Visual Tokens

Fine-grained visual perception in MLLMs is commonly improved by raising the resolution, but the added visual tokens inflate vision-encoding and language-model prefilling costs. We show that the two operations underlying fine-grained perception, localizing the region of interest (RoI) and recognizing its content, have different resolution requirements. In a controlled diagnostic, localization tolerates roughly 3 to 4 times stronger token compression than recognition, which motivates localizing from a coarse view and concentrating resolution on the selected evidence. Decoding coordinates with the MLLM can be trained end-to-end from answers, but costs a full model pass per query and depends on grounding ability. A lightweight proposal network distilled from the model's attention is fast, but inherits the noise of its attention targets. The RoI from the proposal network reaches the answer through a discrete region choice, so its faithfulness to the answer cannot supervise the network. We therefore optimize the proposal network with region-level reinforcement learning, which we call Vision-RL2. It treats coherent regions as actions, and a frozen MLLM reader scores each one by how its removal changes the answer likelihood. Complementary subtractive and additive objectives suppress distracting proposals and recover missing evidence, updating only the predictor without region annotations, response sampling, or reasoning trajectories. The refined proposal further enables a sparse encoding that magnifies evidence and excludes background tokens. Across six fine-grained benchmarks and four MLLM backbones, Vision-RL2 improves accuracy over the base model at every token budget and surpasses its largest-budget accuracy with about 4 times fewer visual tokens. Code is available at https://github.com/YuHengsss/VisionRL2 .

Autoregressive image generation has emerged as a paradigm for multimodal AI systems due to its compatibility with transformer-based LLM serving infrastructures. However, generating thousands of visual tokens per request makes decoding increasingly bottlenecked by KV cache accesses during attention computation. Sparse attention is particularly attractive for this workload because many visual generation applications tolerate moderate quality degradation in exchange for improved performance and efficiency. While sparse attention has been extensively explored for text-based LLM inference, it remains unclear whether its sparsity assumptions generalize effectively to autoregressive image generation. We present the first systematic characterization of attention sparsity in autoregressive image generation across diverse workloads and representative open-source models. Our analysis reveals several distinguishing properties, including a pronounced prefill-decode asymmetry, strong attention concentration on prompt and local tokens, and a unique diagonal attention sparsity pattern arising from the spatial locality of visual tokens. Motivated by these observations, we propose a diagonal-aware sparse attention mechanism that selectively skips KV entries along the diagonal attention direction within a recent window. Implemented on top of a GPU-based serving system using FlexGen, FlashAttention-2, and custom kernels, our approach achieves up to 3.1x throughput and 1.19x latency improvements with less than 2% quality degradation compared to dense inference.

Recent advancements have successfully adapted autoregressive language models to process multimodal signals, such as images and actions. Since raw action signals are continuous, effective tokenization is essential to map high-dimensional inputs into compact discrete tokens for autoregressive processing. However, existing discrete action tokenizers often suffer from high reconstruction loss, failing to preserve the fine-grained dynamics required for precise control. This "discretization bottleneck" significantly limits the performance ceiling of downstream Vision-Language-Action (VLA) models. To address this, we propose ${M}^2$Tok, a Multi-head Multi-codebook Action Tokenizer designed to minimize reconstruction error and enhance policy performance. Our approach introduces two key structural innovations: (1) we decompose the latent action features into multiple heads, enabling the model to implicitly align specific heads with distinct action dimensions; (2) we assign independent codebooks to each head for quantization. By leveraging the combinatorial nature of multiple codebooks, we significantly expand the representational expressivity of the tokenizer, leading to substantially lower reconstruction loss compared to previous methods. We evaluate the ${M}^2$Tok-based VLA on the RoboTwin, Simpler-Env, and 3 zero-shot real-world tasks. Experimental results demonstrate our method not only achieves superior reconstruction fidelity but also significantly boosts the success rate of VLA models. Comprehensive ablation studies further confirm the effectiveness of the multi-head and multi-codebook mechanisms. Code is available at https://github.com/cpaaax/M2Tok.

Increasing image resolution produces ever-longer visual-token sequences in vision-language models (VLMs), substantially raising their inference cost. To reduce this overhead without retraining, existing methods select compact token subsets that prioritize query relevance, visual coverage, or a fixed trade-off between them. The appropriate balance, however, varies across queries and token budgets: localized questions favor relevance, whereas holistic questions demand broader visual coverage. We introduce StackTok, a training-free selector that treats query relevance as the objective and visual coverage as budget-calibrated support. StackTok builds a size-indexed coverage reference from a coverage-only greedy sequence and adjusts its support target using query--vision affinity entropy. A reference-gated interleaved selection policy then switches between relevance- and coverage-oriented additions according to the current subset's support deficit. For high-resolution inputs, StackTok allocates one shared token budget across crops according to the combined marginal gain of locally nominated tokens. Evaluated with five VLMs over ten distinct image-understanding benchmarks, StackTok ranks first among training-free selectors in every tested model--budget setting. On high-resolution LLaVA-NeXT-7B, it retains 95.26% of full-token performance with only 160 of 2{,}880 (5.6%) visual tokens.

3D Gaussian language fields provide an explicit, spatially grounded representation for 3D visual question answering (VQA), but their dense semantic features can require tens of thousands of embeddings per scene, resulting in substantial storage, memory, and inference costs. We investigate how much of this representation is actually necessary for downstream reasoning. Starting from a full embedding representation, we systematically sparsify its semantic embeddings, including the previously underexplored regime below a single image-equivalent block down to 8 visual tokens. We compare random, geometric, semantic, and joint spatial-semantic selection strategies and introduce an object-based sparsification method that distributes the token budget across detected object instances while retaining background context. Experiments on ScanQA and MV-ScanQA reveal substantial redundancy in dense Gaussian language fields. Strong VQA performance is retained with only a few hundred semantic embeddings, corresponding to less than 1% of the original representation. Object-based selection performs well relative to others, with only modest observed changes down to 256 tokens. At this budget, SparseTalk retains 0.80% of SplatTalk's 32,076-token inference input and 0.332% of the mean 77,207-Gaussian dense field, increasing inference throughput while reducing decoded-feature memory 125-fold.

Multimodal large language models (MLLMs) require substantial computation to process numerous visual tokens across all transformer layers. Most methods for efficient MLLM inference exploit horizontal redundancy by compressing visual tokens. Beyond token reduction, recent studies exploit vertical redundancy through early exit or fixed-layer skipping. However, we find that the extent and distribution of this redundancy vary across inputs and differ between self-attention and MLP modules. Motivated by these observations, we propose AdaVSkip, which equips each layer with two lightweight routers that independently determine whether visual tokens pass through by or skip the self-attention and MLP modules. These decisions collectively define an input-specific visual-computation path, but their discrete and non-differentiable nature makes learning effective paths challenging. To address this challenge, we develop a progressive two-stage training framework that updates only the routers while keeping the backbone frozen. Stage I establishes an initial routing policy through supervised training with input-specific targets derived from module-wise necessity scores. To further align the routing policy with task performance, Stage II uses reinforcement learning to optimize routing decisions with direct feedback from generated answers. It combines an answer correctness reward with a skip-consistency reward that discourages excessive retention of visual-token computation. Across three MLLM backbones, AdaVSkip maintains strong task performance with substantially less computation. On LLaVA-NeXT-7B, AdaVSkip reduces FLOPs by 53.2% while preserving the original model's average performance. Combining it with visual token compression increases this reduction to 91.2%, while retaining 97.2% of the original performance on average.

Video tokenizers have emerged as a cornerstone of modern video modeling, underpinning progress in compression, reconstruction and generation by mapping high-dimensional visual signals into compact latent spaces. However, despite this progress, current tokenization paradigms largely remain within the 2D visual domain, treating videos as image sequences rather than observations of an underlying dynamic 3D world. Consequently, the learned tokens inherit this observation-centric bias, limiting their capacity to compactly represent real-world 4D scenes. To mitigate this issue, we propose VideoTok4D, a novel 4D-aware video tokenizer for compact world representation. Specifically, our approach comprises three key designs: 1) a spatiotemporal disentanglement strategy that factorizes videos into static and dynamic tokens for holistic world modeling; 2) a track-aware dynamic attention mechanism that aggregates trajectory-aligned cues to promote cross-view motion consistency; and 3) Co4DGen, a diffusion prior learned over the resulting VideoTok4D token space for efficient 4D scene generation. Extensive experiments have demonstrated that our proposed method achieves state-of-the-art performance while requiring up to 4 orders of magnitude less storage than dense 4D representations. Moreover, the compact token space substantially shortens diffusion sequences, enabling efficient generation.

