Vision Encoders

In recent years, pre-training has become fundamental to learning effective video representations, enabling strong transfer to downstream tasks. A popular framework in pre-training involves aligning features of a video encoder with that of another modality, for example, language or audio. We introduce Video-STLayout pre-training, a novel strategy for obtaining rich video representations informed by spatio-temporal layout of object bounding boxes. Object layouts can easily be obtained by applying an off-the-shelf object detector on the video frames. Our method uses a contrastive loss to align video features with the layout features from a trained layout encoder. We show the effectiveness of our approach in the task of activity recognition in complex scenes.

Fine-tuning a Vision-Language-Action (VLA) model for a new deployment environment is expensive, yet most methods apply uniform-capacity adapters to every network region as if every region requires equal adjustment. This paper tests that assumption on five architecturally diverse VLAs (OpenVLA-OFT, $π_0$, SmolVLA, DTP, Octo; 93M-7B parameters). Measuring per-region adaptation cost as normalized parameter displacement under region-isolated fine-tuning reveals an adaptation spectrum in which appearance shifts concentrate cost in the vision encoder, instruction shifts in the language backbone, and novel-object shifts in the vision encoder together with the action head, across all five architectures. To exploit this structure, we introduce a pipeline that observes, diagnoses, allocates, and adapts. From ten unlabeled target observations and without fine-tuning, the diagnostic estimates per-region cost by combining reference-free gradient and Monte Carlo Dropout signals with a Centered Kernel Alignment score against a cached source reference; the allocator converts the estimates into variable-rank LoRA adapters under a parameter budget and freezes well-calibrated regions; and standard LoRA fine-tuning trains the resulting adapters. The diagnostic ranks regions within each deployment at a median Spearman of 0.91, and the allocation matches or exceeds uniform LoRA at every budget we tested on LIBERO and CALVIN. On a physical xArm-7, the pipeline matches full fine-tuning under an instruction-wording shift with 0.04% of its trainable parameters, and on five held-out scenes evaluated without retraining it leads every baseline, with 11-23 successes of 30 rollouts against 8-18 for the strongest parameter-efficient baseline at equal or larger budgets and 2-11 for full fine-tuning. These results suggest that adaptation cost in VLAs is structured enough to measure before fine-tuning begins.

Table understanding is a core task in document intelligence, encompassing two key subtasks: table reconstruction and table visual question answering (TabVQA). While recent approaches predominantly rely on vision- language models (VLMs) operating on table images, we propose a more scalable and effective alternative based on structured textual representations. These representations are easier to process, align more naturally with LLMs, and eliminate the need for language-specific visual encoders, making them particularly suitable for multilingual documents. We present DELTA, which separates physical structure recognition, logical structure recognition, and OCR to extract both layout and content accurately. DELTA outputs tables in Optimised Table Structure Language (OTSL), a compact and unified format that encodes cell arrangements and textual content. On table structure recognition (TSR), DELTA achieves TEDS- Structure scores comparable with state-of-the-art methods across FinTabNet, PubTabNet, and PubTables-1M. We further establish its robustness on non-English tables through our curated Hindi benchmark, TORQUE. Building on this, we introduce TARQA, an LLM fine-tuned on OTSL sequences. Our approach yields gains of 9.3 p.p. on WTQ (TabQA) and 9.2 p.p. on FinTabNetQA (TabVQA), respectively. On TORQUE, our method ranks second among all VLMs and DELTA + LLM variants. We release our code, models, and benchmark at: https://github.com/Tihiitborg/Tables-Decoded

No-reference video quality assessment (NR VQA) has recently seen promising progress with deep learning. However, video data is inherently large, and processing them with deep models incurs high computational cost. This challenge is particularly acute in VQA, where preserving original-resolution cues and dense temporal information is critical for accuracy. Existing efficiency-driven preprocessing strategies, such as fragmenting, reduce computation but alter the input data distribution, limiting effective reuse of pretrained video foundation models (ViFMs). To address these challenges, we propose \textbf{H}igh-\textbf{F}idelity \textbf{V}ideo \textbf{Q}uality \textbf{A}ssessment (\textbf{HFVQA}), a framework built on fixed-size spatio-temporal (ST) patches that is fully compatible with pretrained ViFMs. HFVQA samples ST patches across multiple scales, including the original resolution, with minimal temporal subsampling to preserve low-level quality cues and semantic context. To limit computation, HFVQA introduces a lightweight auxiliary network trained end-to-end with the ViFM encoder to learn \textit{VQA-specific saliency}. Distilled directly from quality supervision, this saliency captures task-specific importance patterns, reflecting that video quality perception is dominated by a small subset of spatio-temporal regions. By combining high-fidelity spatio-temporal cues with learned, task-specific saliency, HFVQA achieves SOTA performance on standard NR VQA benchmarks while processing as little as 12% of candidate ST patches, making high-fidelity ViFM-based VQA computationally tractable.

We propose a pre-fine-tuning probing method for Parameter-Efficient Fine-Tuning (PEFT) layer selection, aiming to obtain more stable and higher gains with fewer trainable parameters when adapting large vision--language models (VLMs). Unlike the common practice of applying LoRA and other adapters to all layers at once---where layer selection often relies on heuristic rules---we focus on the vision encoder and directly evaluate the "adaptability'' of each Transformer layer. Specifically, we characterize each layer from two perspectives: (i) the statistical properties of its Q/K/V projection weights (e.g., norms and condition numbers); (ii) robustness under controlled parameter perturbations. We then systematically compare these indicators with the downstream performance gains brought by applying PEFT to a single layer. Across experiments covering seven benchmarks and five PEFT variants, we observe a consistent correlation: layers (or matrices) with larger weight norms and higher condition numbers are usually more robust to perturbations and are more likely to yield larger fine-tuning gains. These results show that distribution-statistics analysis and perturbation tests before fine-tuning can provide practical signals for adaptation-layer selection, thereby maintaining or improving performance while reducing trainable parameters.

Large multimodal models tend to hallucinate visual detail fluently, which limits their deployment for fine-grained interpretation. We present HALDETECT, our system for the English hallucination-detection track (Task 1b) of ImageEval 2026, in which a system must identify, from an image and three culturally plausible statements, the single visually grounded one. We frame the item as one contrastive decision, emit the answer before its explanation, and structure reasoning around colour/texture, shape/form, and context. Our best submitted adapter fine-tunes Qwen2.5-VL-7B-Instruct with 4-bit QLoRA while freezing the vision encoder and reaches Contrastive Instability (CI) 0.035 on the 1,000-item test set; we placed third of eight teams. Development experiments show that answer order can matter more than model scale and that adaptation beats prompting alone. Retrospective paired analysis of the released gold labels confirms the QLoRA gain over the best prompt but not the small gap between the devtest-selected and best-test adapters, and reseeding all four training sizes shows that the apparent data-scaling curve does not survive a seed change. The 35 residual errors are culturally plausible function, material, and recognition distinctions; naive adapter voting does not help.

Segmenting bruises is a challenging task in medical imaging due to limited data and annotations, diffuse boundaries, and highly variable appearance. In this work, we propose BruNet, a segmentation framework that combines a ViT-based visual encoder (a self-supervised DINOv3 or a pretrained LingBot-Vision backbone) with a SAM-based mask decoder. BruNet is trained on the HAM10000 skin lesion dataset and evaluated on a separate bruise dataset without additional fine-tuning. Although a small number of prior studies have explored machine learning and computer vision for bruise analysis, existing work has primarily focused on detection, classification, or colour analysis rather than pixel-level localisation. To the best of our knowledge, this is the first study to address automatic bruise segmentation. Our results show that BruNet outperforms CNN-based models, state-of-the-art segmentation models, ChatGPT-4o/5-assisted SAM2 zero-shot baselines, and the medical-oriented MedSAM model, demonstrating strong cross-domain generalisation to bruise segmentation.

