Paper deep dive
RoRA: Role-Oriented Regional Allocation for Visual Token Pruning in MLLMs
Qiyanhui Lu, Han Wu, Rongjian Xu, Tingzhang Luo, Cheng Fan, Xinghao Chen, Minjing Dong, Jufeng Yang, Jianyuan Guo
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/10/2026, 4:13:10 AM
Summary
The paper introduces RoRA, a training-free framework for visual token pruning in Multimodal Large Language Models (MLLMs). RoRA allocates a fixed token budget into three roles: a protected semantic core, complementary context, and fine-grained detail. It uses Attention-Anchored Regions (AARs) to track covered object support, enabling efficient redundancy filtering and improved accuracy over baselines like D2Pruner and FastV, achieving significant speedups on NVIDIA H800.
Entities (12)
Relation Signals (9)
RoRA → partitions → Complementary Context
confidence 95% · RoRA partitions tokens into a protected semantic core, complementary context, and fine-grained detail.
RoRA → partitions → Fine-grained Detail
confidence 95% · RoRA partitions tokens into a protected semantic core, complementary context, and fine-grained detail.
RoRA → partitions → Semantic Core
confidence 95% · RoRA partitions tokens into a protected semantic core, complementary context, and fine-grained detail.
RoRA → uses → Attention-Anchored Regions
confidence 95% · builds Attention-Anchored Regions (AARs) from high-confidence anchors as lightweight proxies for covered object support.
RoRA → achievesspeedupon → NVIDIA H800
confidence 90% · corresponding to a 1.33x speedup over unpruned inference on an NVIDIA H800.
RoRA → appliedto → LLaVA-1.5
confidence 90% · We evaluate RoRA on several representative MLLMs including LLaVA-1.5
RoRA → appliedto → Qwen-VL
confidence 90% · We evaluate RoRA on several representative MLLMs including ... Qwen-VL
RoRA → outperforms → FastV
confidence 90% · Under matched budgets, RoRA consistently outperforms strong training-free baselines... FastV is affected by positional attention sinks
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal large language models (MLLMs) encode images as long visual token sequences, making prefilling and KV-cache storage expensive. Existing training-free pruning methods select tokens by importance, diversity, or spatial coverage, but treat retained tokens as interchangeable and do not explicitly track which object-related regions are already covered. We present RoRA, a training-free framework that casts visual token pruning as role-oriented regional evidence allocation. Given a fixed budget, RoRA partitions tokens into a protected semantic core, complementary context, and fine-grained detail. It first calibrates text-conditioned attention with a positional prior and a prompt-calibrated object prior, then builds Attention-Anchored Regions (AARs) from high-confidence anchors as lightweight proxies for covered object support. Context is explored mainly outside AARs, while a small AAR-guided budget restores local detail; pairwise similarity is used only for context-stage redundancy filtering. Under matched budgets, RoRA consistently outperforms strong training-free baselines across LLaVA and Qwen-VL families, retaining most of the unpruned accuracy even at aggressive pruning ratios, e.g., 96.5% of full performance at 88.9% pruning on LLaVA-1.5, and improving over D2Pruner by about 5% on Qwen3-VL at 75-90% pruning. At a 66.7% pruning ratio, RoRA requires only 0.7 ms for token selection and reduces end-to-end inference time by 24.6%, corresponding to a 1.33x speedup over unpruned inference on an NVIDIA H800.
Tags
Links
- Source: https://arxiv.org/abs/2608.07088v1
- Canonical: https://arxiv.org/abs/2608.07088v1
Trouble viewing inline? Open PDF directly →
Full Text
48,569 characters extracted from source content.
Expand or collapse full text
RoRA: Role-Oriented Regional Allocation for Visual Token Pruning in MLLMs Qiyanhui Lu1, Han Wu2, Rongjian Xu1, Tingzhang Luo1, Cheng Fan1, Xinghao Chen3, Minjing Dong1, Jufeng Yang4, Jianyuan Guo1 Abstract Multimodal large language models (MLLMs) encode images as long visual token sequences, making prefilling and KV-cache storage expensive. Existing training-free pruning methods select tokens by importance, diversity, or spatial coverage, but treat retained tokens as interchangeable and do not explicitly track which object-related regions are already covered. We present RoRA, a training-free framework that casts visual token pruning as role-oriented regional evidence allocation. Given a fixed budget, RoRA partitions tokens into a protected semantic core, complementary context, and fine-grained detail. It first calibrates text-conditioned attention with a positional prior and a prompt-calibrated object prior, then builds Attention-Anchored Regions (AARs) from high-confidence anchors as lightweight proxies for covered object support. Context is explored mainly outside AARs, while a small AAR-guided budget restores local detail; pairwise similarity is used only for context-stage redundancy filtering. Under matched budgets, RoRA consistently outperforms strong training-free baselines across LLaVA and Qwen-VL families, retaining most of the unpruned accuracy even at aggressive pruning ratios, e.g., 96.5% of full performance at 88.9% pruning on LLaVA-1.5, and improving over D2Pruner by about 5% on Qwen3-VL at 75–90% pruning. At a 66.7% pruning ratio, RoRA requires only 0.7ms for token selection and reduces end-to-end inference time by 24.6%, corresponding to a 1.33× speedup over unpruned inference on an NVIDIA H800. Code: https://github.com/LukieLuu/RoRA Introduction Multimodal large language models (MLLMs) represent images as long sequences of visual tokens. High-resolution images, multi-image inputs, and videos can easily yield hundreds to thousands of tokens, accounting for most of the prefilling cost, KV-cache size, and latency in subsequent LLM layers. Training-free visual token pruning therefore aims to retain only a small subset of visual tokens