Multimodal large language models (MLLMs) process hundreds or thousands of visual tokens per image, incurring prohibitive inference costs. While existing vision token pruning methods mitigate this overhead, they implicitly assume that a single fixed pruning strategy can be applied uniformly across all inputs. Our analysis further reveals that ranking pruning methods by average benchmark accuracy conceals substantial sample-wise complementarity: although the average-best strategy excels overall, alternative strategies prove superior on a significant fraction of individual samples. To harness this diversity, we propose VIP-Router, a lightweight VIsion Pruning Router that adaptively selects the pruning strategy predicted to be best suited to each input at a specified pruning level. Conditioned on low-cost visual and textual features, VIP-Router identifies the most suitable candidate strategy while retaining full-token inference as an option when pruning is predicted to be unfavorable. Evaluated on a curated suite of pruning-sensitive visual perception benchmarks, VTC-Bench Group A, VIP-Router consistently outperforms the best fixed strategy baseline across all reduction ratios, achieving a 26.9% relative improvement in average accuracy, and a 22.0% relative increase in average utility after accounting for realized token cost. Crucially, VIP-Router operates in a plug-and-play manner without modifying underlying pruning algorithms or model weights, introducing trainable parameters equivalent to merely 0.017% of the backbone. Furthermore, VIP-Router proves effective across various MLLM backbones and yields consistent gains on unseen benchmarks, highlighting the potential of sample adaptive routing for visual token pruning.

Long-video language models cannot look at every frame: an hour sampled once per second is 3,600 images, and a system keeps only a small fixed slice of that pool. Which frames survive that slice is usually treated as a preprocessing detail; we test whether it should be. Published selectors make the comparison hard because they change the frame scorer, the prompt boundary, the resolution policy, and the answering model all at once. We hold each fixed and vary one decision at a time: selection, spatial compression, and reinvestment of the savings, across six training-free selection rules, three long-video benchmarks, and two answering models. Selection is the largest single lever: on LongVideoBench's hour-long bin, eight query-selected frames beat sixteen uniformly spaced ones by 6.9 points, and Orthogonal Matching Pursuit, an unmodified decades-old sparse-approximation algorithm, matches or comes within a point of every purpose-built selector we compare it against, across all three benchmarks. Compression is close to free: halving each frame's spatial budget at fixed timestamps costs at most 0.44 points. Reinvestment is where that budget turns back into accuracy: spending the freed tokens on twice as many compressed frames, at a measured cost no higher than the original eight, returns a further two to three points; compression only pays off once its savings are spent this way. Along the way, an implementation bug in our own AKS baseline and a 0.07 to 3.74 point gap between two harnesses running the same published rules at the same budget show why these comparisons need to happen inside one controlled harness rather than across papers.

When the visual encoder and the language decoder of a vision-language model (VLM) run on different compute nodes, the intermediate visual-token embeddings become a communicated payload rather than an internal activation. We call such machine-consumed intermediate tensors AI traffic and ask how far they can be compressed with a standardized, training-free codec. We insert ISO/IEC 15938-17 Neural Network Coding (NNC) round trips on the complete visual interface of a Qwen3-VL-8B-Instruct video question answering pipeline, comprising the main visual-token representation and the DeepStack feature streams, while leaving weights, prompts, and generation untouched, and sweep the quantization parameter (QP) over a wide rate range. Closed-ended Video-MME accuracy remains close to the uncompressed reference up to a 98% reduction of the transmitted BF16 tensor and only then collapses; open-ended MLVU generation shows the same plateau-and-collapse profile under an LLM judge. This robustness is not due to near-lossless reconstruction: the decoded tensor is heavily discretized, carries substantial row-wise relative L2 error, and has a visibly steeper singular-value decay than its source. Downstream reasoning therefore depends on coarse structure and relative geometry rather than exact floating-point values, which argues for rate-task rather than rate-distortion optimization of AI traffic codecs.

Despite their strong multimodal understanding ability, multimodal large language models (MLLMs) incur substantial computational overhead when processing long visual token sequences. To reduce inference costs, recent studies have explored visual token pruning through vision-centric or text-guided strategies. However, these methods often overlook high-norm outlier tokens, i.e., tokens with abnormally large feature norms, leading to suboptimal pruning decisions. In this work, we show that such high-norm outlier tokens are highly redundant in both feature and spatial dimensions, yet are often mistakenly preserved as informative cues by existing methods. Motivated by this observation, we propose SinkPruner, a training-free visual token pruning framework for efficient MLLM inference. SinkPruner follows a coarse-to-fine design with two key modules: a visual sanitizer that filters high-norm redundancies and alleviates attention sink and attention dispersion, and a text-guided pruner that further retains tokens semantically aligned with the text query. Extensive experiments on twelve image-language and four video-language benchmarks demonstrate the effectiveness, efficiency, and generalizability of our framework. Notably, SinkPruner preserves 96.5% (91.8%) of the original performance of LLaVA-1.5 (Qwen2.5-VL) under an 89% token reduction. Experiments further indicate that our visual sanitizer exhibits promising transferability in enhancing the performance of existing pruning methods. Our code is available at https://github.com/LaVi-Lab/SinkPruner.

Whether attention weights faithfully reflect model reasoning has been actively debated in NLP, yet this question remains largely unexplored for the visual modality in Vision-Language Models (VLMs). We address this gap through causal perturbation analysis on current VLMs, evaluating both the comprehensiveness and sufficiency gap of attention-ranked visual tokens. Our analysis reveals that visual attention faithfulness is heterogeneous, manifesting in three distinct processing modes: Faithful-Sufficient, where top-$k$ attention tokens are both necessary and sufficient for prediction; Faithful-Distributed, where they are necessary but broader visual context remains required; and Non-Focal, where no localized attention region is individually necessary while visual information remains an essential trigger for prediction. Furthermore, human-annotated ground-truth regions satisfy comprehensiveness in only $\sim 60$% of cases compared with model attention rankings, revealing systematic divergence between model visual reliance and human intuition. We demonstrate these patterns across both general VQA on VQAv2 and document tasks on VRDU and ChartQA, showing that visual attention faithfulness varies systematically with processing demands and model architectures rather than being uniformly faithful or unfaithful.

Large vision-language models used as listwise rerankers must jointly process visual tokens from tens of candidates per query, making token pruning essential for practical deployment. Existing pruning methods retain tokens by attention saliency, yet we show that saliency is systematically misaligned with ranking contribution: visually prominent tokens often capture order-neutral patterns shared across candidates. This mismatch is layer-dependent: saliency becomes informative only where attention is concentrated, and normalized attention entropy diagnoses the reliability shift (Pearson r=0.87). We propose RaDiCal (Rank-Discriminative Calibration), a training-free framework that uses normalized attention entropy to decide when saliency can be trusted, fusing it with an attention-free rank-discriminative prior and selecting pruning layers from the same trust landscape. Across three retrieval benchmarks and multiple VLM architectures, RaDiCal matches Dense MRR@10 on Flickr30K and surpasses it on MSCOCO at a 20% token budget, ranks first among all pruning methods on FashionIQ, and holds within 1.2 pp on Flickr30K and MSCOCO at 10% retention. It cuts FLOPs by 39--45% and delivers 1.28--1.45$\times$ measured speedups across two VLM architectures without dataset-specific retuning.

Multimodal large language models (MLLMs) struggle with fine-grained Visual Search, the task of locating small or rare objects in high-resolution images. Existing remedies fall into two families: (1) Training-free methods based on attention or confidence scores are accurate but slow, since they require multiple MLLM queries per example. (2) Reinforcement Learning (RL) trained tool-use models are faster at inference but opaque, since their tool calls remain uncontrollable and hard to interpret. To overcome this, we propose \emph{VisLens} (Visual Focus via Logit Lens), a Visual Search method built on the logit lens, which decodes the semantics held in a hidden state by projecting it through the LLM head. VisLens further uses a lightweight tuned-lens that maps early hidden states into the final hidden state space, so visual tokens can be read out from early layers. These tokens are matched to target words in the query to generate a crop of the relevant region, which is fed back in alongside the original image to produce the final answer. The whole process, from decoding to the final answer, completes in a single forward pass without repeated queries. VisLens matches or exceeds prior baselines while delivering a substantial latency advantage, running $8.5$--$9.9\times$ faster than Thyme and up to $22.2\times$ faster than training-free multi-pass search methods.