Video understanding has rapidly evolved toward video large language models (VideoLLMs): systems that couple video representations with pretrained large language models and condition generation on a textual prompt. Their strong performance on captioning, question answering, retrieval and temporal grounding comes at a computation and memory cost that grows with frame count and context length, limiting deployment in real-time, mobile and resource-constrained settings. This survey covers inference-efficiency mechanisms for visual and audiovisual VideoLLMs that report concrete reductions in parameter count, FLOPs per input, latency, memory, or visual and audio token count. We analyze bottlenecks across frame sampling, modality encoding, connector-level token reduction, and LLM prefilling and decoding. We organize methods by the pipeline stage at which they act, covering VideoLLMs developed since late 2022 together with earlier frame-sampling and vision-encoder mechanisms that remain components of current pipelines. We assemble literature-reported accuracy--cost comparisons under shared host models and input protocols wherever available, distinguish them from heterogeneous cross-paper evidence, and identify gaps in audiovisual efficiency and standardized evaluation. We maintain a repository at https://github.com/momentslab/awesome-efficient-videollm.

Vision-language models (VLMs) show promise for agricultural classification, but zero-shot performance on disease, pest, damage, quality, and species identification remains poor, and it is unclear whether this reflects weak visual features or a failure to connect them to domain knowledge. We build a benchmark of 116 datasets, 834 classes, and 8,324 images spanning these tasks to isolate where the gap arises. Linear probing shows VLM vision encoders already encode agricultural features nearly as separable as a self-supervised DINOv3 baseline, ruling out weak visual representations as the primary bottleneck. Conditioning each model on an oracle reference description (an upper bound on its parametric knowledge) closes most of the gap left by an unaided lower bound, showing VLMs already know more about agriculture than they show. To close this gap without an oracle description at inference time, we structure test-time reasoning around a fixed, per-task diagnostic rubric: the model generates $K$ candidate responses and a Probabilistic Pivot Tournament (PPT) verifier, scored pairwise against the rubric, selects the best one. This nearly doubles judged F1 over the lower bound and matches or exceeds the upper bound on several tasks, notably pushing Gemma 4 E4B-it's disease F1 to 0.71, above its own upper bound of 0.60. However, the verifier's letter-scale confidence score has the opposite of its intended effect: filtering to its most confident predictions does not improve accuracy and correlates negatively with correctness across every model and pool size tested, so the score cannot serve as a measure of predictive uncertainty, and most of the observed gain likely comes from rubric-grounded generation rather than pairwise verification.

Visual encoders construct a representation of the image input for Vision-Language models. How much conceptual, as opposed to immediately visible, information does this representation contain? We use canonical color as a controlled test case to ask whether vision encoders make canonical-color information linearly accessible, even when color is removed from the input image. We construct a dataset of objects with canonical colors, and probe vision encoders for both color and object identity using color and grayscale images. We find that canonical color remains decodable from grayscale images, and is tied to predicted object identity, indicating a conceptual link. Extending this analysis to full VLMs, we find that VLM post-training can have a surprisingly large effect on color decodability in the vision encoder. Overall, canonical color provides a usefully controllable lens for tracing object-level conceptual semantic information in vision encoders and VLMs.

Streaming video understanding requires Vision Language Models (VLLMs) to process growing video streams and answer user questions under tight latency constraints. Existing methods improve efficiency through token pruning and memory-bank schemes, but mainly reduce visual tokens after visual encoding. Consequently, downstream token pruning alone cannot substantially reduce end-to-end latency because the expensive frame encoding cost has already been incurred. We propose CoFiE, a Coarse-to-Fine Evidence Selection framework that decouples evidence selection into a coarse, query-agnostic filtering stage before the vision encoder and a fine, query-specific refinement stage during LLM prefill. CoFiE introduces Novelty-Guided Frame Filtering to retain visually distinctive candidate frames and Query-Specific Evidence Refinement to select the frames most relevant to the user query. This design removes substantial redundancy before frame encoding while preserving query-specific refinement once semantic information becomes available. Experiments show that CoFiE establishes a new state-of-the-art accuracy-efficiency trade-off across multiple video understanding benchmarks, reaching 78.86% accuracy on StreamingBench and 68.72% on OvO-Bench, with improvements of up to 3.15% over prior methods. Even with up to 80% evidence-frame filtering, CoFiE outperforms strong open-source multimodal models while improving end-to-end inference latency by up to 2.54 times.

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.

Joint-Embedding Predictive Architectures (JEPAs) for world modeling typically employ fixed-size Vision Transformer encoders that are over-provisioned for simple tasks and under-provisioned for complex ones, with significant redundancy across attention heads. We propose Successive Capacity Growth (SCG), a method that starts from a minimal encoder (1 head, 2 layers, 283K parameters) and grows incrementally in width (adding attention heads for low-level semantic capacity) or depth (adding transformer blocks for higher-order semantic abstraction), driven by a task-agnostic test-and-verify mechanism that exploits function-preserving expansion to safely trial architectural changes and roll back if they do not improve prediction loss. The Sketched Isotropic Gaussian Regularizer (SIGReg) ensures that all learned semantic dimensions remain statistically independent and aligned with the predictive objective, preventing collapse even as the architecture grows. On a 60-dimensional multi-object dynamics task, SCG naturally triggers depth expansion, improving prediction loss by 20.3% over the fixed small baseline with 56 times greater parameter efficiency than scaling to the fixed large model; on a 2D navigation task, a single width expansion yields even an 23% improvement over the fixed large model. Across all three tested environments of increasing complexity, the adaptive encoder matches or exceeds the fixed small baseline, with zero false-positive expansions and bit-exact function preservation (ratio = 1.0, absolute difference = 0.0). The take-away is that JEPA world model encoders need not be pre-allocated at maximum capacity - they can grow successively as the task demands, achieving significant compute and data efficiency while maintaining representation quality.

4D understanding and reasoning is a fundamental capability for embodied AI agents operating in dynamic physical environments. However, existing vision encoders are largely limited to static 2D images or 3D point clouds without temporal modeling, or to 2D videos that lack accurate geometric depth reasoning. Consequently, current approaches fail to jointly capture spatial structure and motion evolution in dynamic scenes. We present CL4D, the first foundational 4D vision encoder that directly operates on dynamic point clouds, trained with a contrastive learning objective to align spatio-temporal geometric representations with natural language descriptions. By learning a shared embedding space between text and 4D scene dynamics, CL4D enables zero-shot motion-to-text and text-to-motion retrieval in dynamic environments and serves as a foundational 4D vision encoder for downstream 4D vision-language tasks. Building on this encoder, we introduce 4DVLM, a 4D vision-language model that conditions language generation on dynamic geometric representations. 4DVLM is the first VLM designed to operate directly on 4D point clouds without relying on 2D images, 2D videos, or static 3D point clouds. We train CL4D and subsequently 4DVLM on a newly constructed dataset termed DynAction4D capturing diverse human motions across varying object interactions and scene environments. Extensive experiments across multiple 4D human action benchmarks demonstrate that CL4D achieves state-of-the-art performance, with improvements of approximately ~16.75% over prior methods. Furthermore, 4DVLM outperforms frontier video VLMs such as Gemini and GPT-5 even when these models are provided with RGB video sequences corresponding to the same scenes represented as 4D point clouds for 4DVLM.