while preserving the information essential for downstream tasks under a fixed retention ratio, e.g., keeping r%r\% of the visual sequence. One natural signal for token selection is the text-conditioned attention produced by LLM layers. However, raw text-conditioned attention is often positionally biased and concentrates around image borders or other systematic hotspots (Zou et al. 2026). Selecting tokens solely based on attention, as in FastV (Chen et al. 2024), therefore leads the retained set to cluster around biased regions while overlooking the true semantic evidence (Figure 1). As a result, subsequent training-free methods are largely motivated by the same objective: retain tokens that are informative for downstream understanding but would otherwise be discarded by attention-only selection. Figure 1: Qualitative comparison of FastV, HoloV, and RoRA under the same visual token budget. FastV is affected by positional attention sinks, while HoloV’s crop-wise allocation assigns tokens to high-variance but task-irrelevant regions. In contrast, RoRA uses AARs to allocate tokens across semantic core, complementary context, and fine-grained detail, thereby preserving the evidence needed to identify the jersey number and document year correctly. Existing remedies fall into two main branches. The first branch does not explicitly correct the attention bias. Instead, it performs image-level allocation that promotes non-redundancy and spatial coverage. DART (Wen et al. 2025) selects diversified tokens relative to pivots, while HoloV (Zou et al. 2026) partitions the image into crops and assigns per-crop quotas using attention and feature variance. These strategies improve coverage over naive top-k attention. However, these strategies remain agnostic to the semantic role of each retained token. In particular, HoloV can over-allocate budget to cluttered background crops: cluttered regions often exhibit high local variance and thus high crop scores, causing many retained tokens to be allocated to spectators or texture-rich edges rather than the queried object (Figure 1). The second branch explicitly corrects positional bias before ranking tokens by the debiased attention. D2Pruner (Zhang et al. 2026) debiases attention and then suppresses redundancy with a dense token–token similarity graph (e.g., 576×576576× 576). Debiasing produces a more reliable attention ranking, but global pairwise modeling is expensive and largely structure-agnostic: without explicitly modeling which regions are already represented by object evidence, the method must compare almost everything to everything to approximate coverage. Across both branches, a common limitation remains: tokens are selected by importance, diversity, or spatial coverage, but not by the role they should play under a fixed token budget. We argue that answering a multimodal query typically requires three distinct yet complementary types of visual evidence. First, a small set of high-confidence tokens must lock onto the queried object (core); without them, the model loses the primary referent. Second, once the core object is secured, additional tokens are more useful as complementary scene or secondary-object cues outside the support (context), rather than as near-duplicates around the same object. Third, a few fine-grained tokens are still needed for text, boundaries, and small parts (detail), which are easy to miss under coarse importance ranking or crop-level quotas. Because these roles complement rather than replace one another, allocating the budget with one homogeneous criterion, or with crop quotas that ignore object coverage, cannot deliberately preserve all three. We propose RoRA, a training-free framework for role-oriented regional evidence allocation. After obtaining debiased attention, we calibrate sample-specific relevance with a lightweight role-aware prompt and protect a semantic core. We then leverage the remaining positional structure as a useful prior: high-confidence anchors and their local neighborhoods on the 2D token grid form Attention-Anchored Regions (AARs). An AAR is not an exact object mask; it is a lightweight spatial proxy of object-related support already covered by the core. Conditioned on AARs, RoRA allocates the residual budget asymmetrically: explore complementary context mainly outside AARs, and preserve detail mainly inside AARs. Because coverage is made explicit by AARs, redundancy control reduces to comparing candidate context tokens against the protected core, instead of building a full pairwise redundancy matrix over the whole image. This yields both clearer evidence allocation and substantially lower selection overhead, as illustrated in Figure 1. We evaluate RoRA on several representative MLLMs including LLaVA-1.5, LLaVA-NeXT, Qwen2.5-VL, and Qwen3-VL. RoRA consistently achieves higher normalized average scores than prior training-free methods under equal retention budgets, while maintaining negligible selection overhead, e.g., 0.7ms per image. Our contributions are summarized as follows: • Role-aware budget allocation. We organize retained visual tokens into three complementary roles, and apply different selection criteria under a shared budget. • Object region prior. Beyond debiasing, we exploit a prompt-calibrated signal to identify object-centric regions and stabilize semantic-core protection. • Attention-anchored regions for structured pruning. AARs identify object-related spatial support, enabling outside-AAR context exploration and inside-AAR detail preservation with core-referenced redundancy checks, without requiring dense global similarity graphs. • Extensive experiments. Comparisons and ablations on several MLLMs verify RoRA’s accuracy–efficiency gains. Related Work Visual Token Compression in MLLMs. Visual token compression has become an important technique for reducing the high inference cost of MLLMs caused by long visual sequences. Existing approaches can be performed before visual features enter