Latent video generation relies on autoencoders to define a compact space in which generative models operate. Although video autoencoder architectures have evolved substantially, their latent spaces are still optimized primarily for pixel-level reconstruction and provide limited high-level semantic organization. A reconstruction-optimal latent space, however, need not be well suited to generative modeling. We propose V-RAE, a video representation autoencoder that builds compact generative latents on top of frozen vision foundation model representations. A lightweight temporal pooling module removes temporal redundancy while preserving semantic structure, and a video decoder reconstructs continuous motion from the compressed features. We evaluate V-RAE with four representative frozen encoders on video reconstruction, semantic probing, and class-conditional generation. V-RAE achieves 2.13 rFVD on K600, outperforming all evaluated large-scale pretrained video VAEs. Its latents retain substantially more semantic information than conventional video tokenizer latents. Under matched generation settings, our best variant achieves gFVD scores of 117.86 and 19.16 on UCF101 and K600, respectively, while converging up to 6x faster}. We further show that reconstruction quality alone is insufficient to characterize generative utility and introduce tFVD, a temporal-coherence diagnostic that correlates more reliably with downstream generation quality. Beyond video generation, V-RAE also improves future video prediction on Cityscapes over the Wan 2.2 VAE latent space under matched prediction settings. Taken together, the experiments show that frozen semantic representations can support video reconstruction, generation, and predictive modeling. The project page: https://v-rae.github.io/.

Vision-Language-Action(VLA) models map camera images and language instructions directly to motor commands through a single autoregressive transformer. In real-time control, they still spend substantial compute recomputing key-value(KV) representations for visual tokens that barely change across neighboring frames. Recent work such as VLA-Cache reduces that cost by reusing KV states for visually static patches, but its policy relies only on observation-space heuristics and does not account for the model's own uncertainty. We propose Gated VLA-Cache, a lightweight, training-free extension that augments visual-similarity caching with neural introspection. The method monitors the logit margin between the top two predicted action tokens, a zero-cost confidence signal available during decoding. When the margin drops below a threshold, the cache is invalidated and a full recompute is triggered. Evaluated on four LIBERO benchmark suites with both OpenVLA and OpenVLA-OFT, Gated VLA-Cache improves reliability when blind caching hurts. On LIBERO-Goal and LIBERO-Long, it recovers over 100% of the lost accuracy while retaining 80% of the compute savings.

Visual token compression for vision--language models (VLMs) has largely relied on criteria such as attention, redundancy, and uncertainty to maximize average accuracy under a fixed compute budget, implicitly assuming that all errors carry equal cost. However, the consequence of an incorrect prediction on downstream tasks is rarely symmetric: misreading an invoice amount can be far more costly than misclassifying a background color. Motivated by this, we introduce consequence-sensitive visual token compression, which allocates visual computation across requests according to their potential error costs. Our method follows a calibrate-then-allocate procedure, estimating consequence-specific error-budget curves offline and applying the calibrated token budgets online using consequence signals available from question or task information. On a controlled within-task benchmark, high- and low-consequence questions are drawn from the same document images, so content alone cannot reveal which questions are costly to get wrong. In this setting, our method reduces high-stakes errors from 0.300 to 0.133 under the same total token budget, whereas a content-driven allocator performs no better than uniform allocation. Measuring how error rates change with token budget across different cost ratios, we derive an allocation frontier: uniform allocation is optimal when errors are equally costly, and token transfer toward high-consequence questions becomes increasingly beneficial as the cost gap grows. This allocation principle generalizes well across three dense vision-language benchmarks, two budget realization mechanisms (token deletion and resolution reallocation), two VLM architectures, and multiple token selection strategies. On a realistic mixed workload, consequence-sensitive allocation reduces cost-weighted error by 38% while achieving approximately 21% lower latency than full-resolution inference.

Next-scale visual autoregressive models (VARs) have emerged as a powerful generative paradigm, producing high-quality images through efficient coarse-to-fine prediction. However, their potential for text-guided image editing remains largely underexplored. Existing training-free VAR editing approaches often formulate editing as target-conditioned regeneration guided or constrained by the source image, and may rely on inversion, test-time optimization, attention control, or user-provided masks. This generation-centric formulation does not fully exploit the multiscale source representations provided by VARs and may introduce additional computation or intervention. We instead take a source-centric perspective on VAR editing, in which the encoded source image tokens serve as the primary visual state and the editing process focuses on condition-induced changes. Based on this perspective, we propose \textbf{EditMod}, which compares source- and target-conditioned predictions under a shared autoregressive context, treats their difference as a scale-wise editing direction, and applies it as a residual update to source tokens at selected scales. Experiments show that EditMod achieves leading source-image fidelity while maintaining strong text alignment, and completes end-to-end editing of a 1K image in only 1.57 seconds on a single A100 GPU without per-image preparation.

Latent diffusion modeling (LDM), a prominent paradigm, utilizes tokenizers to map input signal to compressed representation. This dependency positions tokenizer as an integral part of generation process itself, since it affects learning speed, quality of synthesized samples and lay foundation for later applications. This report presents series of KVAE tokenizers for audio, image and video, all designed for subsequent text-conditioned generation: KVAE-Audio, a continuous full-band 48 kHz tokenizer with a 50 Hz latent of 64 channels; KVAE-3D -- two causal video tokenizers for 4x16x16 and 4x8x8 compression; KVAE-2D, an image model, compressing input by factor of 8 with 32 channels. We demonstrate that reconstruction (PSNR, LPIPS, PESQ, etc.) and generation results on objective (Frechet Distance, CLIP score, CLAP score, etc.) and subjective (side-by-side evaluation) metrics matches or surpasses frontier opensource tokenizers, such as VAEs from Wan-2.2, HunyuanVideo-1.5, FLUX.2, MovieGen, StableAudio and MMAudio. Considering difficulty of development, we share with community training details, model selection method and ablation on design choices. The code is publicly available at https://github.com/kandinskylab/kvae and https://github.com/kandinskylab/kvae-audio.

Video large language models (Video-LLMs) have made strong progress in open-ended video understanding. However, their visual interfaces remain token-intensive and provide limited explicit structure for linking recurring object evidence across time. We introduce SlotNarrative, a slot-based interface that organizes a video into persistent object narratives represented by compact object-state tokens. Rather than compressing frame-wise features before establishing temporal correspondence, SlotNarrative first groups visual features into object-like slots and then associates recurring observations with clip-level object entries through a lightweight, parameter-free memory that integrates multiple complementary matching cues. Each retained entry is serialized into two token types: an identity token that summarizes persistent object appearance and a set of state tokens that encode segment-level appearance, geometry, visibility, and trajectory information. This design yields an interface of only 144 allocated visual-token positions for a frozen Video-LLM, independent of the number of sampled frames. Across multiple datasets, SlotNarrative achieves a favorable trade-off between accuracy and visual-token count compared with prior compact Video-LLM interfaces. Experimental results establish persistent object narratives as a compact, structured, and temporally organized visual interface for Video-LLMs. Our code will be made publicly available.

Slice-based MLLMs leverage mature 2D encoders by representing 3D volumes as sequences of 2D slices. However, this slice-wise formulation produces thousands of visual tokens that burden the LLM backbone, many of which capture overlapping visual evidence across adjacent slices. To understand how effectively a growing visual token budget improves performance, we perform scaling analyses on two 3D medical VQA benchmarks and find diminishing returns: cost keeps rising while accuracy saturates, and improving in-plane resolution is more effective than adding slices at comparable budgets. The budget should therefore be allocated more selectively rather than simply enlarged, yet most token compression methods are designed for 2D images or videos, where redundancy arises from spatial layout or temporal motion rather than from near-duplicate content along the depth axis. We present CARVE, a training-free framework that compresses visual tokens prior to LLM inference and casts token reduction as budget-constrained 2.5D allocation. CARVE partitions the depth axis into coherent windows and allocates tokens non-uniformly according to normalized cross-slice evidence. Under a shared budget, CARVE builds spatial anchors on representative slices and retrieves locally varying evidence from the full volume, then merges remaining eligible tokens into nearby anchors within each window. Removing roughly 80% of the visual tokens on Hulu-Med-7B, CARVE leads all compression baselines on every AMOS-MM report-generation metric, with 6.2 points higher retention of full-token quality than the strongest baseline, and preserves 98.1% of full-token performance across three VQA benchmarks.

High-resolution images and long videos provide vision-language models with rich context for multimodal reasoning and fine-grained perception, but the resulting long visual token sequences make large language model-side computation and memory costly. Existing visual token reducers often operate at prescribed rates, while recent methods adapt token counts across inputs using method-specific learned thresholds or importance predictors. We introduce RUTA, a principled Rate-Utility Token Allocation method that performs pre-LLM reduction by jointly learning which tokens to retain and how many to allocate to each image-query pair. RUTA constructs query-conditioned candidate tokens and predicts a retention probability for each candidate. During training, these probabilities parameterize independent Bernoulli gates, while their sum provides a differentiable training-time estimate of the token count for each pair. Retained tokens serve as anchors that aggregate information from non-retained tokens according to semantic affinity and spatial proximity. RUTA is optimized with a penalized rate-utility objective that balances downstream task loss against expected token usage. Averaged across five benchmarks and measured relative to each backbone's full-token baseline, RUTA uses only $2.0\%$ and $4.2\%$ of visual tokens while preserving $88.2\%$ and $94.4\%$ of task performance on LLaVA-NeXT-7B and Qwen3-VL-8B, respectively.