Creative agents still lack an effective way to learn from high-quality human films, limiting their ability to produce cinematic-grade videos. A key challenge is the absence of a structured video representation that is both faithful to film content and directly usable for agentic reasoning and manipulation. To address the challenge, we propose the Agentic Video Auto-Encoder (AVA-Encoder), a framework for learning agent-native video representations via agentic auto-encoding. AVA-Encoder transforms a video into a knowledge graph (KG) representation and then reconstructs it back into video. Its hierarchy and state nodes store structured text, while a linked asset layer holds generated images, audio, and video. Typed edges preserve the relations between these text descriptions and assets in a form that agents can easily understand, query, and edit. The video reconstruction differences drive a textual-gradient optimization framework, which expresses evaluation feedback as natural-language update directions for Data-Independent Encoding Policy Pseudo-Training in the outer loop and optional Data-Dependent KG Representation Refinement in the test-time inner loop. Extensive experiments show that AVA-Encoder improves by 20.7 percentage points over the strongest external baseline. In the controlled policy-only setting, its pseudo-trained shot-level Agentic Video Encoder policy also outperforms a carefully human-tuned policy while using 74.3% fewer system-prompt tokens. We release the complete AVA-Encoder framework, a reliable agentic video reconstruction benchmark, and the first dataset of high-quality film KG representations.

The prevailing recipe for Vision-Language-Action (VLA) models couples a pretrained VLM with a separately trained flow-matching action expert. This makes the VLM a context encoder rather than a decision-maker. We introduce G0.5, a pretrained autoregressive VLA in which a single transformer decoder emits reasoning and action tokens under a single objective. Three components make this tractable at foundation-model scale: a learnable cross-embodiment action tokenizer that maps heterogeneous robot actions into a shared vocabulary; a native chain-of-thought stream interleaving task decomposition, object grounding, and action hints with action tokens; and a visual memory module that injects multi-second history through the vision encoder. Because reasoning and action share a single set of weights, the pretrained VLM's capabilities carry over to physical behavior: the model follows instructions closely, and prompts directly steer action granularity, task horizon, and out-of-distribution scene handling without further training. Pretrained on a large collection of robot datasets together with VQA samples, G0.5 surpasses state-of-the-art models across 7 independent regimes: real-world fine-tuning on R1lite and R1pro robots (76.7% vs.\ 53.3% for $π_{0.5}$ and 24.4% for GR00T-N1.7), the 2025 BEHAVIOR Challenge on 50 long-horizon household mobile manipulation tasks using a generalist policy (31.4% vs.\ 26.3% for $π_{0.5}$ and 26.1% for the challenge winner), DROID post-training followed by zero-shot transfer to an unseen environment and objects (82.5%), a language-following Pick-and-Place benchmark, LIBERO (98.9%), RoboTwin 2.0 (93.3%), and SimplerEnv-Bridge (87.3%).

Learned image compression has seen significant progress in recent years with the development of end-to-end learned models that achieve better compression efficiency than state-of-the-art conventional methods. Recently, Mixture of Experts (MoE) approaches have seen promising results in NLP and computer vision tasks. In this paper, we introduce the MoE approach to learned image compression. We propose a MoE-based Entropy model (MoEE) for learned image compression, allowing the model to selectively activate only the subset of parameters required for the input image. Our model achieves a BD-Rate improvement over VVC of -16.85% on the Kodak dataset.

Vision-language models can describe an image with remarkable accuracy, yet a more fundamental question remains unanswered: what visual information actually drives their answers? In this work, we investigate this question through causal tracing, and we observe that highly causal vision tokens often lie outside the target region. Extending the analysis to larger vision-language models reveals a similar pattern across models and corruption settings, suggesting that strong multimodal performance does not necessarily imply spatially localized causal representations. We further investigate: can these models preserve visual structure when appearance cues are removed? and find that visual cues are exploited to understand visual structures. Together, our experiments expose a gap between seeing, using, and reasoning over visual structure, and provide a causal framework for studying how visual information is transformed, preserved, and ultimately used by modern vision-language models.

Video coding is advancing into the low and ultra-low bitrate regime, driven by end-to-end codecs that replace the hand-crafted pipeline with jointly optimized neural networks and generative codecs that exploit the priors of video generation models. Yet the dominant metrics, LPIPS and DISTS, measure feature and texture similarity rather than content fidelity: a reconstruction that hallucinates a wrong face or blurs text into convincing strokes can still score well, even when a human rejects it instantly. To address this, we propose CodecArena, the first vision-language framework for video coding quality assessment, casting codec evaluation as source-conditioned comparative reasoning between a reference and its reconstructions. We optimize CodecArena with Facet-GRPO, a visual reinforcement learning scheme that aligns pairwise codec preferences while grounding the verdict in five fidelity facets: identity, objects, text, texture, and temporal consistency. Its facet-anchored reward uses automatically derived facet directions as weak anchors, rather than human per-facet labels, to prevent any single sub-score from dominating the holistic preference and to yield interpretable fine-grained quality judgments. To support training and evaluation in this underexplored regime, we construct two complementary resources: CodecArena-1K, a fully automatic preference dataset of 1,500 comparison groups built from traditional, neural, and generative codec reconstructions with fused vision-language and objective supervision; and CodecArena-Bench, a human-ranked benchmark with source-disjoint videos for fair out-of-domain evaluation. Extensive experiments demonstrate that CodecArena achieves state-of-the-art agreement with human judgments on source-disjoint content across diverse codecs and bitrates, surpassing perceptual metrics and prior vision-language evaluators.

Parkinson's disease (PD) is the second most common neurodegenerative disorder. Typical machine learning screening methods require PD labels, but the available data is limited by privacy concerns and the need for expert annotation. We propose a label-free face-plus-voice PD screen built entirely on frozen pretrained encoders--a face-expression Vision Transformer and HuBERT--in which no PD label touches any fit; the reference is training controls only. The voice modality uses a synthetic-dysarthria contrastive activation addition (CAA) direction built from time-stretch and breathy degradation of healthy speech; the face modality uses a k-nearest-neighbor anomaly score to the control embedding cluster. We introduce the alignment principle, a post-hoc analysis showing that a synthetic-degradation CAA detector works when the cosine similarity between the synthetic and real disease directions exceeds zero. Measured on the YouTubePD benchmark, this cosine is +0.37 for voice (CAA works, AUROC 0.765) and -0.48 for face (CAA fails; anomaly succeeds, AUROC 0.751). Equal-weight late fusion reaches AUROC 0.802 (95% CI [0.70,0.89]) with NPV 0.95, supporting a rule-out triage interpretation. An overfitting audit shows the voice detector transfers cleanly, while the face-side--and thus fused--AUROC is potentially optimistic pending external validation.

Semantic vision encoders have become a central visual interface for multimodal understanding and semantic conditioning in image generation. However, their final tokens discard fine-grained visual details, leading to poor pixel reconstruction and limiting their use in reconstruction-sensitive tasks such as image generation and editing. In this work, we ask whether understanding, generation, and editing can be modeled in a single visual representation space built from a pretrained semantic ViT. We show that the frozen Transformer blocks of a semantic ViT are not intrinsically unable to preserve visual details. Instead, the original patch parameterization drives the representation toward semantic abstraction, making fine-grained information difficult to recover from the final tokens. Based on this observation, we introduce \emph{Patch Reparameterization}, which preserves the original semantic pathway while adding a reconstruction-aware patch embedding that provides fine-grained visual information to the same frozen ViT blocks. The resulting unified representation preserves multimodal understanding while enabling high-fidelity image reconstruction and a favorable reconstruction--generation trade-off. We further scale this representation into \emph{UniSpace}, an 8B Mixture-of-Transformer-Experts model that performs understanding, generation, and editing in the same visual space without a separate VAE pathway. System-level evaluations demonstrate practical text-to-image generation and instruction-based image editing, showing that a reparameterized pretrained ViT can serve as a unified visual interface for scalable multimodal modeling.

The recent advances in neural language models have also spurred much work in computational psycholinguistics, asking whether neural LMs are also promising models of human language processing. However, work has been overwhelmingly focused on the unimodal case of written or spoken language. In contrast, multimodal experimental paradigms, like visual world studies that present participants with both visual and linguistic input simultaneously, have been neglected. In this paper, we present a novel approach that predicts gaze behavior in visual world studies. It does so by combining a simple multi-modal bi-encoder model of the CLIP family with a bimodal attribution method. We demonstrate the ability of this approach to robustly replicate the results of a seminal English visual world study which shows hu- man predictive processing. Remarkably, it does so without a generative architecture and without the need for fine-tuning, despite not being trained for this task.

