Xiaoke Huang

Wanhua Li, Xiaoke Huang, Zheng Zhu et al. · tsinghua

This paper presents a language-powered paradigm for ordinal regression. Existing methods usually treat each rank as a category and employ a set of weights to learn these concepts. These methods are easy to overfit and usually attain unsatisfactory performance as the learned concepts are mainly derived from the training set. Recent large pre-trained vision-language models like CLIP have shown impressive performance on various visual tasks. In this paper, we propose to learn the rank concepts from the rich semantic CLIP latent space. Specifically, we reformulate this task as an image-language matching problem with a contrastive objective, which regards labels as text and obtains a language prototype from a text encoder for each rank. While prompt engineering for CLIP is extremely time-consuming, we propose OrdinalCLIP, a differentiable prompting method for adapting CLIP for ordinal regression. OrdinalCLIP consists of learnable context tokens and learnable rank embeddings; The learnable rank embeddings are constructed by explicitly modeling numerical continuity, resulting in well-ordered, compact language prototypes in the CLIP space. Once learned, we can only save the language prototypes and discard the huge language model, resulting in zero additional computational overhead compared with the linear head counterpart. Experimental results show that our paradigm achieves competitive performance in general ordinal regression tasks, and gains improvements in few-shot and distribution shift settings for age estimation. The code is available at https://github.com/xk-huang/OrdinalCLIP.

Haixing Dai, Zhengliang Liu, Wenxiong Liao et al.

Text data augmentation is an effective strategy for overcoming the challenge of limited sample sizes in many natural language processing (NLP) tasks. This challenge is especially prominent in the few-shot learning scenario, where the data in the target domain is generally much scarcer and of lowered quality. A natural and widely-used strategy to mitigate such challenges is to perform data augmentation to better capture the data invariance and increase the sample size. However, current text data augmentation methods either can't ensure the correct labeling of the generated data (lacking faithfulness) or can't ensure sufficient diversity in the generated data (lacking compactness), or both. Inspired by the recent success of large language models, especially the development of ChatGPT, which demonstrated improved language comprehension abilities, in this work, we propose a text data augmentation approach based on ChatGPT (named AugGPT). AugGPT rephrases each sentence in the training samples into multiple conceptually similar but semantically different samples. The augmented samples can then be used in downstream model training. Experiment results on few-shot learning text classification tasks show the superior performance of the proposed AugGPT approach over state-of-the-art text data augmentation methods in terms of testing accuracy and distribution of the augmented samples.

Hongmin Cai, Wenxiong Liao, Zhengliang Liu et al. · harvard

Modern supervised learning neural network models require a large amount of manually labeled data, which makes the construction of domain-specific knowledge graphs time-consuming and labor-intensive. In parallel, although there has been much research on named entity recognition and relation extraction based on distantly supervised learning, constructing a domain-specific knowledge graph from large collections of textual data without manual annotations is still an urgent problem to be solved. In response, we propose an integrated framework for adapting and re-learning knowledge graphs from one coarse domain (biomedical) to a finer-define domain (oncology). In this framework, we apply distant-supervision on cross-domain knowledge graph adaptation. Consequently, no manual data annotation is required to train the model. We introduce a novel iterative training strategy to facilitate the discovery of domain-specific named entities and triples. Experimental results indicate that the proposed framework can perform domain adaptation and construction of knowledge graph efficiently.

Xiaoke Huang, Yiji Cheng, Yansong Tang et al. · tsinghua

Efficiently digitizing high-fidelity animatable human avatars from videos is a challenging and active research topic. Recent volume rendering-based neural representations open a new way for human digitization with their friendly usability and photo-realistic reconstruction quality. However, they are inefficient for long optimization times and slow inference speed; their implicit nature results in entangled geometry, materials, and dynamics of humans, which are hard to edit afterward. Such drawbacks prevent their direct applicability to downstream applications, especially the prominent rasterization-based graphic ones. We present EMA, a method that Efficiently learns Meshy neural fields to reconstruct animatable human Avatars. It jointly optimizes explicit triangular canonical mesh, spatial-varying material, and motion dynamics, via inverse rendering in an end-to-end fashion. Each above component is derived from separate neural fields, relaxing the requirement of a template, or rigging. The mesh representation is highly compatible with the efficient rasterization-based renderer, thus our method only takes about an hour of training and can render in real-time. Moreover, only minutes of optimization is enough for plausible reconstruction results. The disentanglement of meshes enables direct downstream applications. Extensive experiments illustrate the very competitive performance and significant speed boost against previous methods. We also showcase applications including novel pose synthesis, material editing, and relighting. The project page: https://xk-huang.github.io/ema/.