While Vision-Language Models (VLMs) have demonstrated remarkable performance in processing and understanding both text and images, their large parameter sizes lead to significant computational overhead, limiting their deployment on resource-constrained devices. While pruning has been effective for compressing Large Language Models (LLMs), directly applying it to VLMs leads to significant performance drops, largely due to redundant visual tokens interfering with importance estimation. To this end, we propose SlimVLM, a structured pruning framework designed to compress VLMs while preserving their task performance. We introduce an adaptive visual token selection strategy for VLMs that leverages average text-to-visual attention scores to assess the importance of visual tokens, removing redundant ones during pruning based on a set threshold, thereby optimizing the importance calculation. Recognizing the varying tolerance to sparsity across different modules, we also propose a Sensitivity-aware dynamic pruning mechanism that determines the appropriate pruning ratio for each module by calculating the linear reconstruction error between the outputs of the pruned and unpruned modules, ensuring overall performance stability. Experimental results show that SlimVLM outperforms existing methods across multiple multimodal benchmarks, achieving state-of-the-art performance.

Generative Vision-Language Models (VLMs) commonly treat bounding-box coordinates as independent output symbols, leaving numerical order and axis semantics implicit. We identify this representation as an important source of error in visual grounding. Hi-Token encodes each coordinate with axis-specific tokens for the hundreds, tens, and ones digits, which adds coarse-to-fine structure and increases token reuse while retaining the existing VLM architecture. Hi-GAR complements this representation with a geometry-based reward for Group Relative Policy Optimization (GRPO), using box overlap and coordinate accuracy at multiple scales. Controlled comparisons under matched training conditions show that Hi-Token improves localization throughout the evaluated IoU range. Hi-GAR further reduces low-overlap predictions and is used only during training. Experiments on three VLM backbones and the RefCOCO family show consistent gains across models and benchmarks. Hi-R1 achieves higher values than strong specialist baselines on most reported metrics. Analyses of token frequency, digit boundaries, object scale, and IoU distributions explain the effects of coordinate representation and reward training. The results show that structured coordinate generation provides an effective approach to generative visual grounding.

A 3D CT scan entering a vision-language model produces a long sequence of visual tokens, often thousands to tens of thousands per volume, and this sequence must be compressed before a language model can consume it. Token compression is well studied in general vision, but little of it targets 3D CT specifically. A common baseline is grid average, which pools regular grid cells and can blend distinct anatomy, lesion, and air into one token. We present \textbf{ORCA} (ORgan-Centroid Aggregation), a token compressor for 3D CT. It merges adjacent tokens with organ guidance and adds a sinusoidal encoding of each region's centroid to preserve spatial layout. This preserves the anatomical information a downstream model needs. ORCA is training-free and plug-and-play, producing an adjustable token set without any model change or text query. We evaluate it across two datasets (CT-RATE and Merlin) and five encoders. The evaluation spans two task types: attribute prediction over five families (size, density, location, texture, and disease) and text generation (visual question answering and report generation). At matched token budgets, ORCA improves consistently over existing compression methods. It shrinks the visual context $64\times$ and its KV-cache $50\times$, and is $31\times$ faster to process each volume. Code released at https://github.com/renjie-liang/ORCA-3DCT.

Single-image 3D hand avatar reconstruction is fundamentally ill-posed and particularly challenging due to limited visual evidence under severe self-occlusion and the complex pose-dependent deformation of highly articulated hands. Existing methods predominantly rely on implicit NeRF-style representations, whose volumetric fitting is computationally expensive and often struggles to preserve fine-grained hand details. In this work, we present OASIS, a tailored 3D Gaussian Splatting framework for single-image hand avatar reconstruction. To faithfully encode sparse image-specific appearance cues in single-view reconstruction, we construct geometry-aligned visual evidence tokens by explicitly aligning input image observations with 3D hand geometry and context-adaptively tokenizing the resulting visual evidence. Since severe self-occlusion makes the reliability of image evidence inherently visibility-dependent, we introduce a visibility-conditioned point-image attention to reliably transfer visual evidence to geometric tokens, yielding occlusion-aware Gaussian features for faithful and robust reconstruction. To further capture non-rigid deformation of articulated hands, we introduce a Feature-on-Mesh representation to enable Gaussian deformation to be guided by local surface stretching. Under this framework, we adopt a one-shot adaptation scheme that learns a shared hand prior from multi-identity training data and then fits it to a target image for target-specific reconstruction. Extensive experiments show that OASIS outperforms existing baselines in both visual fidelity and efficiency across challenging poses and in-the-wild scenarios, and further demonstrates strong versatility in downstream applications such as text-to-avatar generation and texture editing.

Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. We present ReToken, a single learnable embedding trained as an explicit retrieval target that selects a sparse set of query-relevant visual tokens from a pre-filled visual KV cache. Trained on only a small image-QA dataset, ReToken yields consistent gains across image and video benchmarks: on Visual Haystacks it improves Qwen3VL-8B by 13.4 points and InternVL3.5 by 12.4 points (>20% relative), and on LVBench it transfers zero-shot to long video for an 8.0-point gain with Qwen3VL-8B. Thanks to its lightweight design, both training and long-video inference fit on a single H100. Code is available at: https://github.com/avaxiao/ReToken

Long-video understanding commonly compresses videos into a small set of frames or visual tokens for answer generation. Existing compact pipelines focus on retaining relevant visual content as explicit evidence. Yet making evidence available does not ensure that complementary cues across moments are integrated for answering. Our key idea is to organize selected frames into query-relevant cross-frame evidence before generation. We formulate this post-selection stage as a latent evidence interface and instantiate it with GenEvA ($\textbf{Gen}erative$ $Latent$ $\textbf{Ev}idence$ $\textbf{A}ggregation$), a distribution-guided latent evidence aggregation framework. Specifically, GenEvA uses a query-conditioned evidence distribution to focus aggregation on relevant frames, forming compact cross-frame latent evidence from their frame-specific information. Since cross-frame integration is not always needed, the same distribution determines whether to insert this latent complement. Across four benchmarks and two Video-MLLM backbones, GenEvA consistently improves matched-frame baselines. At 8 frames, it raises the four-benchmark LLaVA-Video average by $+5.2$ points and Qwen2.5-VL accuracy on LVBench by $+10.1$ points. These gains require only $0.11\%$--$0.40\%$ average video-token overhead; analyses further show task-aware allocation and benefits from Adaptive Evidence Invocation.

Streaming video understanding requires models to continuously retain useful visual evidence before future questions are known. Existing approaches primarily manage the growing visual context according to token importance, temporal redundancy, or segment-level relevance, but rarely organize evidence around objects that persist and evolve over time. Thus, in this paper, we introduce ObjectStream, a training-free framework that treats latent objects as memory anchors for streaming video understanding. ObjectStream induces spatially coherent latent objects directly from frozen Video-LLM representations, links them across frames into persistent anchors, and maintains their histories under a bounded memory budget, without requiring external object detectors or segmentation models. Built on these anchors, ObjectStream preserves three complementary forms of evidence: persistent object histories, transient object changes, and recent visual context. This design enables existing Video Large Language Models (Video-LLMs) to reason over object identities, interactions, and state changes while leaving the underlying model unchanged. Extensive experiments on online streaming and offline long-video benchmarks demonstrate both effectiveness and efficiency. In online streaming evaluation, ObjectStream improves Qwen2.5-VL-7B by 10.0 points on OVO-Bench Real-Time Visual Perception, while reducing peak GPU mem-ory and TTFT by approximately 50%. On offline long-video benchmarks, it surpasses the full-token baseline while discarding 82.5% of visual tokens. These results highlight latent objects as a practical and effective organizing principle for compact streaming video memory.

Multimodal foundation models are reshaping edge-cloud visual intelligence from task-specific feature pipelines into token-based interfaces, where edge devices encode visual inputs into tokens for a general-purpose cloud MLLM. However, dense visual-token sequences increase cloud-side inference costs. Existing pruning methods mainly target centralized inference: vision-driven methods can operate before cloud execution but are typically query-agnostic, whereas query-guided methods often rely on internal states of the target MLLM and cannot determine token relevance before transmission. Compact guidance models offer an alternative, but existing designs may require costly attention aggregation or auxiliary generation. We propose LAST, a training-free framework for query-dependent visual token pruning in edge-cloud collaborative MLLM inference. LAST uses a compact edge-side VLM as a guidance proxy and derives a lightweight importance signal from the last query token's attention to visual tokens. Under causal attention, the last query token can attend to the full visual sequence and the entire query context, enabling query-aware pruning without cloud-model access, autoregressive generation, or costly aggregation over multiple query positions. LAST then retains a diverse set of query-relevant visual tokens under a fixed token budget. We evaluate LAST on 11 multimodal benchmarks under multiple token budgets against pruning methods with different guidance strategies. Experiments show that LAST consistently achieves the strongest performance, preserving 95.4% of the full-token accuracy while retaining only 12.5% of the visual tokens, with low edge-side selection overhead and reduced cloud-side computation.