Deep vision models exploit shortcuts, relying on cues that correlate with supervision signals. Prior work has focused on visible biases, such as object-background or texture correlations. We identify a different source of shortcut learning: invisible metadata traces embedded at the pixel level, for metadata such as image processing and photo acquisition. We hypothesize that large-scale semantic supervision, whether through categorical labels (ImageNet) or billion-scale captions (LAION), naturally induces metadata-semantics correlations during pretraining, leading models to convert low-level signals into predictive features. By introducing controlled metadata-semantics correlations, we show that stronger ones produce systematically higher sensitivity to metadata traces and larger performance degradation under metadata distribution shifts. We further explore mitigation strategies applied during and after pretraining that reduce sensitivity not only to targeted metadata but also to unseen ones, without sacrificing performance on downstream tasks. Metadata sensitivity also has a positive side: it partly explains the strong generated-image detection ability of some encoders, while its mitigation can improve out-of-distribution generalization. Code: https://github.com/ryan-caesar-ramos/visual-encoder-traces

Reachability analysis for visuomotor policies is difficult because large visual encoders make end-to-end set propagation computationally expensive and excessively conservative. We therefore freeze the visual encoder and confine set propagation to a low-dimensional interface between it and the downstream policy, with the interface set calibrated from held-out camera-pose perturbations. Propagating this set through the policy with zonotopes yields a terminal output-enclosure width that set-based training optimizes directly. During evaluation, camera-pose perturbations are sampled from the prescribed distribution, and rollout-level split conformal calibration converts the resulting action-deviation scores into a probabilistic reachable-action radius with finite-sample coverage. In controlled manipulation experiments, set-based training reduces this radius while preserving closed-loop task capability, and matched behavior-only, observational-consistency, and pointwise-adversarial controls all leave a larger radius.

Visual representations of VLA models remain unreliable for spatially precise robotic manipulation. We uncover that vision encoders in VLAs also exhibit attention artifacts previously documented in generic Vision Transformers, and further show that, in embodied policies, these artifacts are closely associated with spatial perception capabilities acquired during post-training. As the encoder learns task-relevant information such as object location, depth ordering, and local geometry, limited global-token capacity causes part of this information to spill into low-information patch tokens. We introduce AtVLA, a framework that inserts learnable register tokens into the visual encoder. Trained end-to-end using only embodied data and the original action objective, these registers emerge as dedicated carriers of embodied spatial information, while the remaining patch tokens recover clean and spatially faithful attention distributions crucial for precise target localization and fine-grained contact. Clean attention restores reliable localization, but cannot recover geometric details lost in low-resolution observations. AtVLA therefore couples attention rectification with uncertainty-gated local refinement. The action expert samples multiple action chunks and estimates uncertainty from their disagreement; only for uncertain predictions, action-conditioned attention rollout identifies the task-relevant region, which is cropped, re-encoded at high resolution, and appended to the cached prefix for refined action generation. Across LIBERO, SimplerEnv, and a challenging single-view real-world benchmark, AtVLA improves the average LIBERO success rate from 94.2% to 98.4% and real-world success from 46.5% to 69.0%. The cropping is triggered on approximately 30% of replanning steps, resulting in only 1.4-1.6x the total computation of the base model under the representative deployment setting.

Vision-language models (VLMs) remain unreliable when predictions require fine-grained visual evidence. We identify a previously overlooked cause: spectral response rigidity. Despite substantial frequency variation across images and tasks, pretrained vision encoders exhibit persistent, encoder-specific layerwise spectral profiles that change only marginally under downstream fine-tuning. Since pretrained vision encoders only receive images, they cannot adapt spectral extraction to the evidence required by the current query. We therefore propose HAFI-VLM, which introduces a task-conditioned frequency pathway while preserving the pretrained semantic representation. Hierarchical Adaptive Frequency Injection (HAFI) retrieves complementary low-, mid-, and high-frequency evidence at multiple encoder depths using text-modulated, spatially aligned cross-attention. A Visual Enrichment Layer Adapter further recalibrates shallow LLM attention to effectively utilize the enriched visual tokens. Experiments on LLaVA-1.5 and Qwen2.5-VL demonstrate consistent improvements in general VQA, text-rich understanding, and hallucination robustness, outperforming representation-level enhancement methods and most resolution- or cropping-based approaches without additional high-resolution encoding. Mechanistic analyses show that HAFI restores task-dependent spectral allocation while retaining semantic attention, establishing frequency enrichment as a distinct and effective route for improving VLM perception.

Vision-language-action models (VLAs), which leverage the cognition of multimodal information to infer physical-world actions, provide a generalized solution for embodied AI applications. Conventional VLAs usually concentrate on current digital cognition. While some efforts are made to enhance VLAs' reasoning capabilities by capturing temporal information, encoding the long-context history causes an efficiency-decreasing issue. To reconcile the conflict between capturing temporal information and maintaining inference efficiency in VLAs, this paper introduces FibVLA, an efficient framework featuring temporal perception of long-context history. Specifically, we leverage logarithmic hindsight sampling to both proprioceptive states and visual frames to capture long-term temporal dependencies with minimal redundancy. For the action expert, we introduce the flow matching to produce action distributions, and the Fibonacci recurrent inference strategy to generate long-range planning steps based on real-time closed-loop feedback. Experiments demonstrate that FibVLA significantly improves action smoothness and success rates without retraining large-scale visual encoders. Efficiency analysis demonstrates superior real-time responsiveness compared to video-based baselines in real-world evaluations.

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.

Model merging aims to combine multiple domain-specialized experts trained from a shared foundation model into a single multi-task model. Existing approaches largely focus on improving the merging procedure itself and typically assume experts obtained through full-parameter fine-tuning. In this work, we revisit expert training for model merging. We first show that prompt-based adaptation provides a strong baseline: independently learned prompts can be exploited across tasks while keeping the backbone fixed, avoiding the interference introduced by weight merging. Building on this observation, we introduce Dual-Tuned Experts (DTEs), a two-stage training strategy that first learns prompts and then fine-tunes the vision encoder. This reduces the magnitude of task-specific parameter updates and produces experts with higher merge compatibility. Experiments across multiple CLIP architectures, full fine-tuning, and LoRA experts show that DTEs consistently improve merged performance of standard merging approaches and remain effective even when combining heterogeneous sets of experts.

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.

Large Vision-Language Models (LVLMs) remain bottlenecked by massive computational footprints, precluding their deployment on resource-constrained edge devices. While efforts to compress LVLMs focus heavily on vision token reduction or smaller language models, the vision encoder is largely overlooked, typically deployed as a monolithic, computationally heavy feature extractor. Moreover, there is no previous effort that designs a vision encoder for LVLMs directly optimized for on-device latency. In this paper, we present UltraViT, a vision encoder for LVLMs, explicitly designed and optimized for on-device performance. Specifically, by taking into account real on-device latencies, we systematically design a pyramidal architecture that strategically integrates and adapts heterogeneous spatial mixers at the macro-block level. Furthermore, to pre-train UltraViT, we propose a novel two-stage generative pre-training strategy: cultivating rich spatial features via dense distillation, followed by direct generative supervision from a capacity-mixed frozen LLM. Compared to standard contrastive and SSL, we show that our pre-training is much more effective for achieving high-level semantic grounding for UltraViT needed for the subsequent generative multimodal alignment of LVLM training. Extensive experiments demonstrate that our on-device latency-informed design combined with our tailored training strategy establishes a new state-of-the-art for efficient LVLM encoding, significantly outperforming existing encoder-centric baselines while operating on-device at nearly 1.7xthe speed.