the language model or during LLM inference. Vision-side methods reduce tokens through pooling, query-based compression, or token merging before multimodal fusion (Li et al. 2023a; Yao et al. 2024; Bolya et al. 2022; Shang et al. 2025; Chen et al. 2025). However, these methods generally lack direct interaction with the textual instruction. In contrast, training-free LLM-side pruning exploits text-conditioned signals during inference to remove visual tokens after they have acquired multimodal semantics (Chen et al. 2024; Wen et al. 2025; Zou et al. 2026; Zhang et al. 2026). We focus on this setting and aim to improve visual token selection under a fixed retention budget. Training-Free Visual Token Selection. Existing training-free visual token selection methods mainly rely on three types of signals: (i) Attention-based methods. FastV and related approaches estimate token importance from text-conditioned attention (Chen et al. 2024; Zhang et al. 2025a). However, original text-conditioned attention can contain positional bias, causing important tokens to be confused with model-specific attention patterns. D2Pruner addresses this issue by introducing attention debiasing and combining it with structural diversity modeling (Zhang et al. 2026). (i) Redundancy-aware methods. Instead of relying only on importance scores, redundancy-aware methods select tokens according to feature similarity or duplication. DART and DivPrune reduce repeated visual representations by encouraging diverse token subsets (Wen et al. 2025; Alvar et al. 2025). However, these approaches mainly focus on redundancy reduction and do not explicitly distinguish the semantic roles of retained tokens. (i) Region-aware methods. HoloV improves spatial coverage by dividing images into regions and allocating tokens according to regional attention and feature variance (Zou et al. 2026). Nevertheless, fixed regional allocation does not explicitly model which object-related evidence has already been covered. Different from previous methods, RoRA treats retained tokens as different evidence roles and performs role-aware allocation among semantic core, complementary context, and fine-grained detail. Methods Figure 2: Overview of RoRA. RoRA decomposes the retained visual token budget into a protected semantic core, complementary context, and fine-grained details. Attention-Anchored Regions represent object-related spatial support already covered by high-confidence evidence, guiding outside-AAR context exploration and inside-AAR detail repair. In this section, we present RoRA, a training-free framework for role-oriented regional evidence allocation. As illustrated in Figure 2, RoRA first calibrates text-conditioned visual relevance and protects a high-confidence semantic core. It then constructs Attention-Anchored Regions (AARs) to represent object-related spatial support already covered by reliable evidence and allocates the remaining budget toward complementary context. Finally, a small-budget detail-repair stage restores fine-grained visual cues. RoRA Framework Overview Method GQA MMB MMB-CN MME POPE SQA VQAV2^V2 VQAText^Text VizWiz Avg. LLaVA-1.5-7B Upper Bound, 576 Tokens (100%) Vanilla 61.9 64.7 58.1 1862 85.9 69.5 78.4 58.2 50.0 100.0% LLaVA-1.5-7B Retain 192 Tokens (↓66.7%)( 66.7\%) ToMe (ICLR23) 54.3 60.5 - 1563 72.4 65.2 68.0 52.1 - 88.5% FastV (ECCV24) 52.7 61.2 57.0 1612 64.8 67.3 67.1 52.5 50.8 90.5% MustDrop (arXiv24) 58.2 62.3 55.8 1787 82.6 69.2 76.0 56.5 51.4 97.2% LLaVA-PruMerge (ICCV25) 54.3 59.6 52.9 1632 71.3 67.9 70.6 54.3 50.1 91.4% PDrop (CVPR25) 57.1 63.2 56.8 1766 82.3 68.8 75.1 56.1 51.1 96.7% VisionZip (CVPR25) 59.3 64.5 57.3 1767 86.4 68.9 76.8 57.3 51.6 98.1% SparseVLM (ICML25) 57.6 62.5 53.7 1721 83.6 69.1 75.6 56.1 50.5 96.1% DART (arxiv25) 60.3 64.1 56.3 1839.8 85.2 69.4 76.5 57.9 51.3 99.3% HoloV (NeurIPS25) 58.7 63.5 53.6 1782.9 86.3 68.5 74.8 56.1 51.8 97.5% D2Pruner (AAAI25) 60.9 64.2 55.7 1854.7 85.5 69.1 77.0 58.6 51.2 99.5% RoRA (Ours) 61.2 64.5 56.2 1861.0 85.5 69.1 77.1 58.7 51.4 99.8% LLaVA-1.5-7B Retain 128 Tokens (↓77.8%)( 77.8\%) ToMe (ICLR23) 52.4 53.3 - 1343 62.8 59.6 63.0 49.1 - 80.4% FastV (ECCV24) 49.6 56.1 56.4 1490 59.6 60.2 61.8 50.6 51.3 85.4% MustDrop (arXiv24) 56.9 61.1 55.2 1745 78.7 68.5 74.6 56.3 52.1 95.7% LLaVA-PruMerge (ICCV25) 53.3 58.1 51.7 1554 67.2 67.1 68.8 54.3 50.3 89.4% PDrop (CVPR25) 56.0 61.1 56.6 1644 82.3 68.3 72.9 55.1 51.0 94.9% VisionZip (CVPR25) 57.6 63.4 56.7 1768 84.7 68.8 75.6 56.8 52.0 97.2% SparseVLM (ICML25) 56.0 60.0 51.1 1696 80.5 67.1 73.8 54.9 51.4 93.8% DART (arxiv25) 58.7 63.8 54.9 1826 85.0 69.1 75.2 56.5 51.6 98.1% HoloV (NeurIPS25) 57.3 63.2 52.8 1757.8 84.0 67.8 73.5 56.1 52.0 96.3% D2Pruner (AAAI25) 60.0 63.4 54.3 1846.8 85.7 69.1 76.3 58.3 51.5 98.8% RoRA (Ours) 60.0 63.7 54.9 1849.6 86.1 69.2 76.2 58.4 51.8 99.1% LLaVA-1.5-7B Retain 64 Tokens (↓88.9%)( 88.9\%) ToMe (ICLR23) 48.6 43.7 - 1138 52.5 50.0 57.1 45.3 - 70.1% FastV (ECCV24) 46.1 48.0 52.7 1256 48.0 51.1 55.0 47.8 50.8 76.7% MustDrop (arXiv24) 53.1 60.0 53.1 1612 68.0 63.4 69.3 54.2 51.2 90.1% LLaVA-PruMerge (ICCV25) 51.9 55.3 49.1 1549 65.3 68.1 67.4 54.0 50.1 87.7% PDrop (CVPR25) 41.9 33.3 50.5 1092 55.9 68.6 69.2 45.9 50.7 77.5% VisionZip (CVPR25) 55.1 60.1 55.4 1690 77.0 69.0 72.4 55.5 52.9 94.5% SparseVLM (ICML25) 52.7 56.2 46.1 1505 75.1 62.2 68.2 51.8 50.1 87.3% DART (arxiv25) 56.1 61.6 53.4 1755 82.8 69.1 72.0 52.7 52.0 95.2% HoloV (NeurIPS25) 55.1 60.0 51.3 1696.8 80.5 68.9 70.8 54.6 53.2 94.1% D2Pruner (AAAI25) 56.9 62.4 52.4 1771.9 83.3 69.0 73.2 56.2 51.7 96.2% RoRA (Ours) 57.5 62.0 53.1 1773.2 84.2 69.0 73.8 56.3 51.6 96.5% LLaVA-Next-7B Upper Bound, 2880 Tokens (100%) Vanilla 64.2 67.4 60.6 1851 86.5 70.1 81.8 64.9 57.6 100.0% LLaVA-Next-7B Retain 320 Tokens (↓88.9%)( 88.9\%) FastV (ECCV24) 55.9 61.6 51.9 1661 71.7 62.8 71.9 55.7 53.1 88.0% LLaVA-PruMerge (ICCV25) 53.6 61.3 55.3 1534 60.8 66.4 69.7 50.6 54.0 85.6% PDrop (CVPR25) 56.4 63.4 56.2 1663 77.6 67.5 73.5 54.4 