Integrating 3D medical images with vision-language models (VLMs) holds substantial promise for computer-aided diagnosis. However, volumetric images generate prohibitively long visual-token sequences with considerable spatial and inter-slice redundancy. Existing token compression methods typically apply uniform reduction or rely on a single importance signal, increasing the risk of removing regions that are clinically relevant to the query or structurally distinctive. To address this limitation, we propose MedARC, a unified, training-free framework for Adaptive Redundancy Compression of visual tokens in 3D medical VLMs. MedARC estimates token importance by integrating three complementary cues: self-attention from the VLM vision encoder, which reflects the model's intrinsic visual focus; similarity between projected visual tokens and text embeddings, which identifies query-relevant regions; and deviations of local visual foundation model features from the volume-level feature center, which highlight structurally distinctive anatomy. The resulting importance distribution guides a saliency-aware merging strategy that preserves informative tokens while consolidating redundant ones rather than simply discarding them. Experiments on CT-RATE and MR-RATE show that MedARC reduces visual-token overhead and inference time while preserving or improving diagnostic performance. Its multi-cue scoring cost is outweighed by the savings from processing fewer tokens, with greater benefits expected for larger language models.

Recent multimodal large language models (MLLMs), such as Qwen2.5-VL and InternVL3, generate large numbers of vision tokens for high-resolution inputs, leading to substantial computational cost. Existing vision token pruning methods either depend on cross-modal attention and cannot prune before the prefill stage, or rely on diversity estimation with high computational overhead. We observe that attention scores from both vision and text tokens peak at modality separator tokens, suggesting that these separators bridge the two modalities. Based on this observation, we propose SepPrune, an efficient, training-free, plug-and-play pruning method that uses the separator token as a unified query to rank and select informative vision tokens. SepPrune reuses the LLM's built-in projection parameters and requires no architectural changes. Experiments on Qwen2.5-VL-7B show that SepPrune achieves state-of-the-art performance, retaining 96.3% of the original accuracy while removing 80.2% of vision tokens.

Vision-language models commonly project all tokens produced by a pretrained vision encoder into a large language model. However, final-layer features can discard text, local attributes, and spatial relationships, while high-resolution inputs substantially increase context length and inference latency. We introduce \method, a Multi-scale Adaptive Vision Encoder. \method uses position-dependent gates to fuse shallow, intermediate, and deep features from a vision Transformer, preserving global semantics while enhancing edges, text, and local structure. It then performs question-conditioned token routing according to question relevance, local information content, global semantics, and spatial coverage, with a token budget that adapts to image complexity. To mitigate compression loss, we further introduce full-to-compressed representation distillation and a spatial diversity regularizer. In an illustrative simulation under a unified 7B language-model framework, \method reduces the average number of SigLIP-SO400M visual tokens from 729 to 146 (approximately 80.0%) and improves the mean score on VQAv2, GQA, TextVQA, ScienceQA-IMG, and MMBench by 2.2 percentage points, while reducing single-image time to first token from 228,ms to 129,ms. We provide the full model design and evaluation protocol. All reported numbers currently serve only as placeholders for paper organization and experimental design; formal claims require real training runs, independent replications, and official benchmark evaluation.

Existing Large Vision-Language Models (LVLMs) struggle with long-form video understanding due to the quadratic computational cost of visual tokens. While recent efficient methods attempt to compress tokens via hard pruning or uniform merging, they operate strictly in the spatial feature domain, where robust structural context and discriminative semantic details are inherently entangled. In this work, we propose WaveZip, a joint signal-frequency-domain framework for efficient video inference. Driven by the insight that temporal redundancy resides in low-pass approximation scales while spatial saliency strongly correlates with high-frequency components, WaveZip leverages Discrete Wavelet Transforms (DWT) to disentangle these signals. Temporally, it employs 1D DWT to analyze query-frame relevance, and the resulting high-frequency coefficients are further gated by inter-frame differences, with both signals jointly driving the dynamic allocation of a precise frame-level token budget. Spatially, a 2D DWT decomposes features into low-frequency approximations and high-frequency detail components, where the high-frequency coefficients are modulated within query-salient regions to regulate spatial reconstruction. Importantly, WaveZip requires no task-specific training and can be seamlessly integrated into off-the-shelf LVLMs to boost inference efficiency. Extensive experiments on long video understanding benchmarks demonstrate that WaveZip retains 99.6% of the full performance under an extreme 10x compression ratio, consistently outperforming state-of-the-art methods.

Unified multimodal models seek a shared visual token space that supports both multimodal understanding and image generation. Discrete methods unify the interface via a shared codebook, whereas continuous pipelines often rely on two disparate representations -- semantic features (e.g., ViT) for understanding and low-level latents (e.g., VAE) for synthesis -- resulting in mismatched latent spaces. We propose Twins, a unified continuous token space formed by channel-wise concatenating ViT and VAE features on the same token grid, so the sequence length is unchanged and attention cost does not increase. However, jointly modeling Twins in a Diffusion Transformer exposes a severe optimization imbalance: the model fits the ViT component well but struggles to match the VAE latent distribution. We trace this imbalance to three sources of heterogeneity: frequency bias, intrinsic dimensionality, and condition-aligned vs condition-independent uncertainty. To address it, we adapt a focal regression objective for flow matching that upweights large-error VAE dimensions, better balancing optimization across the ViT and VAE components. On ImageNet, this yields up to 10.57 gFID gain over naive MSE loss without classifier-free guidance. Twins also performs competitively on multimodal understanding benchmarks and improves reconstruction fidelity, narrowing the gap between understanding- and generation-oriented representations.

The goal of this paper is to reduce the input token cost of Omni-modal large language models (Omni-LLMs) at inference time. Omni-LLMs reason jointly over audio, video and text, but the cost of the three streams is highly unbalanced: visual tokens account for the vast majority of the input, and are highly redundant. In this paper, we propose ReMo, a training-free framework that compresses visual tokens by redistributing their information across modalities: a visual token is kept only if its information appears nowhere else. ReMo achieves this in two ways: (i) it aligns audio and video in a common embedding space, and removes visual tokens already explained by the audio or by other visual tokens; and (ii) it replaces object-level visual tokens with compact text proxies, short descriptions of each object and its location, conveying the same content in far fewer tokens. On Qwen2.5-Omni at two model scales, ReMo removes 54% of the input tokens with no loss in accuracy. Indeed, it slightly exceeds the full-token model, reaching 101.2% and 101.3% of its average accuracy over five audio-visual benchmarks.

Vector Quantization (VQ) underpins modern discrete visual tokenization. However, training quantization modules for state-of-the-art VQ-based models requires significant computational resources which, in practice, all but prevents the development of novel, cutting-edge VQ techniques under resource constraints. To address this limitation, we propose {\bf VQ-Transplant}, a simple framework that enables plug-and-play integration of new VQ modules into frozen, pre-trained tokenizers by replacing their native VQ modules. Crucially, the proposed transplantation process preserves all encoder-decoder parameters, obviating the need for costly end-to-end retraining when modifying the quantization method. To mitigate decoder-quantization mismatch, we introduce a lightweight decoder adaptation strategy (trained for only 5 epochs on ImageNet-1k) to align feature priors with the new quantization space. In our empirical evaluation, we find that VQ-Transplant allows obtaining near state-of-the-art reconstruction fidelity for industry-level models like VAR while reducing the training cost by 95%. VQ-Transplant democratizes quantization research by enabling resource-efficient integration of novel VQ techniques while matching industry-level reconstruction performance.

Long-video question answering requires a model to preserve visual evidence over time without repeatedly reprocessing the same video. A practical approach is to store the vision-language model's internal key-value (KV) cache for each video chunk and retrieve that state at query time. However, independently cached video chunks do not compose correctly: every chunk is prefilled from local rotary position zero, so naive concatenation collides temporal phases and removes the global order required for questions about what happened first, how often events occurred, or what changed across the video. This paper presents ChronoStitch, a training-free method for composing independently stored visual KV memories. The method first re-bases stored post-rotary keys onto a global three-axis multimodal RoPE coordinate system that preserves time, height, and width structure. We show why a one-dimensional scalar re-indexing is geometrically inconsistent for visual tokens because it turns spatial order within a frame into false temporal displacement. We then address the residual content gap left by positional repair: later chunks were originally encoded without attending to earlier chunks. ChronoStitch therefore selectively recomputes a small fraction of high-deviation later-chunk visual tokens while allowing them to attend over the composed cache. On Qwen2.5-VL-3B and the temporal split of TempCompass, ChronoStitch outperforms naive composition and position-only variants, improving event-ordering accuracy while running 3.3x faster than full joint re-prefilling.

