on edge-device show that VPRune effectively reduces end-to-end inference latency while maintaining superior task performance, demonstrating its practicality for resource-constrained LVLM deployment. Index Terms— Vision-language models, visual token pruning, token reduction, efficient inference, edge AI 1. INTRODUCTION Large vision-language models (LVLMs) achieve strong mul- timodal understanding capability by jointly processing visual and textual tokens [1, 2, 3]. However, visual representations often contain substantial redundancy, while long visual-token sequences increase inference latency and hinder deployment on resource-constrained edge devices. To this end, visual to- ken pruning alleviates this cost by removing redundant visual tokens aggressively [4, 5], but such reduction often causes severe performance degradation, particularly when pruning is performed before the LLM. Existing visual-token reduction methods generally im- prove efficiency by removing redundant tokens before the LLM or within early decoder layers [4, 6, 7, 8]. To identify informative tokens, they typically rely on visual redundancy, attention responses, or cross-modal relevance [5, 9]. How- ever, many existing pruning methods are developed around token-rich LVLM settings, while increasingly efficient edge- oriented models are designed with compact multimodal repre- ∗ Corresponding author: dianxingshi10@gmail.com (a) VPRuneHiPruneCDPruner Normalized Average (↑) VPRune 94.87% HiPrune 81.93% CDPruner 82.71% (b) 100% 94.87% Baseline 6.13s 9.88s (c) Consistent TTFT reduction 80 100 10 30 60 80 100 Retained visual tokens (%) 79.7 85.0 91.2 96.0 Baseline VPRune VPRune 246810120 LLM prefilltime (seconds) Lower Prefilling timeon QCS8550 CPU 100 VPRune Baseline 90 70 38.0% lower TextVQA GQA POPE-adv ChartQA 66.36 62.13 88.13 66.00 70.80 AI2D-lite 36.08 42.36 86.23 87.33 67.40 66.00 57.27 54.41 58.45 58.12 Fig. 1. Performance and edge efficiency of VPRune. At 30% token retention, VPRune achieves the highest normalized av- erage performance of 94.87% in (a), while reducing LLM pre- fill latency by 38.0% on the QCS8550 CPU in (b); (c) further shows consistent end-to-end TTFT reductions across differ- ent visual-token budgets. sentations for practical deployment [3, 10, 11]. In this regime, applying the same retention ratio yields a much tighter abso- lute token budget, making further pruning more sensitive to information loss and representation distortion. To understand these degradations, we analyze pre-LLM pruning under tight token budgets and identify three em- pirical observations: i) text-conditioned relevance can over- concentrate token selection and reduce visual coverage, i) discarded tokens are often redundant rather than uninforma- tive and may still contain complementary visual evidence, and i) sequence compaction reassigns retained tokens to consec- utive position indices, distorting the spatial relationships seen by the pretrained LLM. Motivated by these observations, we propose VPRune, a training-free pre-LLM framework with three corresponding components. Specifically, we first introduce Visual-only Diversity Selection (VDS) to removes text conditioning and preserves visual diversity. We then employ Similarity-guided Token Recycling (STR) to transfer complementary information from discarded tokens to similar retained tokens. Moreover, we propose Position-Preserving arXiv:2609.24485v1 [cs.CV] 21 Sep 2026 Restoration (PPR) to preserve their original raster order and position indices. Together, these components preserve visual evidence and positional structure under high compression without modifying the underlying LVLM. As highlighted in Fig. 1, VPRune provides a favorable accuracy–efficiency trade-off under aggressive compression. At 30% token retention, it achieves a normalized average per- formance of 94.87%, substantially outperforming representa- tive methods. At the same retention ratio, VPRune reduces LLM prefill latency from 9.88 s to 6.13 s (38.0%) on a Qual- comm QCS8550 platform, while also providing consistent TTFT reductions across different token budgets. Our contributions are: i) We analyze three key causes of performance degradation under aggressive visual-token prun- ing. i) We propose VPRune, a training-free pre-LLM frame- work integrating VDS, STR, and PPR to advance low-budget token reduction performance.i) Extensive experiments demonstrate strong accuracy under aggressive compression together with practical prefill and end-to-end latency reduc- tions on a CPU-only edge platform. 