End-to-end OCR systems based on vision-language models have achieved strong performance in complex document OCR, but their efficiency is limited by the large number of visual tokens produced from document images. Many of these tokens correspond to blank margins or visually redundant regions, yet directly applying generic visual token compression methods may remove OCR-critical fine-grained details. In this paper, we propose LayoutLite, a lightweight plug-and-play module for efficient document OCR. Instead of relying on explicit document layout detection, LayoutLite performs implicit layout analysis at the token level between the vision encoder and the language decoder. It aggregates multi-layer visual representations from the vision encoder, and predicts an importance score for each visual token with a lightweight scoring network. Low-information tokens are then removed before entering the language decoder while preserving the original spatial positional information of retained tokens. To train LayoutLite without human annotations, we cast token selection as a reinforcement learning problem and optimize it with a group-relative policy optimization objective driven by OCR output consistency, together with an auxiliary layout supervision signal to stabilize training. Experiments on OmniDocBench demonstrate that LayoutLite can substantially reduce visual token length and inference cost with negligible degradation in recognition quality. We further evaluate LayoutLite on two OCR-specialized VLMs, FireRed-OCR and Logics-Parsing-V2. Under up to 50% token compression, LayoutLite preserves almost the same score on both models while reducing prefill latency, FLOPs, and KV cache memory by over 40%, with only a small additional inference overhead. These results show that token-level implicit layout analysis is an effective and practical approach for accelerating VLM-based OCR systems.

Electroencephalography (EEG) models used for epilepsy are often limited to specific datasets and tasks. This limited approach can make it challenging to apply these models across different datasets or in various situations. However, recent studies in foundation models and self-supervised learning suggest that an adaptable EEG backbone could support a range of EEG related tasks. In this study, we have developed a multimodal EEG foundation model that combines a raw signal encoder based on the Mamba architecture, a Vision Transformer (ViT)-style encoder for time-frequency data, and a lightweight encoder for text, all within a shared embedding space. The pretraining process relies on several innovative techniques, such as masked modeling, cross-view contrastive alignment, and temporal consistency losses. These methods are designed to create rich, seizure-relevant representations without requiring labeled data. To assess the efficacy and generalization of our pretrained model, we fine-tuned it on the canonical CHB-MIT seizure detection benchmark and additional seizure detection datasets, and conducted extensive experiments comparing different model variants. On the standard CHB-MIT split, our best single model achieved an AUROC of 0.874, and an ensemble variant reached 0.878 AUROC, representing state-of-the-art performance on this benchmark. In addition to standard train-test splits, we evaluated performance under a leave-one-subject-out (LOSO) protocol, which is rarely reported in prior EEG seizure modeling work and highlights the difficulty of patient-independent seizure detection, with a mean LOSO balanced accuracy of 0.558 across 19 subjects. Across datasets and evaluation settings, our multimodal foundation model enabled robust seizure detection and straightforward adaptation to new seizure detection scenarios, while also supporting interpretable seizure localization.

Large vision-language models are becoming increasingly dominant in 3D medical image interpretation, but we rarely know <i>which</i> internal units encode clinical findings or <i>where</i> that information lives in the representation. We first study this on a 3D chest vision-language model (Pillar-0) by probing its frozen vision embeddings. We show that (i) each radiological finding is encoded by a <i>sparse</i> set of ~10 vision-encoder channels that match full-feature classification performance and far exceed a zero-shot text prompting; (ii) turning off the channels tied to one finding, that finding's score collapses while unrelated labels stay stable; and (iii) the same sparse probe <i>replicates</i> on an architecturally unrelated 3D abdominal VLM (Merlin) suggesting a general property of frozen medical encoders. Our training-free concept channel probe (CCP) method, paired with a corpus-derived report template, outperforms published CT-CHAT on clinical efficacy and NLG metrics (F1 0.549 vs. 0.184; BLEU 0.483 vs. 0.373) at 22x lower latency. Our results provide a clear, reproducible characterization of how frozen medical encoders represent findings, demonstrating direct applicability across models.

Understanding instrument-tissue interactions is essential for context-aware surgical AI and autonomous robotic surgery. Pretrained vision-language models (VLMs) and vision encoders offer an alternative to conventional interaction classifiers by transferring broad visual and semantic knowledge. However, adapting them to fine-grained surgical interactions remains challenging: (1) freezing the vision encoder depends entirely on pretrained representations that may retain noise and provide weak spatial localization, while (2) full fine-tuning can improve global semantic alignment without ensuring that the encoder learns meaningful features in the correct action region. We address these limitations by introducing LAViFiT, an end-to-end latent-action-guided framework for vision-language fine-tuning. An inverse dynamics model captures the visual changes induced by each action, while a forward world model drives the encoder to represent action-relevant regions. A patch-level SIG Regularizer further prevents local feature collapse without additional supervision, such as bounding boxes or pseudo-labels. Experiments across multiple encoders and datasets improve recognition and image-text alignment, while representation analyses show stronger grounding over the complete instrument-tissue interaction region and more spatially coherent features.

Despite advances in Video Large Language Models (VLLMs) that have displayed promising outcomes in video understanding, the redundancy in the long-duration frames remains a hindrance to efficient reasoning. This paper introduces a training-free $\mathbf{P}$ersistence-Aware $\mathbf{C}$ompression and $\mathbf{A}$ggregation (PCA) method designed to preserve high-fidelity raw visual information before the encoding stage. PCA can be built on arbitrary VLLMs and consists of two modules: 1) A Dynamic Downsampling (DD) module that adaptively removes redundant frames by analyzing frame-wise similarity. 2) A Persistence-Aware Motion Enhancement (PAME) module that enriches each selected keyframe by aggregating the temporal context of its neighbors, ensuring that essential information is preserved even after aggressive frame reduction. Our approach substantially reduces the computation of long-context modeling, while enhancing the performance of the baseline model. Extensive experiments demonstrate that PCA consistently outperforms existing state-of-the-art approaches in both efficiency and accuracy, achieving a speedup of 1.8$\times$ to 2.5$\times$ compared to the baseline VLLM. The code is open-sourced at https://github.com/Heisenberg10110/PCA.

While Large Vision-Language Models (LVLMs), represented by LLaVA and GPT-4V, have demonstrated remarkable capabilities, their visual inputs remain vulnerable to adversarial attacks, posing significant security risks. Existing defense methods predominantly target single-task scenarios (e.g., zero-shot classification) and consequently lack generalizability across various multimodal tasks. To address this limitation, we propose a dual adversarial fine-tuning framework that jointly optimizes visual and semantic supervision signals from two modalities, enhancing model robustness while generalizing across multiple downstream tasks. The proposed framework comprises two core components, i.e., $\textbf{Visual}$ supervision branch and $\textbf{Semantic}$ supervision branch. The former branch leverages features from clean images, extracted via a frozen original vision encoder, to guide adversarial robustness while the latter incorporates caption-image alignment as a contextual signal to preserve semantic coherence under attack. Moreover, our method achieves cross-task robustness by simply replacing the CLIP vision encoder in the original model, with no need of separate task-specific retraining or architecture modifications.Extensive experiments demonstrate that our approach outperforms the state-of-the-art method in adversarial robustness evaluation across zero-shot classification, image captioning, and visual question answering (VQA) tasks.

Vision-Language Navigation (VLN) requires an embodied agent to interpret a natural-language instruction and predict actions from temporally ordered visual observations. Adapting a multimodal large language model to VLN requires visual-language alignment, compact temporal inputs, action-space grounding, and stable training on the target hardware. This technical report presents PGN (Pangu Navigator), an offline VLN action-prediction system built on OpenPangu-7B. Training proceeds in two stages. First, PGMM aligns a frozen EVA-ViT-G/14 vision encoder with the frozen language backbone by training a Q-Former and a two-layer MLP projector. Second, PGN adapts the aligned model to expert navigation trajectories using five-observation windows, epoch-dependent temporal sampling, and a reasoning-then-action output format; this stage freezes the aligned visual pathway and updates three structural-token embeddings and LoRA adapters. The implementation combines mixed-precision computation, selective FP32 computation, and DeepSpeed ZeRO-2 on eight Ascend 910B NPUs. Under teacher-forced, open-loop evaluation on 500 held-out expert trajectories, V9 reports a 62.29% Normalized Action Match (NAM) and a 100.00% Non-empty Rate (NER). These metrics quantify offline expert-action alignment rather than closed-loop navigation success; evaluating error accumulation, path efficiency, and goal completion remains future work.