Wenxiong Liao, Zhengliang Liu, Haixing Dai et al.

Background: Large language models such as ChatGPT are capable of generating grammatically perfect and human-like text content, and a large number of ChatGPT-generated texts have appeared on the Internet. However, medical texts such as clinical notes and diagnoses require rigorous validation, and erroneous medical content generated by ChatGPT could potentially lead to disinformation that poses significant harm to healthcare and the general public. Objective: This research is among the first studies on responsible and ethical AIGC (Artificial Intelligence Generated Content) in medicine. We focus on analyzing the differences between medical texts written by human experts and generated by ChatGPT, and designing machine learning workflows to effectively detect and differentiate medical texts generated by ChatGPT. Methods: We first construct a suite of datasets containing medical texts written by human experts and generated by ChatGPT. In the next step, we analyze the linguistic features of these two types of content and uncover differences in vocabulary, part-of-speech, dependency, sentiment, perplexity, etc. Finally, we design and implement machine learning methods to detect medical text generated by ChatGPT. Results: Medical texts written by humans are more concrete, more diverse, and typically contain more useful information, while medical texts generated by ChatGPT pay more attention to fluency and logic, and usually express general terminologies rather than effective information specific to the context of the problem. A BERT-based model can effectively detect medical texts generated by ChatGPT, and the F1 exceeds 95%.

Wenxiong Liao, Zhengliang Liu, Haixing Dai et al.

Transformer-based language models have achieved significant success in various domains. However, the data-intensive nature of the transformer architecture requires much labeled data, which is challenging in low-resource scenarios (i.e., few-shot learning (FSL)). The main challenge of FSL is the difficulty of training robust models on small amounts of samples, which frequently leads to overfitting. Here we present Mask-BERT, a simple and modular framework to help BERT-based architectures tackle FSL. The proposed approach fundamentally differs from existing FSL strategies such as prompt tuning and meta-learning. The core idea is to selectively apply masks on text inputs and filter out irrelevant information, which guides the model to focus on discriminative tokens that influence prediction results. In addition, to make the text representations from different categories more separable and the text representations from the same category more compact, we introduce a contrastive learning loss function. Experimental results on public-domain benchmark datasets demonstrate the effectiveness of Mask-BERT.

Junqi Liu, Salena Song, Yuhan Wang et al.

Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spanning five research tracks: segmentation, image enhancement, visual question answering (VQA), report generation, and lesion detection. Each task is evaluated under two difficulty tiers, Lite and Standard, which use the same data and metrics but differ in the amount of task-brief scaffolding, and each run is scored using both final task performance and S1-S5 stage scores, enabling stage-level analysis from the initial task brief to the final submitted artifact. Across thousands of recorded runs, stage-level scoring reveals that Validate is the weakest workflow stage on average, whereas Setup is the strongest, suggesting that current agents are better at making pipelines executable than at verifying their reliability. Post-run error analysis further shows that verification and submission failures dominate tagged errors, accounting for 37.7% and 38.1% of fired codes respectively, whereas task-understanding errors are rare at 0.9%, and runs with one fired error code have a 48% lower overall score than runs with no error code on average.

Hongmin Cai, Xiaoke Huang, Zhengliang Liu et al.

Alzheimer's disease (AD) is a common form of dementia that severely impacts patient health. As AD impairs the patient's language understanding and expression ability, the speech of AD patients can serve as an indicator of this disease. This study investigates various methods for detecting AD using patients' speech and transcripts data from the DementiaBank Pitt database. The proposed approach involves pre-trained language models and Graph Neural Network (GNN) that constructs a graph from the speech transcript, and extracts features using GNN for AD detection. Data augmentation techniques, including synonym replacement, GPT-based augmenter, and so on, were used to address the small dataset size. Audio data was also introduced, and WavLM model was used to extract audio features. These features were then fused with text features using various methods. Finally, a contrastive learning approach was attempted by converting speech transcripts back to audio and using it for contrastive learning with the original audio. We conducted intensive experiments and analysis on the above methods. Our findings shed light on the challenges and potential solutions in AD detection using speech and audio data.

Yiji Cheng, Fei Yin, Xiaoke Huang et al.