Continuous diffusion models have become the dominant paradigm for photo-realistic image Super-Resolution (SR), but they typically formulate reconstruction as continuous signal-level denoising and incorporate semantic priors through external conditioning modules. This makes it less direct to exploit the unified token-based scaling paradigm of modern multimodal models. Autoregressive models provide a more native semantic representation by modeling images as discrete visual tokens, yet their causal decoding is inefficient for high-resolution reconstruction. Discrete diffusion offers a promising middle ground by enabling non-causal, parallel prediction over visual tokens. However, directly adapting discrete diffusion to SR remains non-trivial due to two task-specific challenges: (1) the long-tailed distribution of visual tokens, which under-represents rare but perceptually critical textures; and (2) spatially inconsistent parallel decoding, which may introduce isolated artifacts. To address these issues, we propose DiMOO-SR, a rarity-aware multimodal discrete diffusion framework for photo-realistic image SR. During training, Inverse Frequency Sampling (IFS) prioritizes under-represented but information-rich tokens. During inference, Spatial Consistency Ranking (SCR) refines token confidence using local neighborhood agreement to improve structural coherence. Extensive experiments on widely used real-world SR benchmarks demonstrate that DiMOO-SR achieves competitive perceptual quality with only a few parallel decoding steps, highlighting the potential of discrete diffusion for generative image super-resolution. The code will be released upon publication.

Latent diffusion models (LDMs) enable efficient high-resolution image synthesis by denoising in a VAE-compressed latent space. However, fixed visual tokenizers can discard fine textures and structural details, while separate representation and diffusion training creates a mismatch between reconstruction and generation objectives. These limitations have renewed interest in pixel-space diffusion, which models raw pixels directly, removes the VAE bottleneck, and supports end-to-end optimization. This formulation better matches the demands of high-fidelity generation but introduces challenges in high-dimensional modeling, including noise scheduling, loss weighting, token efficiency, and scalable architecture design. Pixel-space modeling also offers a promising basis for unified multimodal systems: raw pixels, text, and task conditions can be represented in a shared token space and jointly processed by a single Transformer, narrowing the gap between visual understanding and generation. This paper reviews Pixel-Space Diffusion Transformers (pDiTs) from the perspectives of model architecture, continuous generative mechanisms, and unified multimodal modeling. We summarize representative methods, identify key technical challenges, and discuss future directions toward high-fidelity, end-to-end vision foundation models that integrate generation and understanding.

Understanding how vision-language models (VLMs) interpret data visualizations remains an open problem, and is increasingly important as these models are used for analytical tasks where reliable reasoning is essential. We introduce a lightweight, diagnostic saliency map method tailored for text generation over images using transformer models, the current state-of-the-art models in visualization interpretation. Our approach aggregates the language model's attention over the visual tokens across all heads and layers, then maps this attention back onto the vision encoder's patch grid to localise it over the image, producing a direct correspondence between each generated answer token and the image regions it attended to. This yields fast, gradient-free saliency maps that expose how VLMs allocate focus across visual elements during answer generation, enabling inspection of whether model attention aligns with semantically relevant components. We evaluate our approach using a deletion metric which validates the causal faithfulness of our saliency maps to the model's behavior.

Recent visual place recognition (VPR) methods based on vision transformers, particularly foundation models, have achieved remarkable recognition performance. However, these models process all visual tokens throughout the entire network, resulting in substantial computational overhead, which hinders their deployment in real-time and resource-constrained scenarios. A natural question thus arises: are all visual tokens necessary for VPR? To answer this question, we present the first systematic benchmark of token reduction for efficient visual place recognition. Our benchmark comprehensively evaluates representative token pruning, token merging, and hybrid pruning-merging methods across multiple state-of-the-art VPR models and diverse benchmark datasets covering urban, suburban, and natural environments. We further investigate token reduction from multiple perspectives, including recognition performance under different reduction configurations, computational complexity, inference speed, qualitative visualization, and deployment efficiency on edge devices. Through extensive experiments and in-depth analysis, our benchmark reveals multiple important characteristics of token reduction in VPR and provides several practical insights into the trade-offs between accuracy and inference efficiency. For example, token reduction can reduce computational cost by up to 29% and improve throughput by up to 44%, while incurring less than 1% degradation in recognition accuracy. Overall, this work establishes a comprehensive foundation for future research on token-efficient VPR and efficient visual retrieval systems. Our codes and models will be available at https://github.com/Tong-Jin01/TokenReduction4VPR

Vision-Language Models (VLMs) have achieved strong performance in multimodal understanding, yet remain challenging to deploy on resource-constrained edge devices due to the substantial computational overhead of processing numerous visual tokens. Token reduction is a promising direction for accelerating VLMs inference, but existing approaches either rely on attention maps that are incompatible with modern acceleration frameworks or depend on computationally intensive pairwise similarity comparisons, which undermine scalability and negate their practical benefits in deployment. In this paper, we propose an attention-free and lightweight token reduction framework as a plug-and-play module for VLMs, which preserves both important and diverse tokens to produce a compact visual representation. First, to enable attention-free importance estimation, we adopt an information-theoretic perspective and quantify token information using a novel entropy-based criterion, retaining those with more expressive and less degenerate feature representations. Second, to ensure diverse visual coverage in a lightweight manner, we introduce a transformation-induced consistency signal where similar tokens yield similar signals, such that sorting by this signal places similar tokens close to each other and enables stride-based selection to produce a diverse token set. Extensive experiments across multiple VLMs benchmarks demonstrate that our framework achieves a favorable accuracy-efficiency trade-off, maintaining competitive performance under aggressive compression.

Large Vision-Language Models (LVLMs) face significant challenges in long video understanding due to the excessive computational cost and information loss associated with uniform sampling. Existing keyframe selection methods often treat video frames as atomic entities and allocate visual budgets equally, thereby overlooking high-level semantic structures and introducing substantial redundancy. To address these limitations, we propose GMM-EVA (Gaussian Mixture Modeling for Event-Aware Visual Allocation), which leverages Gaussian Mixture Models to model event-level structure from discrete frame-wise observations. A differentiated allocation strategy is then applied to preserve one primary high-resolution keyframe per event for high-fidelity detail, while utilizing lower-resolution secondary keyframes to maintain temporal context and optimize token budgets. GMM-EVA is a training-free, plug-and-play framework that generalizes robustly across various relevance measures and downstream LVLMs. Extensive experiments on multiple long video benchmarks demonstrate that our method significantly outperforms uniform sampling. Notably, GMM-EVA achieves comparable performance to baseline selection methods while utilizing only approximately half of the visual token budget, highlighting its superior efficiency and effectiveness.

Vision-Language Models (VLMs) are costly at inference time because they must process long sequences of visual tokens. Existing token pruning methods often degrade under high compression by blindly discarding information, breaking spatial structure or collapsing diversity. We propose SpecFlow, a training-free framework that shifts the paradigm from destructive pruning to conservative condensation, strictly enforcing spatial coverage and statistical conservation to ensure stability. Treating visual tokens as nodes in a $k$NN graph, SpecFlow (i) computes a stable importance field via spectral heat flow to preserve structural coherence, (ii) allocates budgets via adaptive spatial partitioning to guarantee coverage, and (iii) aggregates discarded information into coreset sinks to maintain statistical conservation. The method is plug-and-play, requires no fine-tuning, and is compatible with FlashAttention. Experiments confirm that our SpecFlow outperforms SOTA methods across tasks, VLM architectures, and pruning ratios. Notably, LLaVA-1.5 with SpecFlow retains 95.6% of original performance despite pruning 88.9% of visual tokens, offering an exceptional efficiency-accuracy balance. Code is available at https://github.com/Lzy-dot/SpecFlow

Ambivalence and hesitancy are subtle behavioral states that are expressed through a combination of verbal content, facial behavior, visual context, and acoustic cues. Effective recognition therefore requires not only extracting informative unimodal representations, but also modeling how temporally aligned behavioral evidence interacts across modalities. In this paper, we propose a synchronized visual-facial cross-refinement framework (SVF-CR) with pairwise multimodal evidence fusion for ambivalence and hesitancy recognition. The proposed method first extracts whole-video segment tokens and cropped-face segment tokens using the same temporal partition. The synchronized visual and facial tokens are refined through intra-modal self-attention and bidirectional visual-facial cross-attention, allowing whole-video context and local facial behavior to mutually refine each other before evidence construction. We then construct segment-level visual-facial evidence using consistency and discrepancy modeling, followed by temporal self-attention and attention pooling. Textual and acoustic features are lightly refined through context self-attention and are fused with the enhanced visual-facial evidence at the final decision stage using pairwise evidence fusion. Experiments on the BAH (Behavioral Ambivalence/Hesitancy) public evaluation split show that the proposed synchronized visual-facial cross-refinement improves public macro-F1 over both global visual-face token fusion and synchronized evidence baselines, achieving a public macro-F1 of 0.7156. Code is available at : https://github.com/hiinnnii/BAH-Challenge-ECCV2026\_SVF-CR.