2. RELATED WORK Modern LVLMs typically employ a vision encoder, mul- timodal projector, and LLM to jointly process visual and textual representations [1, 2]. Efficient architectures such as FastVLM [3] reduce visual-processing overhead, while visual token reduction provides a complementary way to remove re- dundant representations. Existing methods prune tokens at different stages: HiPrune [6] and VisionZip [4] exploit vi- sual structure or attention patterns, CDPruner [5] combines diversity-aware DPP selection with text-conditioned rele- vance, FlashVLM [9] performs efficient pre-LLM filtering, VScan [7] incorporates query-aware token reduction, and FastV [8] prunes tokens after early LLM processing. De- spite their effectiveness, preserving visual content, positional consistency, and discarded information remains challeng- ing under aggressive compression. VPRune addresses these issues entirely before LLM computation. 3. METHODOLOGY 3.1. Overview As illustrated in Fig. 2, VPRune is a training-free pre-LLM visual token pruning framework designed for aggressive to- ken budgets. Given the projected visual tokens X =x i N i=1 , where N denotes the number of visual tokens. VPRune con- sists of three stages. First, Visual-only Diversity Selection (VDS) removes the potentially misleading text-relevance bias and selects a compact yet diverse visual subset. Second, Similarity-guided Token Recycling (STR) recovers comple- mentary information from discarded tokens by aggregating Overall Pipeline of VPRune VPRune: Training-free Pre-LLM Pruning Input image Vision Encoder Visual features V Projector Projected visual tokens X VPRune LLM “A dog is running on the beach.” Output “What is happening in this picture?” Text prompt Tokenizer Text tokens T Visual tokenText token Discarded token Recycled visual token Visual-only Diversity Selection (VDS) Selected visual tokens Hᵥ Similarity-guided Token Recycling (STR) Unassigned visual tokens remain unchanged; assigned ones are recycled Recycled visual tokens Ĥᵥ Position-Preserving Restoration (PPR) Final visual tokens Hᵥ* Selected visual tokens Hᵥ Discarded tokens ... ... ... ... ... Projected visual tokens X ... (Text tokens not used) ... Select K diverse tokens via greedy DPP ... × ... ... Ignore text relevance Recycled tokens Ĥᵥ ... 푝 ! =푁 "#$%!& +푞 ! Restore the original raster order and positional indices 푁 !"#$%& : number of tokens before <image> 푞 % : visual token index ... ... ... Final visual tokens Hᵥ* ... (VDS → STR → PPR) ... Position IDs (non-consecutive) 푁 !"#$%& +푞 ' Compact sequence of K tokens 푁 !"#$%& +푞 ( 푁 !"#$%& +푞 ) ... 푁 !"#$%& +푞 * (a) (b) (c) Fig. 2.Overview of VPRune. VDS selects diverse vi- sual tokens, STR recycles complementary information from discarded tokens, and PPR restores their original positional structure before LLM inference. them into their most similar retained tokens. Finally, Position- Preserving Restoration (PPR) restores the original raster or- der and positional indices of the retained tokens before they are fed into the LLM. For a token retention ratio b, we retain K = round(bN) visual tokens. In our setting, N = 256, and pruning is per- formed after the multimodal projector and before the LLM. Thus, the LLM operates on the reduced visual sequence. 3.2. Visual-only Diversity Selection Let L denote the cosine-similarity-based visual diversity ker- nel and D r = diag(ˆr 1 ,..., ˆr N ) is the diagonal matrix of normalized text-relevance scores. The conditioned kernel is therefore e L = D r LD r . For a retained subset S of size K, the corresponding DPP objective [12] can be decomposed as log det( e L S ) = log det(L S ) + 2 X i∈S log ˆr i .(1) This decomposition reveals that text relevance contributes an explicit additive bias to the diversity objective. Under aggres- sive compression, visually complementary tokens may there- fore be strongly suppressed because they receive low text- relevance scores. VDS removes this conditioning and per- forms selection solely according to visual diversity. The re- tained subset is obtained by S ∗ = arg max |S|=K log det(L S ).