Few-shot Test-Time Domain Adaptation (FSTT-DA) seeks to adapt models to novel domains using only a handful of unlabeled target samples. This setting is more realistic than typical domain adaptation setups, which assume access to target data during source training. However, prior FSTT-DA approaches fail to effectively leverage source domain-specific knowledge, relying on shallow batch normalization updates, prompt-based methods that treat the model as a black box, or ensembling strategies that do not capture cross-domain relationships. To address these limitations, we introduce a new FSTT-DA framework that integrates LoRA fine-tuning with model merging. In our approach, separate LoRA modules are fine-tuned on CLIP's vision encoder for each source domain. Since LoRA modifies only a small fraction of the model's parameters, it retains the base model's generalized knowledge while internally learning domain-specific features. To adapt the learned knowledge to a specific target domain, we propose a hypernetwork trained via meta-learning that generates per-column merging factors to combine LoRA modules. Given a small batch of target images, the hypernetwork produces merging weights that fuse source LoRA modules into a single adapted representation. Our results demonstrate state-of-the-art performance across various domain adaptation datasets. Our code is publicly available at https://github.com/nahbois4321/DA-MergeLoRA.

Understanding and characterizing human color perception is a longstanding research goal. One of the most traditional approaches is looking for the human color discrimination thresholds, the minimum chromatic differences perceptible to human observers. In recent years, deep neural networks have become the standard networks for computer vision tasks. In particular, deep vision encoders, foundation models trained on large-scale visual data, map images into latent feature representations. Despite the widespread use of deep vision encoders, few studies have investigated whether their internal representations exhibit human-like discrimination thresholds. In this work, we present a large-scale exploratory study probing the chromatic sensitivity of more than 50 pretrained vision encoders, including convolutional networks and vision transformers, against human discrimination thresholds. Using controlled chromatic stimuli at multiple chroma levels, we compare model-derived chromatic discrimination thresholds with human discrimination ellipses through a region-overlap metric (mIoU). Our analysis reveals generally weak alignment between model representations and human perceptual thresholds across all model families, with the best mIoU < 0.25. Moreover, we find that self-supervised encoders consistently outperform supervised ones, while language-supervised models show the most polarized behavior, occupying both the top and bottom of the ranking. These findings suggest that human-like chromatic sensitivity does not emerge naturally from current large-scale visual training objectives for any of the analyzed architectures.

A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.24031). This report keeps the hardware and asks what a model that fits can do: we deploy MiniCPM-V-4.6, a modern multimodal assistant pairing a SigLIP2 vision encoder and window-attention merger (16x visual token compression) with a compact hybrid gated-delta-net backbone, entirely on the GPU. Three results. (i) An all-GPU engine built on measured foundations: projections that dequantize 8-bit weights once and call the vendor SGEMM still in the last Fermi toolchain (64% of FP32 peak; our best hand-written GEMM hit 37%, wrongly called the ceiling); a chunked delta-rule rewrite of the recurrent layers, 2.8x faster than the sequential scan once attribution exposed one bad kernel; and a measured negative: 4-bit weights make decode slower than 8-bit here, since Fermi issues nibble-unpacking shifts at half rate. (ii) The vision side is a port with a proof obligation: we translate tower, merger, and projector to sm_20 CUDA, validating every stage against a locally generated reference forward (full tower 1.4e-5). One failure, position-embedding bucketization differing on exact rational ties, generalizes to a rule: float tie-breaking in index arithmetic is implementation-defined; call the reference operator, do not reimplement it. (iii) Long context exposes an O(N^2) wall short benchmarks hide: prefill falls from 114 tok/s at 2k tokens to 21 at 10k in a naive attention kernel; per-head vendor-GEMM calls writing into the existing score buffer (zero extra memory) restore a flat profile (408 at 2k, 361 at 10k; 17x), verified by exact needle retrieval from 60% depth. The same rewrite cuts image encoding 6x, to 0.93s. The system answers an image question end-to-end in 1.7s.

There's been a surge in adoption of video conferencing applications for both personal and business use cases. However, the bandwidth limitations faced by many users worldwide may restrict the optimal use of such applications. Although deep learning offers a solution for enhancing low bit rate videos, most models today are either hard to incorporate with modern compression standards or require specialized hardware to run such as significant GPUs making these models impractical. To address these issues, we introduce the Realtime Face Video Enhancement (RTFVE) model which can be easily incorporated with any video decoder and can run in realtime on ordinary CPUs. Experiments show that our model improves perceptual quality over the compressed video baseline at multiple low bitrate settings. The source code will be made available at https://github.com/varun-jois/RTFVE.

Deploying billion-parameter Vision-Language-Action (VLA) models on industrial hardware requires fine-tuning to bridge the embodiment gap. Full Fine-Tuning (FFT) provides maximal plasticity but requires data centre-grade GPUs. We present a systematic study of Low-Rank Adaptation (LoRA) for $π_0$, a flow-matching VLA, evaluated on four precision assembly tasks with a UR5e robotic manipulator. Across a sweep of LoRA ranks (r=8 to 256), allocation strategies, and component-freezing ablations, we find no statistically significant advantage of FFT over certain LoRA configurations. Performance saturates at r=32, and uniform allocation across the Vision-Language-Model (VLM) backbone and action expert proves sufficient. Freezing the VLM or restricting the vision encoder to LoRA significantly degrades performance, indicating that embodiment adaptation requires both semantic and visual plasticity. These results suggest that LoRA at r=32 with full vision encoder fine-tuning is a practical approach, reducing static peak VRAM from 36.2 to 10.8 GiB (parameters and optimizer states, activation memory excluded) without detectable performance loss.

Text-guided medical image segmentation leverages clinical semantics to improve lesion delineation, yet many existing models bind cross-modal fusion, supervision, and decoder design into a task-specific architecture. Such tight coupling makes it difficult to reuse language guidance modules across heterogeneous vision and text backbones, and often requires redesigning the network when the encoder pair changes. This paper presents BTHA, a backbone-transferable hierarchical adapter framework for text-guided medical image segmentation. BTHA is built around a stable feature-level interface: given multi-scale visual features and a text representation, it injects semantic guidance through shape-preserving adapters while maintaining the decoder-side tensor contract. To make this interface effective, we introduce a Hierarchical Coarse-to-Fine Supervision Strategy that decomposes learning into global image-text alignment, multi-scale auxiliary localization, and boundary-aware final mask refinement. We further design a Scale-Adaptive Gated Semantic Guidance (SAGSG) adapter, where resolution-specific gates adaptively control textual injection and channel recalibration suppresses redundant cross-modal responses. Evaluations across diverse vision and text backbones show that the same adapter and supervision design remains effective across convolutional and transformer-based visual encoders as well as different language encoders. Experiments on four public datasets further demonstrate that BTHA improves strong text-guided baselines with modest computational overhead.