We present Subtoken Vision Transformer (SubViT), a selective image tokenization method for fine-grained visual recognition. Standard Vision Transformers compress each fixed-size patch into a single token, although fine-grained distinctions often depend on localized variations within only a few patches. SubViT addresses this mismatch by representing discriminative patches with multiple subtokens while retaining the original token sequence for global context, thereby allocating additional capacity where it is most needed. Since attention heads encode complementary semantics and extracting attention maps at inference requires an extra backbone forward, we adopt a two-stage training strategy. Stage 1 fine-tunes the ViT using subdivision regions sampled from random attention heads, exposing the model to diverse subdivision patterns. Stage 2 identifies informative attention maps through feature-degradation distances and distills them into a lightweight single-map router, which directly predicts deterministic token-importance scores without a separate attention forward. We evaluate SubViT on Generalized Category Discovery (GCD), a challenging task requiring both fine-grained discrimination and generalization to unlabeled novel categories. Across CUB, FGVC-Aircraft, and Stanford Cars, SubViT improves the average novel-category accuracy of DINOv2 from $81.3\%$ to $84.7\%$, with only $0.50$ ms additional latency and $3.4\%$ more FLOPs, while reducing latency by $73.8\%$ relative to Retina Patch. Results on CIFAR-10 and ImageNet-100 demonstrate its broader applicability.

Multi-vector vision-language retrieval preserves fine-grained visual evidence through maximum-similarity late interaction, but dense image-side tokens make storage and scoring expensive. Existing token compression methods reduce this cost, yet they can remove or collapse object- and region-level evidence that future query tokens may need to select. We propose SaMer, an object-aware token merging framework that compresses image-side post-projector tokens into $K$ representative centroids while preserving the original late-interaction interface. SaMer uses object annotations only during training as a merge prior to discourage cross-instance mixing, requires no ground-truth bounding boxes or detectors at inference time, and adapts only the shared projection layer with frozen vision and language backbones. With $K=64$, SaMer removes more than 93% of image-side tokens and reduces ColPali storage by $16.09\times$, while improving R@1 on Flickr30K and MSCOCO. These gains arise because object-aware merging preserves query-selectable object evidence that pruning or feature-only pooling can remove or collapse. SaMer also outperforms compression baselines and shows stronger phrase-level grounding, suggesting that efficient multi-vector retrieval depends not only on reducing token count, but on preserving the evidence future query tokens need to select.

Large Vision-Language Models (LVLMs) have achieved remarkable progress in visual understanding tasks such as image captioning and visual question answering. However, they remain susceptible to hallucinations, generating content that is inconsistent with the actual visual input. Existing methods primarily intervene at the decoding stage, while overlooking a critical source of hallucinations: irrelevant or noisy visual tokens that mislead the decoding process. To address this issue, we propose SeeMe, a training-free framework that introduces the concept of feature engineering from traditional machine learning into LVLMs. SeeMe restructures visual tokens through a three-stage token engineering process to suppress hallucination sources while preserving informative visual evidence. Experiments on MME, POPE, and AMBER benchmarks across four LVLMs demonstrate that SeeMe consistently reduces hallucinations and improves output consistency, providing a novel perspective for mitigating hallucinations in LVLMs.

We present a black-box model-stealing attack that recovers private vision-tokenizer configurations of deployed vision-language models (VLMs), including the visual patch size and input preprocessing pipeline. The key idea is a task-level side channel induced by ViT-style patchification: when a synthetic grid image is aligned with the hidden patch grid, boundary cues are erased at tokenization, causing periodic accuracy drop. By sweeping the grid cell size and measuring these collapses, we infer the patch size; by introducing padding and a consistency-check test, we further identify whether preprocessing is dynamic- or fixed-resolution and recover the target resize resolution. Across open-source Qwen-VL variants and proprietary models including GPT and Claude, we reliably recover tokenizer-related parameters. Finally, we show that such leakage enables preprocessing-aware transfer attacks and model-targeted adversarial manipulation.

Multimodal Large Language Models (MLLMs) incur prohibitive inference costs due to long visual token sequences. Training-free visual token reduction provides an efficient solution. However, existing methods distort attention distributions, giving rise to a phenomenon we term Attention Logit Collapse. To address this issue, we propose ERA, an Entropy-guided visual token pruning framework with Rectified Attention for efficient MLLMs. Specifically, ERA comprises three crucial components: Dual-view Entropy Pruning (DEP), Bias-aware Token Recycling (BTR), and Logit-preserving Attention Rectification (LAR). First, DEP identifies representative anchor tokens by jointly modeling visual diversity and head-wise saliency. BTR then recycles pruned tokens into their corresponding anchors while estimating a cluster-level logit bias. Building upon this, LAR injects the estimated bias into attention logits, effectively rectifying the collapse induced by token reduction. Together, these components preserve visual evidence even under aggressive compression, enabling robust performance across single-image, multi-image, and video settings on a wide range of MLLMs. Beyond delivering practical acceleration, ERA establishes logit-preserving visual token pruning as a principled framework for efficient MLLMs, unifying theoretical foundation, algorithmic design, and practical deployment. The code is at https://github.com/924973292/ERA.

Multimodal large language models (MLLMs) increasingly process long visual-token sequences, increasing the overall inference computation. Existing acceleration methods usually remove visual tokens or skip visual-token updates in entire layers, but these coarse strategies may discard fine-grained evidence or suppress useful operators together with redundant ones. In this paper, we study visual-token computation from an answer-observable perspective and find that late visual-token updates can remain large while having little effect on answer-token representations. Motivated by this answer-silent redundancy, we decompose each Transformer layer into attention and FFN operators and show that useful visual computation is often operator-dominant and layer-dependent. We propose an operator-level visual-token skipping framework that preserves the full visual-token sequence while selectively bypassing redundant attention, FFN, or both. Experiments across three MLLM architectures and 10 VQA benchmarks show that our method achieves strong efficiency-accuracy trade-offs, reducing \textbf{33.7%} TFLOPs on Qwen3-VL while retaining \textbf{99.5%} of the vanilla model performance.

Large Vision Language Models (LVLMs) have achieved remarkable success on vision-language tasks, yet fine-grained perception over high-resolution images and long-context videos remains challenging. As the number of visual tokens increases, the visual attention sink phenomenon becomes increasingly severe, causing irrelevant tokens to absorb a disproportionate amount of attention mass. Recent approaches attempt to mitigate this issue by explicitly predicting bounding boxes or temporal spans and re-encoding the cropped visual regions. Such methods depend on unreliable numeric localization in the discrete token space and incur significant computational overhead due to additional forward passes. In this work, we propose VisReflect, a simple yet effective framework that improves fine-grained perception in long visual contexts through latent visual reflection. Instead of decoding intermediate predictions into discrete tokens, the model generates continuous visual reflection that represents question-relevant visual features in the latent space. These reflections selectively emphasize salient regions or frames, guiding attention towards relevant visual tokens within a single forward pass. We conduct comprehensive evaluations on challenging high-resolution image benchmarks, including BLINK, V*, and HRBench-4K/8K, as well as video understanding benchmarks such as MVBench, VideoMME, and MLVU. Our method consistently improves over strong baselines, achieving gains of 4.1% on image benchmarks and 1.8% on video benchmarks. Compared with zooming-based methods, our model achieves comparable performance while reducing inference time by roughly 44% on video understanding.