(2) We solve this objective with greedy MAP inference [13]. Table 1. Performance comparison on FastVLM-1.5B under different visual-token reduction ratios. Baseline denotes the full- token model; “Reduce (%)” is the percentage of visual tokens removed relative to it. “Avg.” is the mean of the five benchmark scores below normalized by their respective full-token baselines (Baseline = 100%). Best and second-best results among reduction methods in each column are highlighted in bold and underlined, respectively. MethodTextVQAGQAPOPE-advAI2D-liteChartQAAvg. Reduce (%)204070204070204070204070204070204070 Baseline70.5863.5387.9773.0077.32100 (%) CDPruner67.08 64.17 54.41 61.85 60.91 58.45 87.77 87.50 87.3371.00 68.80 66.00 63.88 57.28 42.36 94.41 90.92 82.71 VScan67.62 55.90 32.06 63.21 60.64 54.71 87.53 84.30 77.57 72.2069.80 66.40 76.96 57.04 39.84 98.65 87.97 72.44 FastV69.19 66.67 54.05 63.49 62.9758.6988.1087.9383.23 71.20 70.00 66.40 75.64 73.6454.3698.7096.9384.97 HiPrune69.17 67.7357.27 62.33 61.15 58.12 88.23 88.13 86.23 71.40 70.4067.4067.40 61.08 36.08 96.28 93.57 81.93 VisionZip56.82 55.04 39.78 59.29 59.17 57.64 87.77 87.63 86.43 65.60 66.40 65.20 35.12 35.28 29.12 81.78 81.46 74.46 FlashVLM69.5867.54 57.8762.39 60.85 57.17 88.07 87.77 85.93 71.20 70.00 66.40 68.12 62.08 35.20 96.51 93.49 81.23 VPRune (Ours)70.5769.6466.3663.2963.0762.1388.0087.9388.1372.4070.8070.8076.2074.0866.0099.4798.1494.87 3.3. Similarity-guided Token Recycling Directly removing unselected tokens may discard comple- mentary visual information, especially under tight token budgets.STR therefore recycles discarded features into their most similar retained tokens. For each discarded to- ken i ∈ D, we assign it to j ∗ (i) = arg max j∈S cos(x i ,x j ), reusing the similarities already computed in the DPP kernel. Let G j = i ∈ D | j ∗ (i) = j denote the discarded tokens assigned to retained token j. Their contribution is weighted according to w ij = exp(L ij /τ) P i ′ ∈G j exp(L i ′ j /τ) .(3) Rather than introducing additional tokens, STR directly up- dates the retained representation as ex j = (1− γ)x j + γ X i∈G j w ij x i ,(4) where γ controls the amount of recycled information. If no discarded token is assigned to j, i.e., G j =∅, we simply keep ex j = x j . Thus, STR recovers complementary information without increasing the number of tokens passed to the LLM. 3.4. Position-Preserving Restoration Conventional token pruning typically compacts the retained visual sequence and reassigns consecutive position indices, altering the positional relationships encoded by RoPE-based LLMs [14]. Let the retained tokens correspond to original raster indices S =q i K i=1 with q 1 <· < q K . PPR first re- stores their original sequence order and assigns each retained token the position p vis i = N prefix + q i ,(5) instead of the compact position N prefix + i.Following GAP [15], text tokens after the image start from N prefix +N , preserving the original visual-token span. Importantly, the retained embeddings remain a compact K-token sequence; gaps exist only in the assigned position indices rather than as empty token slots. This preserves the original raster or- der and relative positional distances without reconstructing a length-N visual sequence. 4. EXPERIMENTS 4.1. Experimental Setup Model and Baselines. We evaluate VPRune on the frozen FastVLM-1.5B model [3] and compare it with six repre- sentative visual-token reduction methods: CDPruner [5], VScan [7], FastV [8], HiPrune [6], VisionZip [4], and FlashVLM [9]. Benchmarks. We evaluate five complementary benchmarks: GQA [16] for compositional reasoning, TextVQA [17] for text-rich understanding, POPE-adv [18] for object hallucina- tion, a 500-sample subset of AI2D [19] (AI2D-lite) for dia- gram understanding, and ChartQA [20] for chart reasoning. Implementation Details.FastVLM-1.5B uses FastViT- HD [21] and a Qwen2 LLM. Inputs are padded to 1024 × 1024, yielding 256 visual tokens; VPRune prunes after the projector and before the LLM. We report accuracy at b ∈ 0.3, 0.6, 0.8 and additionally use b = 0.1 for ef- ficiency. Baselines are matched to the target ratios. All runs use greedy decoding (batch size 1) without fine-tuning; VPRune sets γ = 0.3 and τ = 0.1. Edge Platform.Latency is measured on a Qualcomm QCS8550-based development platform using CPU-only in- ference, with GPU and NPU acceleration disabled.The QCS8550 integrates a 64-bit Kryo CPU with a maximum clock frequency of 3.2 GHz, providing a representative resource-constrained platform for evaluating practical edge inference. 