54.1 90.9% MustDrop (arXiv24) 57.3 62.8 55.1 1641 82.1 68.0 73.7 59.9 54.0 92.2% FasterVLM (ICCV25) 56.9 61.6 53.5 1701 83.6 66.5 74.0 56.5 52.6 91.1% SparseVLM (ICML25) 56.1 60.6 54.5 1533 82.4 66.1 71.5 58.4 52.0 89.7% DART (arxiv25) 61.1 66.3 44.1 1699.9 87.7 68.7 77.6 58.4 54.8 95.2% HoloV (NeurIPS25) 59.2 65.1 47.6 1750.2 87.4 66.5 76.3 55.7 55.3 94.9% D2Pruner (AAAI25) 61.5 66.1 41.7 1712.7 88.0 68.4 78.4 60.6 54.3 95.1% RoRA (Ours) 61.6 66.0 42.3 1713.6 88.1 68.9 78.4 60.7 54.9 95.5% Table 1: Performance comparisons on LLaVA-1.5-7B and LLaVA-Next-7B across several understanding benchmarks. RoRA is applied at an early LLM layer, where visual tokens have acquired text-conditioned information while pruning can still reduce the sequence to be processed by all subsequent layers (Chen et al. 2024; Zhang et al. 2026). Let ℓp _p denote the pruning layer. At this LLM layer, the hidden states of visual tokens can be defined as: ℋV=ii=1N.H_V= \h_i \_i=1^N. (1) RoRA retains a total budget of K for visual tokens. Unlike other methods with one homogeneous selection criterion for the entire visual sequence, it assigns the retained visual tokens with three distinct and non-interchangeable evidence roles: =∪˙∪˙,||=K,Kp+Kc+Kd=K. gatheredS=P\, ∪\,C\, ∪\,D,\\[-1.0pt] |S|=K, K_p+K_c+K_d=K. gathered (2) In this decomposition, S denotes the final set of retained visual tokens, while P encapsulates protected semantic-core tokens that preserve the primary evidence required for query answering. The complementary-context set C subsumes secondary objects, relational structures, and background context that fall outside the semantic core. Meanwhile, the detail-token set D recovers fine-grained spatial evidence, including text, small objects, boundaries, and local textures. The symbol ∪˙ ∪ denotes a disjoint union, so each retained token belongs to exactly one of the three sets. Here, Kp=||K_p=|P|, Kc=||K_c=|C|, and Kd=||K_d=|D| denote their respective token counts, while K is the total retained-token budget. Thus, ||=K|S|=K and Kp+Kc+Kd=K_p+K_c+K_d=K. Object-Calibrated Attention and Core Protection Original text-conditioned attention provides an initial estimate of the correlation between the current text instruction and visual content. However, such methods are vulnerable to positional bias, which systematically favors visual tokens at specific spatial positions. Following the debiasing principle of D2Pruner (Zhang et al. 2026), we calibrate the instance-level attention with an offline positional-attention prior. RoRA further incorporates a soft, prompt-calibrated object prior that smoothly steers attention toward regions more likely to contain object-centric evidence. In practice, we sample 1,000 images from the GQA (Hudson and Manning 2019) training split as an unlabeled calibration subset, using only images without questions, answers, or category annotations. For all calibration images, we use the same prompt, “Please describe the image.”, to estimate the model’s positional bias under generic image-description scenarios and obtain a general positional prior. We then repeat the same procedure with the object-inspection prompt, “List all visible objects in the image.”, to derive relatively weak object-aware prior scores for tokens at each visual position. This prior is intentionally designed as a subtle regularization term, which is motivated by the fact that object locations naturally vary across images and an excessively strong prior could compromise generalizability. Yet, a gentle prior suffices to encourage the model to focus on regions with higher semantic likelihood. Specifically, let aia_i denote the attention assigned to the i-th visual token, bib_i the positional-bias prior, and oio_i the object-aware prior. The calibrated attention score is defined as: ri r_i =Norm(aibi+ϵ), =Norm\! ( a_ib_i+ε ), (3) pi p_i =Norm(ri+λobjoi). =Norm\! (r_i+ _objo_i ). Here, ϵε is a numerical stabilizer, Norm(⋅)Norm(·) denotes the token-wise score-normalization operation, and λobj _obj controls the strength of object-aware calibration. Because the object prior is only a weak positional tendency rather than an image-specific object mask, λobj _obj is intentionally set to a modest value to ensure that sample-specific attention remains the dominant signal. We then select the KpK_p tokens with the highest pip_i scores to form the protected semantic-core set: =TopKKp(pii=1N).P=TopK_K_p\! (\p_i\_i=1^N ). (4) These tokens constitute the semantic backbone of the compressed visual sequence. They are excluded from the candidate pools of subsequent allocation stages and preserved in the final retained set. Attention-Anchored Regional Allocation The protected semantic core tokens preserve the most reliable semantic evidence. Consequently, selecting the remaining tokens solely by calibrated attention risks over-allocating the token budget to regions already sufficiently covered. RoRA therefore employs high-confidence tokens as spatial anchors, explicitly modeling the coverage of core evidence to ensure broader context diversity We select MaM_a region anchors according to calibrated attention and expand their local neighborhoods on the two-dimensional visual token grid: =TopKMa(pii=1N),ℛAAR=⋃a∈ra(a). gatheredA=TopK_M_a\! (\p_i\_i=1^N ),\\[-1.0pt] R_AAR= _a N_r_a(a). gathered (5) For each anchor, ra(a)N_r_a(a) represents the local grid region centered at a with radius rar_a, and ℛAARR_AAR denotes the union of all anchor neighborhoods. Rather than enforcing a fixed token quota for each region, Attention-Anchored Regions (AAR) dynamically adjusts the scores of complementary-context candidates. Once primary object evidence is covered by P, candidates outside AAR receive a small exploration boost, encouraging the remaining budget to cover secondary objects, spatial relations, and scene context. Based on AAR, we define the complementary-context score of the i-th token as ci=Norm(pi+λctx[i∉ℛAAR]).c_i=Norm\! (p_i+ _ctxI[\,i _AAR\,] ). (6) The