We propose Nemotron-Labs-Diffusion-Image, a state-of-the-art masked discrete diffusion model (MDM) for high-resolution text-to-image synthesis. Compared with prior work on masked image generation, Nemotron-Labs-Diffusion-Image addresses two key challenges. First, unlike continuous diffusion models which progressively refine latent representations across the entire image, standard MDMs lack self-correcting capability because discrete tokens cannot be modified once they are unmasked. Second, although increasing the vocabulary size of discrete image tokenizers improves reconstruction fidelity, it introduces optimization difficulties for generative modeling as the per-token training signal becomes increasingly sparse. To address the first challenge, Nemotron-Labs-Diffusion-Image incorporates a token-editing mechanism that enables the model to dynamically revise already-unmasked tokens during inference, similar to how a sculptor iteratively refines their work. To tackle the second challenge, we propose a Grouped Cross-Entropy (GCE) objective that assigns positive learning signals to tokens neighboring the ground truth in embedding space, thereby alleviating signal sparsity. To further improve training efficiency, we implement a custom fused operator for GCE that significantly reduces VRAM usage in large-vocabulary settings. Experimental results demonstrate that these innovations substantially improve both training efficiency and image fidelity of masked discrete image generators, achieving a score of 0.90 on GenEval, 86.9 on DPG and 10.76 of HPSv3.

Vision-language models and vision-language action models endow the robot with unprecedented capabilities. However, the input of video and high-resolution images yields a massive number of visual tokens, leading to extremely high inference latency and severely hindering the robot's real-time control. To break through this computational bottleneck, we propose ST-Merge, a plug-and-play, training-free framework that efficiently fuses redundant tokens directly during the visual encoding phase. By explicitly constructing 3D spatiotemporal coordinates, it employs a multi-queue parallel matching and weighted aggregation mechanism to achieve efficient and geometrically consistent fusion of redundant tokens across frames. In addition, we introduce a post-merge positional correction mechanism that effectively eliminates spatial deviation caused by merging by dynamically re-evaluating the rotational position code of the weighted centroid of the vision token, thereby ensuring the high-precision spatial awareness required for dexterous operation. In the Video Question Answering task on the mainstream VLM, Qwen2.5-VL, ST-Merge achieves a 2$\times$ inference speedup with only a tiny 1% loss in precision. When deployed on the $π_{0.5}$ VLA policy, ST-Merge achieves an 8.3$\times$ speedup at 1024 $\times$ 1024 resolution and matches the baseline success rate at this high-resolution setting. At lower resolutions, it introduces a small drop in accuracy.

Multimodal Large Language Models (MLLMs) describe artworks with remarkable fluency, yet the visual reasoning behind their outputs remains opaque. When an MLLM names a style, identifies a subject, or recognizes an iconographic symbol, does it ground each claim in the relevant region of the canvas, draw on an undifferentiated visual signal, or rely primarily on textual priors? We study this using the Token Activation Map (TAM), which produces, for each generated token, a heatmap isolating the visual evidence specific to that token from prior-context interference. Applying TAM to a curated set of paintings spanning multiple periods and genres, we analyze grounding patterns across five semantically distinct token categories: common visual objects, style descriptors, metadata, iconographic tokens, and affective expressions. We find that visual grounding varies substantially with token semantics. We further show that MLLMs attempt to identify artworks and artists, achieving higher accuracy in artist attribution than in title prediction, where hallucinations are more frequent. Finally, we compare TAM with SAM~3 open-vocabulary segmentation. To ensure reproducibility, we release our code, experimental configurations, prompts, and qualitative results on the project page at https://nicolafan.github.io/tamart/.

Recently, self-evolving large multimodal models (LMMs) have received attention for improving visual reasoning in a purely unsupervised setting. However, multi-role self-play and self-consistency reward schemes in existing self-evolving LMMs optimize answer agreement without ensuring the decoder attends to visual content, relying instead on statistical language priors to produce self consistent outputs. This leads to a persistent failure mode we term visual under-conditioning, where the decoder relies on language priors rather than the image during generation, manifesting as insufficient attention to visual tokens. As a result, current self-evolving LMMs struggle on vision--language understanding tasks such as image captioning and visual question answering. To address this, we propose VISE (Visual Invariance Self-Evolution), a purely unsupervised self-evolving framework that directly regularizes the model's visual conditioning policy through two complementary invariance-based rewards: a geometric invariance reward that enforces spatial consistency under known transformations, and a semantic invariance reward that penalizes evidence-agnostic generation by requiring the model to recognize the absence of evidence when predicted regions are perturbed. VISE operates within a single model without specialist roles, external reward models, or annotations, and is trained on raw unlabeled images. Experiments on 18 benchmarks demonstrate the efficacy of our approach. Using Qwen3-VL-2B as the base model, VISE achieves gains of $+16.85$ CIDEr on COCO and $+19.66$ CIDEr on TextCaps, reduces object hallucination by $5.0$ Chair-I points, and generalizes across four model families and scales. Our code and models are available at https://mbzuai-oryx.github.io/VISE

Currently, streaming video understanding is still a daunting task for existing \emph{multimodal large language models} (MLLMs). Its difficulties not only lie in handling the ever-increasing video frames, but also in the unpredictability of future video content and input instructions. In this paper, we study this task from the perspective of constructing a dynamic but fixed-budget memory bank, and propose a novel and training-free approach termed \emph{\textbf{CausalMem}}. CausalMem is dedicated to constructing a dynamic visual memory update mechanism, thereby maximizing the amount of information in streaming video within a limited memory space, much like the human brain. In practice, CausalMem estimates the redundancy of visual tokens and updates the memory bank via an online semantic basis, which models the principal semantics of the observed video stream. To validate CausalMem, we apply it to two representative MLLMs, namely LLaVA-OneVision and Qwen2.5-VL respectively, and conduct extensive experiments on both streaming and offline video understanding benchmarks. The experimental results not only show the great advantages than existing methods under both streaming and offline settings, \emph{e.g.}, $+3.2\%$ and $+3.0\%$ average accuracy gains respectively, but also witness the superior semantic preservation for streaming videos, \emph{e.g.}, using 12$k$ token budgets to memorize hour-long streaming videos, which achieves more than \textbf{20$\times$} visual token compression ratio and only occupies about \textbf{82 MB} storage. \textbf{Our code} is given in \href{https://github.com/hktk07/CausalMem}{CausalMem}.

Generating human-object interactions (HOI) is central to character animation, robotics, AR/VR, and embodied AI. Recent HOI generation methods synthesize motion from text, object geometry, and sparse waypoints, controlling action semantics and object trajectories. However, these signals underspecify interaction: the same prompt and trajectory can produce different grasps, approach directions, body poses, object poses, contacts, and body-object layouts. We address this ambiguity with a reference image as a visual specification of the desired interaction snapshot. However, a single global image representation conflates distinct cues and conditions all frames on identical visual evidence. We therefore introduce IMAGIN-4D, a diffusion-based HOI generator that decomposes image conditioning spatio-temporally. For spatial conditioning, IMAGIN-4D extracts supervised interaction-state tokens for body pose, object pose, body-object contact, and spatial relationships at the depicted frame. For temporal conditioning, it computes frame-aware tokens by querying image patches per generated frame, allowing sequence segments to attend to different visual cues from the same image. To balance image, text, and waypoint cues, IMAGIN-4D uses role-aware conditioning: text, waypoints, and interaction-state tokens use separate AdaLN streams, while frame-aware visual tokens cross-attend with motion tokens. Since HOI motion datasets lack paired images, we build a synthetic motion-to-image rendering pipeline from FullBodyManipulation (FBM) and introduce an image-adherence metric to evaluate whether generated motions match the reference snapshot. Experiments on FBM and BEHAVE show that IMAGIN-4D improves fine-grained interaction control over single-token and uniformly image-conditioned baselines while preserving waypoint-following and motion quality. Code and models will be released at https://imagin4d.github.io.

Unsupervised domain adaptation (UDA) aims to learn a target-domain classifier from labeled source data and unlabeled target data under distribution shift. Recent diffusion-based UDA methods approach this problem by synthesizing labeled target-style images and training on the resulting synthetic data. However, their performance depends heavily on the conditioning design: class prompts provide only coarse guidance, while domain adaptation modules mainly control appearance, which may leave target-style synthesis insufficiently specified. We propose VT-DUDA, a visual-token conditioning framework for diffusion-guided UDA. Instead of relying only on text prompts, VT-DUDA uses source images to provide additional instance-level visual context for target-style synthesis. Specifically, VT-DUDA maps each source image to a compact sequence of visual tokens and forms a hybrid conditioning context by concatenating these tokens with the corresponding text embeddings along the cross-attention context dimension of a latent diffusion model. This provides instance-dependent conditioning beyond text alone, while synthesis is performed with the target-domain adapter branch. Because guidance is represented explicitly as a token sequence, the same interface also permits inference-time manipulation of the conditioning signal through token selection and token-strength adjustment. The proposed method preserves the standard diffusion objective and can be integrated into existing adapter-based diffusion frameworks without modifying the backbone. Across Office-31, Office-Home, and VisDA-2017, VT-DUDA improves average target-domain accuracy over strong discriminative and diffusion-based UDA baselines. The results suggest that, in generation-based UDA, a stronger conditioning interface can improve the downstream usefulness of synthetic target-style data.