Visual navigation policies built on large pretrained models have so far followed a common recipe: a dedicated visual encoder, a bespoke action head, and training on thousands of hours of cross-embodiment datasets. We ask whether this recipe is necessary. In this paper, we introduce GemNav, a visual robot navigation policy that adapts a frozen Multimodal Large Language Model (MLLM) for short-to-medium horizon waypoint navigation using Low-Rank Adaptation (LoRA) on the language tower alone, with no auxiliary visual encoder and no continuous regression head. Waypoints and categorical navigation signals share a single discrete token vocabulary generated by the language-model head, and a soft-decoded auxiliary loss recovers the metric structure that pure cross-entropy training discards. On a single 8.7-hour open corpus, roughly three orders of magnitude smaller than competing training sets, the policy transfers zero-shot to four physically distinct unseen environments and stops within 0.25-0.42m of the goal across 20 real-world trials covering an open carpark, an obstacle carpark, a long outdoor chemical yard, and an indoor warehouse. Conditioning on short image histories improves offline metrics but yields no robot benefit, pointing to a ceiling on what temporal context adds once pretrained vision features are in place. These results indicate that discrete-token adaptation of frozen MLLMs can provide a data-efficient, deployable alternative for foundation model robot navigation.

How can pretrained Vision-Language-Action (VLA) models retain long-horizon visual histories with high-frequency updates without sacrificing efficiency? Existing approaches rely on external memory management, which restrains either the memory horizon or the reactiveness of pretrained policies. To this end, we present NativeMEM, a VLA policy that features long-term and real-time updated memory. At its core is an efficient memory encoding scheme, Native Memory Compression, which repurposes the VLA's own vision encoder to compress each historical frame from each camera view into a single token. Appended to the input sequence, these memory tokens enable the pretrained VLA to attend over long-term history with negligible latency overhead, requiring neither an external planner nor a freshly initialized memory module. To align the memory tokens with the pretrained policy, we first develop a generic memory tokenizer under the supervision of a frozen VLA on memory-demanding data, and then unfreeze the VLA for task-specific fine-tuning. NativeMEM consistently outperforms prior methods, boosting success rates from 32.4% to 84.0% in simulation and up to 98.7% on real robots, while maintaining low inference latency and GPU memory usage. Notably, NativeMEM exhibits high data efficiency by achieving competitive results with prior arts using only 20% of the training data.

This report presents our solutions to the QoMEX 2026 Grand Challenge on Video Quality Assessment for Asymmetric Encoded Videos, comprising a full-reference (FR) model, CompressedVQA-AEV-FR, and a no-reference (NR) model, CompressedVQA-AEV-NR. The FR approach leverages a Swin-B backbone to extract multi-stage similarity statistics between reference and distorted videos for quality prediction. For the NR setting, our model employs complementary frame-level encoders based on SigLIP2 and Swin-B, followed by temporal mean pooling and cross-fold ensembling to estimate perceptual quality without reference data. Our CompressedVQA-AEV-FR achieves first place in the FR track of QoMEX 2026 Grand Challenge, while CompressedVQA-AEV-NR secures fourth place in the NR track, demonstrating the effectiveness of our proposed models. The code is available at https://github.com/sunwei925/CompressedVQA-AEV.

We present DroneIQA-VLE, our solution to the ICME 2026 Drone-IQA Grand Challenge on Target-aware Image Quality Assessment for Low-altitude UAV Images. The framework jointly predicts global, target, and background quality scores by ensembling two complementary pipelines: (1) SigLIP2 vision encoders with multi-task regression heads, and (2) a LoRA-adapted Qwen3.5-9B multimodal large language model for quality score regression. The final global quality prediction is obtained by arithmetically averaging the outputs of both pipelines. Our method achieves 2nd place in the challenge, demonstrating its effectiveness. The code is available at https://github.com/sunwei925/DroneIQA-VLE.

Camera pose estimation from image streams is a critical component of spatial world models that integrate perception into planning and decision-making. Nearly all Visual Odometry (VO) and Simultaneous Localization and Mapping (V-SLAM) systems have focused on datasets containing raw, uncompressed videos. Many working systems instead use ubiquitous hardware units to efficiently compress and decode video streams, saving orders of magnitude in storage and bandwidth. However, this lossy compression introduces visual artifacts that hinder the performance of traditional tracking systems. We present VOCA, a causal stereo visual-odometry method that exploits codec information to improve tracking performance. We achieve state-of-the-art performance on causal VO for relative trajectory error, efficiency, and absolute trajectory error on compressed streams. This work highlights the potential of leveraging widely available video codec information for vision tasks.

Adaptive video streaming is a widely used technique for delivering video content over the internet. One of the key challenges is determining the optimal encoding settings for each video, which can vary significantly based on its content and characteristics. In this paper, we propose Compression Echo Contrastive Learning (CECL), a novel self-supervised learning framework for clustering videos based on their encoding complexity. Our method leverages the response of a video to compression - the Compression Echo - as a supervisory signal, allowing the model to capture underlying encoding characteristics during pretraining. We conduct extensive experiments to demonstrate the effectiveness of our learned representations for the downstream task of clustering videos by their encoding complexity. Our results show that CECL improves upon existing state-of-the-art visual encoders and delivers strong bitrate and quality savings against the fixed bitrate ladder.

Vision-Language-Action (VLA) models have become a dominant paradigm for embodied intelligence. However, most existing approaches are built on large-scale transformers, resulting in substantial inference latency and energy consumption that limit their practical deployment in low-power, real-time scenarios. We propose SpikeVLA, a spiking VLA architecture for embodied navigation with energy-efficient inference, consisting of three key components. (i) A spiking vision encoder, Spike-V, that replaces dense continuous layers with event-driven spiking layers to reduce the energy consumption of visual representation learning. (ii) A multi-modal spiking large language model, Spike-L, that reformulates cross-modal reasoning with spiking dynamics and token-level event-driven sparsity to further lower computational cost. (iii) A spiking action policy network, Spike-A employs Laplacian-kernel population coding with a multi-layer fully connected SNN, and decodes spiking activities into stable and robust continuous control with energy-efficient inference under low-power constraints. Experiments on navigation and robotic control tasks show that SpikeVLA significantly reduces energy consumption and computational cost while maintaining competitive performance, highlighting its potential for low-power, real-time embodied intelligence.

A unified representation for text and vision is a natural pursuit, as it enables simpler multimodal modeling and more efficient training. However, representing images as discrete signals in the same way as text inevitably introduces severe information loss. Existing work struggles to balance low-level details and high-level semantics in discrete representations: reconstruction-oriented representations often lack semantic information, whereas semantically stronger features typically suffer from severe loss of detail. We present ViQ, a Visual Quantized Representations framework, which is designed to balance semantics and details in discrete representations while supporting inputs at native resolutions, thereby enabling it to serve as a unified and general discrete representation for arbitrary visual inputs. Our approach structures quantization learning into two stages: text-aligned pre-training and feature discretization. With text-aligned pre-training, we enhance the visual encoder semantic-rich supervision from the pretrained language model and enable it to process native-resolution visual inputs. During discretization, we propose a proximal representation learning strategy to progressively compact the feature space, along with a position-aware head-wise quantization mechanism that enables flexible processing of arbitrary resolutions. Extensive experiments on multimodal tasks demonstrate that ViQ achieves competitive performance compared to state-of-the-art multimodal vision encoders with continuous and high-dimensional visual features, while maintaining high precision in low-level reconstruction. We also show that multimodal training with visual quantized representations largely improves efficiency, yielding up to 20%-70% acceleration with different base LLMs and training recipes.

Self-supervised learning from large-scale video data has emerged as a dominant paradigm for visual representation learning. Since audio and visual streams naturally co-occur in video data, extending this success to jointly learn from both modalities is a natural next step, yet it remains challenging. Existing audio-visual self-supervised methods rely on modality-specific encoders and complex combinations of contrastive or reconstruction objectives, limiting cross-modal synergy and scalability. Joint Embedding Predictive Architectures (JEPAs) offer a simple, modality-agnostic alternative, but have to date been applied primarily to individual modalities. We introduce MJEPA, a joint-embedding predictive architecture for audio-visual learning that uses a single, unified encoder for both modalities. Our approach uses only a single predictive objective, applied both within and across modalities. We show that cross-modal prediction is critical: without it, a shared encoder degrades below unimodal baselines; with it, each modality's representation benefits from the other. Our frozen ViT-g model outperforms the best prior frozen baseline by over 6.8 mAP on AudioSet-20K, surpasses fully finetuned models on ESC-50 and FSD50K, and is competitive on video benchmarks despite using 10x less video data.