coefficient λctx _ctx controls the exploration strength for candidates outside AAR, and [⋅]I[·] denotes the indicator function. When a token falls outside the AAR, its context score receives a small bonus, preventing the remaining token budget from over-concentrating in salient regions that are already adequately covered. To further reduce redundancy, we use semantic similarity as a lightweight redundancy filter. If the semantic similarity between a candidate token and one already retained token exceeds the threshold τsem _sem, this candidate will be skipped, thereby preventing repeated visual evidence from consuming the budget. The resulting complementary-context set is: =Selectτsem,Kc(ci∣i∉).C=Select_ _sem,K_c\! (\c_i i \ ). (7) The operator SelectτsemSelect_ _sem ranks candidate tokens by descending cic_i and skips any candidate whose cosine similarity to the currently retained tokens exceeds τsem _sem, until KcK_c complementary-context tokens have been selected. AAR-Guided Detail Repair Figure 3: Component-wise ablation on LLaVA-1.5-7B. Left: effects of semantic-core protection and complementary-context allocation on eight RefCOCO-family localization splits at 75%75\% and 90%90\% pruning. Right: ablation about detail tokens at K=128K=128 on TextVQA and VizWiz. While the semantic core and complementary context preserve primary objects, relations, and scene information, fine-grained components—such as text, small objects, boundaries, and local textures—occupy few visual patches and often exhibit low global relevance. We therefore employ a small-budget detail-repair stage to recover this evidence. Let a^i a_i, m^i m_i, and ℓ^i _i denote the attention response, feature magnitude, and local contrast of the i-th visual token, respectively: a^i=Norm(ai),m^i=Norm(‖i‖2),Δij=1−cos(i,j),ℓ^i=Norm(1|8(i)|∑j∈8(i)Δij). gathered a_i=Norm(a_i), m_i=Norm\! (\|h_i\|_2 ),\\[-1.0pt] _ij=1- (h_i,h_j),\\[-1.0pt] _i=Norm\! ( 1|N_8(i)|Σ _j _8(i) _ij ). gathered (8) For each visual token, 8(i)N_8(i) denotes its eight neighborhood on the two-dimensional visual token grid. The attention response represents the attention received by the current visual token; the feature magnitude represents the strength of its hidden-state representation; and the local contrast measures the average feature difference between the current token and its neighboring tokens. Based on these quantities, we define an AAR indicator ηi _i and the detail score of the i-th visual token as ηi=[i∈ℛAAR],di=ωaa^i+ωmm^i+ωℓℓ^i+λdetηi. gathered _i=I[\,i _AAR\,],\\[-1.0pt] d_i= _a a_i+ _m m_i+ _ _i+ _det _i. gathered (9) The coefficients ωa _a, ωm _m, and ωℓ _ control the contributions of attention response, feature magnitude, and local contrast, respectively, while λdet _det provides an additional bonus to detail tokens within AAR. Importantly, AAR is not a hard constraint but a soft preference. We select the KdK_d highest-scoring tokens that have not already been assigned to the semantic-core or complementary-context sets: =TopKKd(di∣i∉∪).D=TopK_K_d\! (\d_i i \ ). (10) Finally, the protected semantic core P, complementary context C, and detail tokens D jointly form the retained visual token set S. Experiments Experimental Settings Method Qwen2.5-VL-7B Qwen3-VL-8B AI2D MME T-VQA NB RW TC SQA Avg. BLK Ill VR POPE T-VQA CV MUIR Avg. Upper Bound, All Tokens (100%) Upper Bound, All Tokens (100%) Vanilla 74.7 2246.6 76.4 24.0 27.1 53.5 73.0 100.0% 69.1 53.7 25.6 89.1 80.9 85.9 50.9 100.0% Retain 25% Tokens in Average (↓75%)( 75\%) Retain 25% Tokens in Average (↓75%)( 75\%) HoloV 71.0 1915.2 64.7 21.8 21.7 46.2 74.8 89.2% 58.3 51.0 26.5 85.0 57.1 81.0 48.4 91.1% FastV 71.3 1915.3 73.3 23.2 30.6 65.6 74.1 101.4% 58.3 53.1 24.0 85.0 69.5 81.8 48.4 92.6% DART 69.3 1900.7 66.9 25.4 27.8 54.4 73.2 96.4% 55.0 51.9 25.9 85.5 69.0 81.6 49.7 93.0% DivPrune 21.5 2014.4 69.8 25.1 26.1 58.4 73.4 88.6% 59.1 49.2 24.5 87.8 72.7 81.3 50.5 93.6% D2Pruner 73.3 2080.3 73.6 28.4 31.7 60.9 75.9 105.6% 58.3 49.2 24.6 85.8 70.4 82.0 49.7 92.6% RoRA 73.9 2106.0 73.8 29.4 32.4 61.3 76.0 106.9% 62.5 53.1 29.5 86.6 72.4 82.8 50.4 98.0% Retain 10% Tokens in Average (↓90%)( 90\%) Retain 10% Tokens in Average (↓90%)( 90\%) HoloV 64.6 1760.3 38.7 14.0 17.3 37.5 75.2 72.9% 45.8 49.4 25.8 75.0 40.7 73.4 45.0 81.0% FastV 65.8 1288.9 65.0 11.6 23.5 66.1 70.7 83.6% 45.8 48.2 22.9 75.0 46.3 74.0 45.0 80.2% DART 64.5 1610.3 56.1 17.5 24.3 44.7 72.3 82.4% 45.0 49.8 23.8 77.8 49.0 72.4 47.9 82.4% DivPrune 19.9 1592.4 56.9 17.7 19.3 44.3 69.2 70.6% 48.3 46.6 26.0 83.3 59.2 74.4 48.5 86.6% D2Pruner 72.5 1778.5 69.5 18.3 27.1 63.9 73.6 94.8% 45.0 49.2 23.0 78.9 49.8 76.6 47.1 82.6% RoRA 72.5 1780.6 69.6 20.8 27.7 64.3 73.9 96.7% 49.2 50.8 28.6 79.4 54.2 77.1 47.0 87.9% Table 2: Performance comparisons on Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct. Models and baselines. To evaluate RoRA across different visual-encoding schemes and visual token sequence lengths, we conduct experiments on four representative MLLM backbones: LLaVA-1.5-7B, LLaVA-NeXT-7B, Qwen2.5-VL-7B-Instruct, and Qwen3-VL-8B-Instruct (Liu et al. 2024a, b; Bai et al. 2025b, a). We compare against a broad set of visual token compression baselines, including ToMe, FastV, MustDrop, LLaVA-PruMerge, PDrop, FiCoCo-V, HiRED, FitPrune, VisionZip, SparseVLM, FasterVLM, DART, DivPrune, HoloV, and D2Pruner (Bolya et al. 2022; Chen et al. 2024; Liu et al. 2024c; Shang et al. 2025; Xing et al. 2025; Han et al. 2026; Arif et al. 2025; Ye et al. 2025; Yang et al. 2025; Zhang et al. 2024, 2025a; Wen et al. 2025; Alvar et al. 2025; Zou et al. 2026; Zhang et al. 2026). Benchmarks. For the LLaVA series, we evaluate on GQA, MMBench, MMBench-CN, MME, POPE, ScienceQA, VQAv2, TextVQA, and VizWiz (Hudson and Manning 2019; Liu et al. 2024d; Fu et al. 2026; Li et al. 2023b; Lu et al. 2022; Goyal et al. 2017; Singh et al. 2019; Gurari et al. 2018). TextVQA and VizWiz are more sensitive to fine-grained visual evidence, whereas benchmarks