4.2. Main Results As shown in Table 1, VPRune achieves the highest normal- ized average performance at all reported reduction ratios, retaining 99.47%, 98.14%, and 94.87% of full-token perfor- mance at 20%, 40%, and 70% reduction, respectively. More importantly, its margin over the second-best method grows from 0.77 and 1.21 percentage points at 20% and 40% reduc- tion to 9.90 points over FastV at 70% reduction. At this set- ting, VPRune ranks first on all five benchmarks. The largest gains occur on TextVQA and ChartQA, where it exceeds the corresponding second-best results by 8.49 and 11.64 raw points. These tasks rely heavily on fine-grained textual or structured visual evidence and are therefore more sensitive to information removed by aggressive pruning. This pattern is consistent with VPRune’s design: VDS and STR preserve distributed and complementary visual evidence, while PPR maintains spatial relationships that are particularly important for structured tasks such as ChartQA. The smaller gap on POPE-adv, whose scores remain near the full-token baseline for several methods, further suggests that the main benefit appears on content-dense reasoning tasks rather than from a uniform score shift. 4.3. Ablation Study We conduct ablations at a token retention ratio of b = 0.3. As shown in Fig. 3, removing text-guided relevance yields a substantial gain on TextVQA and a modest improvement on GQA. Position-preserving restoration further improves both benchmarks, while similarity-guided token recycling provides consistent additional gains by recovering comple- mentary information from discarded tokens. We further study the effect of the recycling strength γ on TextVQA and GQA at b = 0.3. Performance remains rel- atively stable over moderate values of γ but degrades when excessive discarded information is introduced. We therefore set γ = 0.3, which provides the best overall trade-off across the two benchmarks. 4.4. Efficiency Analysis We evaluate practical inference efficiency on the QCS8550 CPU-only edge platform. As shown in Fig. 4, VPRune per- forms pruning after the multimodal projector and therefore maintains near-baseline vision-encoder latency, while VScan, VisionZip, and HiPrune incur higher vision-side overhead. More importantly, reducing the visual sequence be- fore LLM inference provides consistent end-to-end bene- fits. VPRune reduces TTFT from 25.551 s for the full-token baseline to 20.365 s at b = 0.1, corresponding to a 20.30% reduction. The reductions remain 14.97%, 8.82%, and 4.02% GQATextVQA 52 54 56 58 60 62 64 66 Accuracy (%) Baseline +V +V+P +V+P+R (a) Component ablation 0.00.10.20.30.40.50.60.70.8 62 63 64 65 66 Accuracy (%) TextVQA GQA (b) Recycling strength γ Fig. 3. Ablation studies under 30% token retention. (a) Cu- mulative ablation of V, P, and R improves performance, where V, P, and R denote visual-only diversity selection, position- preserving restoration, and similarity-guided token recycling, respectively. (b) γ = 0.3 gives the best overall trade-off on TextVQA and GQA. 0.80.60.30.1 Token retention ratio 14.6 14.8 15.0 15.2 15.4 15.6 15.8 16.0 16.2 16.4 Latency (s) VPRune HiPrune VScan VisionZip (a) Vision encoder 0.80.60.30.1 Token retention ratio 20.0 21.0 22.0 23.0 24.0 25.0 26.0 26.8 TTFT (s) VPRune HiPrune VScan VisionZip (b) End-to-end TTFT Fig. 4.Efficiency on the QCS8550 CPU-only edge plat- form. VPRune maintains near-baseline vision-encoder la- tency in (a) and consistently reduces end-to-end TTFT across token-retention ratios in (b). at b = 0.3, 0.6, and 0.8, respectively. Its TTFT remains com- parable to the fastest competing methods while maintaining substantially higher accuracy under aggressive compression, demonstrating a favorable accuracy–latency trade-off for edge deployment. 5. CONCLUSION In this work, we investigate the performance degradation of visual token pruning under aggressive compression and propose VPRune, a training-free pre-LLM framework for efficient LVLM inference. VPRune addresses three key lim- itations of low-budget pruning through visual-only diversity selection, similarity-guided token recycling, and position- preserving restoration.Experiments on FastVLM-1.5B across five vision-language benchmarks demonstrate that VPRune achieves a favorable accuracy–compression trade- off, particularly under tight token budgets. CPU-only edge evaluations further show reduced end-to-end latency while maintaining strong task performance, supporting efficient resource-constrained LVLM deployment. 