Representation alignment has emerged as an effective approach to improve Multimodal Large Language Models (MLLMs) by regularizing their internal representations toward those of an external vision encoder. However, existing methods typically align a fixed layer of the language backbone, overlooking the fine-grained structure of Transformer models. In this work, we propose Head-Wise Representation Alignment (HeRA), a method that enforces cross-modal alignment at the level of individual attention heads. Our approach is grounded in the Platonic Representation Hypothesis, focusing on preserving the topological structure of representations (i.e., their local neighborhood relationships) across modalities. Following the Mutual K-Nearest Neighbor (MKNN) alignment metric, we introduce a contrastive objective that acts as a differentiable proxy for matching local structures. HeRA applies this objective during multimodal training to specific attention heads in the LLM, selected by their alignment score according to the MKNN metric. Counterintuitively, we find that aligning the least aligned heads yields the largest gains. Extensive evaluations across multiple MLLMs and 18 benchmarks demonstrate that HeRA consistently improves performance on challenging vision-centric tasks and serves as an effective regularizer against visual hallucinations by naturally curbing the over-reliance on linguistic priors. Our code is publicly released.

This report presents our solution for the ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge, which requires parsing unstructured outdoor scenes from four camera platforms into 56 fine-grained categories. Our approach pairs foundation vision encoders (including DINOv3, SigLIP2, and InternImage) with a Mask2Former decoder, and trains them with a strong recipe including long training schedules, exponential moving average, a larger crop size, and multi-scale plus flip test-time augmentation. The three encoders, chosen for their complementary pretraining objectives, are combined into a pretraining-diverse ensemble through per-class validation-IoU weighting. Evaluated on the official GOOSE test set, our submission achieves 75.40% composite mIoU and wins the second place of the challenge. Our study further shows that the encoder's pretraining recipe, rather than its parameter count or the decoder design, is the dominant factor for accuracy on this benchmark.

Vision-language-action (VLA) models rely on discrete action tokenizers to bridge continuous robot control and autoregressive sequence modeling, yet existing tokenizers often trade off between compression, latency, and downstream performance. We revisit this design through the lens of neural audio codecs-convolutional encoder-decoder architectures with residual vector quantization that serve as the standard front end for audio foundation models. Motivated by their success, we introduce the Neural Action Codec (NAC), which treats short robot action trajectories as multi-channel 1D signals and compresses them using a multi-scale RVQGAN architecture. We observe that audio-specific mel-spectrogram objectives are ill-suited for kinematic signals; however, by replacing them with simple time-domain and non-mel spectral reconstruction losses, audio-codec-style models can autoencode actions with high fidelity without substantial architectural changes. NAC provides a compact, ordered token space via offset codebooks, enabling standard autoregressive policies to operate over short, structured sequences. Meanwhile, a Vocos-style decoder with an ISTFT head and adversarial discriminators recovers smooth, detailed trajectories. Across LIBERO-10, RoboMimic, and a suite of real-world manipulation tasks, NAC achieves lower reconstruction error and higher success rates than binning, FAST, and prior VQ-based tokenizers at comparable or better compression rates. These results demonstrate that repurposed neural audio codecs offer a strong, practical backbone for learned action tokenization in modern VLAs.

Egocentric video understanding is inherently limited by the narrow perspective of wearable cameras: a single viewpoint, a single modality, a single model cannot capture the full richness of human action. We argue that a truly expressive egocentric representation must subsume complementary knowledge across viewpoints, modalities, and foundation model representations, yet remain deployable from egocentric video alone. To this end, we introduce a hierarchical multi-teacher distillation framework that produces UNIEGO, a unified egocentric encoder trained with nine teachers spanning ego-exo viewpoints, RGB, depth, and skeleton modalities, and four foundation models. Rather than distilling directly from heterogeneous teachers whose incompatible architectures and feature geometries induce conflicting gradients, our framework interposes a layer of representation-specific Proxy models that translate diverse teacher knowledge into a homogeneous egocentric space. A second distillation stage, Selective Proxy Distillation (SPD), then adaptively selects, for each training sample, the subset of proxies that are both correct and confident, distilling exclusively from reliable supervision and suppressing erroneous signals. SPD is further stabilized by initializing UNIEGO as a learned convex combination of proxy parameters, placing the unified model in a well-conditioned region of the loss landscape before distillation begins. UNIEGO achieves state-of-the-art performance across three egocentric video understanding tasks - action recognition, video retrieval, and action segmentation on three challenging ego-exo benchmarks, outperforming naive multi-teacher distillation baselines and demonstrating that structured, proxy-mediated knowledge transfer yields richer and more discriminative egocentric representations.

Vision foundation models are typically trained as static feature extractors, placing the burden of task adaptation onto large downstream models. We propose an alternative paradigm: instead of solely feeding visual features into language models, we use language itself to dynamically guide the vision encoder. Our method, Language-Instructed Vision Embeddings (LIVE), leverages language as high-level guidance to produce task-centric embeddings at inference time, removing the need for task-specific retraining. This enables the encoder to focus on contextually relevant aspects of the input, yielding more controllable and generalizable representations. Empirically, LIVE reduces visual hallucinations (+34 points on MMVP), surpasses vision-language models with orders of magnitude more parameters on visual question answering, and generalizes to unseen instructions and tasks -- offering a direct path toward adaptive, instruction-driven visual intelligence.

Existing vision encoders for robotics face a fundamental bottleneck: robotic datasets lack the scale necessary for large-scale pre-training. Prior work circumvents this data scarcity by turning to internet-scale image and language data or egocentric human video. While these models show promise, neither paradigm learns from paired vision and action data, which downstream visuomotor control policies require. However, robot trajectories, the most direct source of this paired signal, are not available at pre-training scale, motivating us to extract action signals from abundant human video instead. To this end, we introduce CAIP (Contrastive Action-Image Pre-training), a vision encoder that treats human hand poses from large-scale egocentric video as a proxy for end-effector actions. By extracting 3D hand keypoints, a representation that aligns naturally with downstream robot action spaces, CAIP learns a unified action-image representation through a contrastive objective. Leveraging 32,041 hours of egocentric human video and only 88 hours of robotic manipulation data, CAIP outperforms state-of-the-art vision encoders including DINOv2, SigLIP, MVP, and R3M. Evaluated on a challenging real-world dexterous manipulation setup using Dexmate Vega and Sharpa Wave hands, CAIP yields performance gains of more than 30% on tasks involving folding, pouring, and fine-grained manipulation. Our results show that our method of contrastive action-centric pre-training yields a scalable path to achieving robust visual representations better suited for physical interaction.

Vision-language-action (VLA) models increasingly rely on high-frequency multi-camera observations, making visual communication a major bottleneck for real-time robotic control in bandwidth-constrained or distributed deployment settings. Existing image and video codecs, however, are designed to preserve generic visual fidelity rather than the control performance of downstream VLA policies. In this work, we introduce SPARC (SPatially Adaptive Rate Control), a learned image compression framework tailored for VLA-driven robots. Our key observation is that the importance of visual information varies substantially across both camera views and spatial regions within an image. Based on this observation, SPARC employs a lightweight temporal mask selector that adaptively allocates bitrate over latent representations according to task relevance while leveraging temporal context. We further introduce a tilted rate loss that stabilizes training by reducing the tendency of entropy-based objectives to over-suppress rare yet task-critical visual patterns. Experiments on diverse robotic benchmarks, including RoboCasa365, VLABench, and LIBERO, show that SPARC consistently achieves stronger control performance than conventional image/video codecs and recent learned compression methods under the same bitrate budget. We additionally demonstrate real-world deployment benefits in remote-control settings, where our method substantially improves the bitrate-success tradeoff.