such as GQA and MME place greater emphasis on global scene understanding and semantic reasoning (Singh et al. 2019; Gurari et al. 2018; Hudson and Manning 2019; Fu et al. 2026). For Qwen2.5-VL-7B-Instruct, we use AI2D, MME, TextVQA, NaturalBench, MME-RealWorld-CN, TextCaps, and ScienceQA-IMG, covering capabilities such as diagram understanding, image-caption generation, fine-grained perception, and multimodal reasoning (Kembhavi et al. 2016; Fu et al. 2026; Singh et al. 2019; Li et al. 2024; Zhang et al. 2025b; Sidorov et al. 2020; Lu et al. 2022). For the ablation, we additionally evaluate on the RefCOCO family (Kazemzadeh et al. 2014; Mao et al. 2016). For Qwen3-VL-8B-Instruct, we further evaluate on BLINK-counting, IllusionVQA, VisRes, POPE, TextVQA, CV-Bench, and MUIRBench to assess pruning robustness in counting, hallucination-sensitive evaluation, visual resolution understanding, and complex multimodal reasoning (Fu et al. 2024; Shahgir et al. 2024; Törtei et al. 2026; Li et al. 2023b; Singh et al. 2019; Tong et al. 2024; Wang et al. 2025). Unless otherwise specified, we normalize each benchmark score by the corresponding unpruned result and average the normalized scores across benchmarks. Implementation details. RoRA uses one fixed configuration for each backbone–pruning-ratio pair. Once selected, this configuration is shared across all tasks and frozen throughout evaluation, without benchmark-specific tuning. For all competing baselines, we follow the configurations provided by their official implementations. Complete hyperparameter settings are reported in the appendix. Main Results Methods Time Select Latency Acc. Time Select Latency Acc. Upper Bound, 576 Tokens 7:35 – 0.05s 100% 7:35 – 0.05s 100% LLaVA-1.5-7B Retain 192 Tokens (↓ 66.70%) Retain 58 Tokens (↓ 90%) FastV (ECCV24) 6:36 0.1ms 0.04s 75.4% 6:44 0.1ms 0.05s 66.8% D2Pruner (AAAI25) 19:58 71.7ms 0.13s 96.8% 19:53 68.1ms 0.13s 87.7% Faster (ICCV25) 5:51 0.1ms 0.04s 99.2% 5:48 0.1ms 0.04s 92.3% MustDrop (arXiv24) 7:24 0.2ms 0.05s 95.4% 9:29 0.2ms 0.06s 86.6% DART (arXiv25) 6:17 3.2ms 0.04s 97.5% 6:09 2.0ms 0.04s 90.6% HoloV (NeurIPS25) 5:48 1.0ms 0.04s 98.9% 5:38 2.4ms 0.04s 92.4% RoRA (Ours) 5:43 0.7ms 0.04s 98.4% 5:22 0.6ms 0.04s 92.5% Table 3: Real inference comparison on LLaVA-1.5 and POPE at 66.7% and 90% pruning ratios, measured on a single NVIDIA H800 GPU with batch size 1. All competing methods are evaluated under the same settings using their official open-source implementations. Results on LLaVA models. The upper block of Table 1 reports the results on LLaVA-1.5-7B. With visual token budgets of K=192K=192, 128128, and 6464, RoRA retains 99.8%99.8\%, 99.1%99.1\%, and 96.5%96.5\% of the unpruned model’s normalized average performance, respectively, achieving the best result among all compared methods at every matched budget. Moreover, the relative advantage of RoRA increases as pruning becomes more aggressive, indicating that role-aware regional allocation preserves visual evidence more effectively under severely constrained token budgets. The lower block of Table 1 further presents results on LLaVA-NeXT-7B, whose any-resolution visual encoder produces a dynamic number of visual tokens according to the input resolution. When retaining a fixed budget of K=320K=320 visual tokens, RoRA achieves a normalized average performance of 95.5%95.5\%, outperforming all compared baselines. Results on Qwen models. Table 2 reports results on the Qwen series. On Qwen2.5-VL-7B, RoRA achieves normalized averages of 106.9%106.9\% and 96.7%96.7\% at 75%75\% and 90%90\% pruning, respectively, outperforming all compared baselines. On Qwen3-VL-8B, it reaches 98.0%98.0\% and 87.9%87.9\%, surpassing D2Pruner by 5.45.4 and 5.35.3 points, respectively. These results demonstrate that RoRA generalizes consistently across newer dynamic-resolution MLLMs. Ablation Studies We evaluate the semantic core and complementary context on localization benchmarks, as these components are designed to preserve the queried object and its surrounding spatial evidence. The detail stage is evaluated on TextVQA and VizWiz, which are more sensitive to OCR cues, small objects, and localized visual details. Localization results are normalized by our unpruned model under the same prompting, parsing, and IoU-based evaluation protocol, while the detail results are normalized by the corresponding vanilla upper bounds. Semantic core and complementary context. Figure 3 reports localization results on eight RefCOCO-family splits (Kazemzadeh et al. 2014; Mao et al. 2016). All variants retain the same number of visual tokens. At K=144K=144, the normalized average increases from 35.44%35.44\% for FastV (Chen et al. 2024) to 74.44%74.44\% after introducing the protected semantic core. At the more aggressive K=58K=58 setting, the average similarly increases from 9.57%9.57\% to 17.36%17.36\%. Core outperforms FastV on every split at both budgets, confirming that protecting calibrated object-related evidence is essential for localization. Adding AAR-guided complementary context further improves the average from 74.44%74.44\% to 75.03%75.03\% at K=144K=144, and from 17.36%17.36\% to 17.47%17.47\% at K=58K=58. It improves seven of eight splits at K=144K=144 and introduces no degradation at K=58K=58. Although the gain is smaller than that of the semantic core, it shows that allocating part of the fixed budget to previously uncovered spatial evidence provides additional localization support. AAR-guided detail repair. We evaluate detail repair at K=128K=128 on TextVQA (Singh et al. 2019) and VizWiz (Gurari et al. 2018). Without detail tokens, the normalized scores are 100.10%100.10\% on TextVQA and 103.12%103.12\% on VizWiz, with an average of 101.61%101.61\%. Adding the standard detail configuration increases the average to 101.80%101.80\%, while allocating 1818 detail tokens further improves it to 101.94%101.94\%. The monotonic gains on both benchmarks indicate that detail