6. REFERENCES [1] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi, “BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language mod- els,” in ICML, 23–29 Jul 2023, vol. 202 of Proceedings of Machine Learning Research, p. 19730–19742. [2] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee, “Visual instruction tuning,” in NeurIPS, 2023, vol. 36, p. 34892–34916. [3] Pavan Kumar Anasosalu Vasu, Fartash Faghri, Chun- Liang Li, Cem Koc, Nate True, Albert Antony, Gokula Santhanam, James Gabriel, Peter Grasch, Oncel Tuzel, and Hadi Pouransari, “Fastvlm: Efficient vision encod- ing for vision language models,” in CVPR, June 2025, p. 19769–19780. [4] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia, “Visionzip: Longer is better but not necessary in vision language models,” in CVPR, June 2025, p. 19792–19802. [5] Qizhe Zhang, Mengzhen Liu, Lichen Li, Ming Lu, Yuan Zhang, Junwen Pan, Qi She, and Shanghang Zhang, “Beyond attention or similarity: Maximizing condi- tional diversity for token pruning in mllms,” in NeurIPS, 2025, vol. 38, Main Conference, p. 25438–25468. [6] Jizhihui Liu, Guangdao Zhu, Feiyi Du, Niu Lian, Jun Li, Bin Chen, Weili Guan, and Yaowei Wang, “HiPrune: Hierarchical attention for efficient token pruning in vision-language models,” in ACL Findings, July 2026, p. 3274–3291. [7] Ce Zhang, Kaixin Ma, Tianqing Fang, Wenhao Yu, Hongming Zhang, Zhisong Zhang, Haitao Mi, and Dong Yu, “VScan: Rethinking visual token reduction for effi- cient large vision-language models,” TMLR, 2026. [8] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Jun- yang Lin, Chang Zhou, and Baobao Chang, “An image is worth 1/2 tokens after layer 2: Plug-and-play infer- ence acceleration for large vision-language models,” in ECCV, 2024, p. 19–35. [9] Kaitong Cai, Jusheng Zhang, Jing Yang, Yijia Fan, Pengtao Xie, Jian Wang, and Keze Wang, “Flashvlm: Text-guided visual token selection for large multimodal models,” 2025. [10] Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Bo- heng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, Yi Zeng, Lei Wu, Liuyang Bian, Zhaox- iong Wang, Long Liu, Yanzhou Yang, Han Xiao, Ao- jun Zhou, Yafei Wen, Xiaoxin Chen, Shuai Ren, and Hongsheng Li, “Bluelm-v-3b: Algorithm and system co-design for multimodal large language models on mo- bile devices,” in CVPR, June 2025, p. 4145–4155. [11] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang, “Token- packer: Efficient visual projector for multimodal LLM,” Int. J. Comput. Vis., vol. 133, no. 10, p. 6794–6812, 2025. [12] Alex Kulesza and Ben Taskar, “Determinantal Point Processes for Machine Learning,” Foundations and Trends in Machine Learning, vol. 5, p. 123–286, 2012. [13] Laming Chen, Guoxin Zhang, and Eric Zhou, “Fast greedy map inference for determinantal point process to improve recommendation diversity,” in NeurIPS, 2018, vol. 31. [14] Jianlin Su, Murtadha H. M. Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu, “Roformer: Enhanced transformer with rotary position embedding,” Neuro- computing, vol. 568, p. 127063, 2024. [15] Tzu-ChunChien,Chieh-KaiLin,Shiang-Feng Tsai, Ruei-Chi Lai, Hung-Jen Chen, and Min Sun, “Grounding-aware token pruning: Recovering from drastic performance drops in visual grounding caused by pruning,” 2025. [16] Drew A. Hudson and Christopher D. Manning, “Gqa: A new dataset for real-world visual reasoning and compo- sitional question answering,” in CVPR, June 2019. [17] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach, “Towards vqa models that can read,” in CVPR, June 2019. [18] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen, “Evaluating object hallucina- tion in large vision-language models,” in EMNLP, Dec. 2023, p. 292–305. [19] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Min- joon Seo, Hannaneh Hajishirzi, and Ali Farhadi, “A di- agram is worth a dozen images,” in ECCV, 2016, p. 235–251. [20] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque, “ChartQA: A benchmark for question answering about charts with visual and logical reasoning,” in ACL, May 2022, p. 2263–2279. [21] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan, “Fastvit: A fast hybrid vision transformer using structural reparameterization,” in ICCV, October 2023, p. 5785–5795.