repair recovers fine-grained evidence not fully captured by the semantic-core and context stages. Its improvement is modest, which is consistent with its intended role as a small-budget complementary module. Efficiency Analysis Methods Prefill ↓ Total ↓ FLOPs ↓ KV Cache ↓ Acc. ↑ (ms/sample) (ms/sample) (T) (MB) (%) Upper Bound, 2928 Tokens 132.8 (1.00×) 133.2 (1.00×) 16.9 (1.00×) 1512 (1.00×) 88.73 LLaVA-Next-7B D2Pruner (AAAI26) (33.4%) 126.3 (1.05×) 125.9 (1.06×) 6.0 (2.82×) 526 (2.87×) 88.82 RoRA (Ours, 33.4%) 108.1 (1.23×) 106.8 (1.25×) 6.0 (2.82×) 526 (2.87×) 88.86 D2Pruner (AAAI26) (11.2%) 69.0 (1.92×) 68.9 (1.93×) 2.8 (6.10×) 198 (7.63×) 88.00 RoRA (Ours, 11.2%) 66.5 (2.00×) 65.7 (2.03×) 2.8 (6.10×) 198 (7.63×) 88.01 Table 4: Accuracy and speed comparison on LLaVA-Next-7B and POPE under 66.7% and 90% pruning ratios. We report POPE accuracy, actual prefilling time, total runtime, theoretical FLOPs, and KV cache. Runtime is measured on a single RTX PRO 6000 GPU with batch size 1 and max new tokens 1. We evaluate LLaVA-1.5-7B on full POPE (9,000 yes-or-no questions) using one NVIDIA H800 GPU, batch size 1, and max new tokens 1. This prefill-dominated protocol measures visual-token selection and LLM prefilling overhead, with K=192K=192 and K=58K=58 reported in Table 3. Table Efficiency Analysis further evaluates LLaVA-NeXT-7B on one RTX PRO 6000 GPU; due to its much longer visual sequences, selector overhead accounts for a smaller share of total runtime. At K=192K=192, RoRA completes the full evaluation in 5min43s and achieves an average per-sample latency of 0.038s0.038\,s, both of which are the lowest measured values among the compared methods. Its selector overhead is only 0.704ms0.704\,ms, while it retains 98.4%98.4\% of the unpruned model’s POPE performance. Compared with the unpruned runtime of 7 min 35 s, this corresponds to a 24.6% reduction in end-to-end inference time. Under the more aggressive K=58K=58 setting, RoRA further reduces the total runtime to 5 min 22 s. It achieves not only the lowest inference latency among all compared methods, but also the best accuracy among the compressed models, retaining 92.5% of the unpruned performance. RoRA is efficient because token similarity is used only as a lightweight redundancy filter in complementary-context selection: residual candidates highly similar to retained evidence are skipped, instead of injecting pairwise similarity into the primary importance score. Let Nt<N_t<N be the residual candidates after fixing the semantic core, and C the number of HoloV crops. RoRA has complexity (N+Nt2)O(N+N_t^2), which is much lighter in practice than D2Pruner’s (N2)O(N^2) full-graph construction and comparable to HoloV’s crop-wise (N2/C)O(N^2/C). Conclusion In this work, we presented RoRA, a novel training-free visual token pruning framework that views context reduction as role-aware regional evidence allocation. By organizing visual tokens into a protected semantic core, complementary context, and fine-grained detail evidence, RoRA explicitly improves coverage of retained tokens through lightweight Attention-Anchored Regions (AARs). This design enables asymmetric budget allocation, expanding complementary evidence outside attended regions while recovering local details inside them, without constructing costly dense pairwise redundancy graphs. Extensive evaluations across the multiple representative MLLM architectures, including LLaVA and Qwen-VL families show that RoRA achieves a superior accuracy–efficiency profile, consistently outperforming existing training-free baselines across different models even at aggressive pruning ratios. References Alvar et al. (2025) Alvar, S. R.; Singh, G.; Akbari, M.; and Zhang, Y. 2025. Divprune: Diversity-based visual token pruning for large multimodal models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 9392–9401. Arif et al. (2025) Arif, K. H. I.; Yoon, J.; Nikolopoulos, D. S.; Vandierendonck, H.; John, D.; and Ji, B. 2025. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 1773–1781. Bai et al. (2025a) Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng, L.; Ding, W.; Gao, C.; Ge, C.; et al. 2025a. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Bai et al. (2025b) Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Cheng, Z.; Zhang, H.; Yang, Z.; Xu, H.; and Lin, J. 2025b. Qwen2.5-VL Technical Report. arXiv:2502.13923. Bolya et al. (2022) Bolya, D.; Fu, C.-Y.; Dai, X.; Zhang, P.; Feichtenhofer, C.; and Hoffman, J. 2022. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461. Chen et al. (2024) Chen, L.; Zhao, H.; Liu, T.; Bai, S.; Lin, J.; Zhou, C.; and Chang, B. 2024. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, 19–35. Springer. Chen et al. (2025) Chen, Y.; Wen, Z.; Wu, Y.; Liu, X.; Chen, S.; Ma, J.; Li, W.; He, C.; and Zhang, L. 2025. IPCV: Information-Preserving Compression for MLLM Visual Encoders. arXiv preprint arXiv:2512.18747. Fu et al. (2026) Fu, C.; Chen, P.; Shen, Y.; Qin, Y.; Zhang, M.; Lin, X.; Yang, J.; Zheng, X.; Li, K.; Sun, X.; et al. 2026. Mme: A comprehensive evaluation benchmark for multimodal large language models. Advances in Neural Information Processing Systems, 38. Fu et al. (2024) Fu, X.; Hu, Y.; Li, B.; Feng, Y.; Wang, H.; Lin, X.; Roth, D.; Smith, N. A.; Ma, W.-C.; and Krishna, R. 2024. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, 148–166. Springer. Goyal et al. (2017) Goyal, Y.; Khot, T.; Summers-Stay, D.; Batra, D.; and Parikh, D. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, 6904–6913. Gurari et al. (2018) Gurari, D.; Li, Q.; Stangl, A. J.; Guo, A.; Lin, C.; Grauman, K.; Luo, J.; and Bigham, J. P. 2018. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3608–3617. Han et al. (2026) Han, Y.; Liu, X.; Zhang, Z.; Ding, P.; Chen, J.; Chen, H.; Wang, D.; Yan, Q.; and Huang, S. 2026. Filter, correlate, compress: Training-free token reduction for mllm acceleration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 4601–4609. Hudson and Manning (2019) Hudson, D. A.; and Manning, C. D. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 6700–6709. Kazemzadeh et al. (2014) Kazemzadeh, S.; Ordonez, V.; Matten, M.; and Berg, T. 2014. ReferItGame: Referring to Objects in Photographs of Natural Scenes. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, 787–798. Association for Computational Linguistics. Kembhavi et al. (2016) Kembhavi, A.; Salvato, M.; Kolve, E.; Seo, M.; Hajishirzi, H.; and Farhadi, A. 2016. A diagram is worth a dozen images. In European conference on computer vision, 235–251. Springer. Li et al. (2024) Li, B.; Lin, Z.; Peng, W.; Nyandwi, J. d. D.; Jiang, D.; Ma, Z.; Khanuja, S.; Krishna, R.; Neubig, G.; and Ramanan, D. 2024. Naturalbench: Evaluating vision-language models on natural adversarial samples. Advances in Neural Information Processing Systems, 37: 17044–17068. Li et al. (2023a) Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023a. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, 19730–19742. PMLR. Li et al. (2023b) Li, Y.; Du, Y.; Zhou, K.; Wang, J.; Zhao, X.; and Wen, J.-R. 2023b. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 conference on empirical methods in natural language processing, 292–305. Liu et al. (2024a) Liu, H.; Li, C.; Li, Y.; and Lee, Y. J. 2024a. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 26296–26306. Liu et al. (2024b) Liu, H.; Li, C.; Li, Y.; Li, B.; Zhang, Y.; Shen, S.; and Lee, Y. J. 2024b. Llavanext: Improved reasoning, ocr, and world knowledge. Liu et al. (2024c) Liu, T.; Shi, L.; Hong, R.; Hu, Y.; Yin, Q.; and Zhang, L. 2024c. Multi-stage vision token dropping: Towards efficient multimodal large language model. arXiv preprint arXiv:2411.10803. Liu et al. (2024d) Liu, Y.; Duan, H.; Zhang, Y.; Li, B.; Zhang, S.; Zhao, W.; Yuan, Y.; Wang, J.; He, C.; Liu, Z.; et al. 2024d. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, 216–233. Springer. Lu et al. (2022) Lu, P.; Mishra, S.; Xia, T.; Qiu, L.; Chang, K.-W.; Zhu, S.-C.; Tafjord, O.; Clark, P.; and Kalyan, A. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in neural information processing systems, 35: 2507–2521. Mao et al. (2016) Mao, J.; Huang, J.; Toshev, A.; Camburu, O.; Yuille, A. L.; and Murphy, K. 2016. Generation and Comprehension of Unambiguous Object Descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 11–20. Shahgir et al. (2024) Shahgir, H. S.; Sayeed, K. S.; Bhattacharjee, A.; Ahmad, W. U.; Dong, Y.; and Shahriyar, R. 2024. Illusionvqa: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952. Shang et al. (2025) Shang, Y.; Cai, M.; Xu, B.; Lee, Y. J.; and Yan, Y. 2025. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 22857–22867. Sidorov et al. (2020) Sidorov, O.; Hu, R.; Rohrbach, M.; and Singh, A. 2020. Textcaps: a dataset for image captioning with reading comprehension. In European conference on computer vision, 742–758. Springer. Singh et al. (2019) Singh, A.; Natarajan, V.; Shah, M.; Jiang, Y.; Chen, X.; Batra, D.; Parikh, D.; and Rohrbach, M. 2019. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8317–8326. Tong et al. (2024) Tong, S.; Brown, E.; Wu, P.; Woo, S.; Middepogu, M.; Akula, S. C.; Yang, J.; Yang, S.; Iyer, A.; Pan, X.; et al. 2024. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37: 87310–87356. Törtei et al. (2026) Törtei, B. M.; Dahou, Y.; Huynh, N. D.; Para, W. R.; Khac, P. H. L.; Singh, A.; Chaybouti, S.; and Narayan, S. 2026. VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 33185–33195. Wang et al. (2025) Wang, F.; Fu, X.; Huang, J. Y.; Li, Z.; Liu, Q.; Liu, X.; Ma, M. D.; Xu, N.; Zhou, W.; Zhang, K.; et al. 2025. Muirbench: A comprehensive benchmark for robust multi-image understanding. In International Conference on Learning Representations, volume 2025, 62624–62650. Wen et al. (2025) Wen, Z.; Gao, Y.; Wang, S.; Zhang, J.; Zhang, Q.; Li, W.; He, C.; and Zhang, L. 2025. Stop Looking for “Important Tokens” in Multimodal Language Models: Duplication Matters More. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 9972–9991. Xing et al. (2025) Xing, L.; Huang, Q.; Dong, X.; Lu, J.; Zhang, P.; Zang, Y.; Cao, Y.; He, C.; Wang, J.; Wu, F.; et al. 2025. Conical visual concentration for efficient large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14593–14603. Yang et al. (2025) Yang, S.; Chen, Y.; Tian, Z.; Wang, C.; Li, J.; Yu, B.; and Jia, J. 2025. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19792–19802. Yao et al. (2024) Yao, L.; Li, L.; Ren, S.; Wang, L.; Liu, Y.; Sun, X.; and Hou, L. 2024. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985. Ye et al. (2025) Ye, W.; Wu, Q.; Lin, W.; and Zhou, Y. 2025. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 22128–22136. Zhang et al. (2026) Zhang, E.; Yu, F.; Wu, A.; Wen, Z.; Yan, K.; Ding, S.; Qi, B.; and Zhang, L. 2026. D2Pruner: Debiased Importance and Structural Diversity for MLLM Token Pruning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 12412–12420. Zhang et al. (2025a) Zhang, Q.; Cheng, A.; Lu, M.; Zhang, R.; Zhuo, Z.; Cao, J.; Guo, S.; She, Q.; and Zhang, S. 2025a. Beyond text-visual attention: Exploiting visual cues for effective token pruning in vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 20857–20867. Zhang et al. (2024) Zhang, Y.; Fan, C.-K.; Ma, J.; Zheng, W.; Huang, T.; Cheng, K.; Gudovskiy, D.; Okuno, T.; Nakata, Y.; Keutzer, K.; et al. 2024. Sparsevlm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417. Zhang et al. (2025b) Zhang, Y.; Zhang, H.; Tian, H.; Fu, C.; Zhang, S.; Wu, J.; Li, F.; Wang, K.; Wen, Q.; Zhang, Z.; et al. 2025b. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? In International Conference on Learning Representations, volume 2025, 89655–89701. Zou et al. (2026) Zou, X.; Lu, D.; Wang, Y.; Yan, Y.; Lyu, Y.; Zheng, X.; Zhang, L.; and Hu, X. 2026. Don’t Just Chase “Highlighted Tokens” in MLLMs: Revisiting Visual Holistic Context Retention. Advances in Neural Information Processing Systems, 38: 39800–39832.