Text-to-3D is an emerging task that allows users to create 3D content with infinite possibilities. Existing works tackle the problem by optimizing a 3D representation with guidance from pre-trained diffusion models. An apparent drawback is that they need to optimize from scratch for each prompt, which is computationally expensive and often yields poor visual fidelity. In this paper, we propose DreamPortrait, which aims to generate text-guided 3D-aware portraits in a single-forward pass for efficiency. To achieve this, we extend Score Distillation Sampling from datapoint to distribution formulation, which injects semantic prior into a 3D distribution. However, the direct extension will lead to the mode collapse problem since the objective only pursues semantic alignment. Hence, we propose to optimize a distribution with hierarchical condition adapters and GAN loss regularization. For better 3D modeling, we further design a 3D-aware gated cross-attention mechanism to explicitly let the model perceive the correspondence between the text and the 3D-aware space. These elaborated designs enable our model to generate portraits with robust multi-view semantic consistency, eliminating the need for optimization-based methods. Extensive experiments demonstrate our model's highly competitive performance and significant speed boost against existing methods.

Zhaochong An, Menglin Jia, Haonan Qiu et al.

Storytelling in real-world videos often unfolds through multiple shots -- discontinuous yet semantically connected clips that together convey a coherent narrative. However, existing multi-shot video generation (MSV) methods struggle to effectively model long-range cross-shot context, as they rely on limited temporal windows or single keyframe conditioning, leading to degraded performance under complex narratives. In this work, we propose OneStory, enabling global yet compact cross-shot context modeling for consistent and scalable narrative generation. OneStory reformulates MSV as a next-shot generation task, enabling autoregressive shot synthesis while leveraging pretrained image-to-video (I2V) models for strong visual conditioning. We introduce two key modules: a Frame Selection module that constructs a semantically-relevant global memory based on informative frames from prior shots, and an Adaptive Conditioner that performs importance-guided patchification to generate compact context for direct conditioning. We further curate a high-quality multi-shot dataset with referential captions to mirror real-world storytelling patterns, and design effective training strategies under the next-shot paradigm. Finetuned from a pretrained I2V model on our curated 60K dataset, OneStory achieves state-of-the-art narrative coherence across diverse and complex scenes in both text- and image-conditioned settings, enabling controllable and immersive long-form video storytelling.

Zhiheng Liu, Weiming Ren, Haozhe Liu et al.

Unified multimodal models (UMMs) aim to jointly perform multimodal understanding and generation within a single framework. We present TUNA, a native UMM that builds a unified continuous visual representation by cascading a VAE encoder with a representation encoder. This unified representation space allows end-to-end processing of images and videos for both understanding and generation tasks. Compared to prior UMMs with decoupled representations, TUNA's unified visual space avoids representation format mismatches introduced by separate encoders, outperforming decoupled alternatives in both understanding and generation. Moreover, we observe that stronger pretrained representation encoders consistently yield better performance across all multimodal tasks, highlighting the importance of the representation encoder. Finally, in this unified setting, jointly training on both understanding and generation data allows the two tasks to benefit from each other rather than interfere. Our extensive experiments on multimodal understanding and generation benchmarks show that TUNA achieves state-of-the-art results in image and video understanding, image and video generation, and image editing, demonstrating the effectiveness and scalability of its unified representation design.

Zijian Zhou, Shikun Liu, Haozhe Liu et al.

Reference-to-video (R2V) generation aims to synthesize videos that align with a text prompt while preserving the subject identity from reference images. However, current R2V methods are hindered by the reliance on explicit reference image-video-text triplets, whose construction is highly expensive and difficult to scale. We bypass this bottleneck by introducing Saber, a scalable zero-shot framework that requires no explicit R2V data. Trained exclusively on video-text pairs, Saber employs a masked training strategy and a tailored attention-based model design to learn identity-consistent and reference-aware representations. Mask augmentation techniques are further integrated to mitigate copy-paste artifacts common in reference-to-video generation. Moreover, Saber demonstrates remarkable generalization capabilities across a varying number of references and achieves superior performance on the OpenS2V-Eval benchmark compared to methods trained with R2V data.

Xiaoke Huang, Juncheng Wu, Hui Liu et al.

Large Reasoning Models (LRMs) have introduced a new paradigm in AI by enabling models to ``think before responding" via chain-of-thought reasoning. However, the absence of open and reproducible recipes for building reasoning-centric medical LMMs hinders community-wide research, analysis, and comparison. In this paper, we present MedVLThinker, a suite of simple yet strong baselines. Our fully open recipe consists of: (1) systematic data curation for both text-only and image-text medical data, filtered according to varying levels of reasoning difficulty, and (2) two training paradigms: Supervised Fine-Tuning (SFT) on distilled reasoning traces and Reinforcement Learning with Verifiable Rewards (RLVR) based on final answer correctness. Across extensive experiments on the Qwen2.5-VL model family (3B, 7B) and six medical QA benchmarks, we find that RLVR consistently and significantly outperforms SFT. Additionally, under the RLVR framework, a key, counter-intuitive finding is that training on our curated text-only reasoning data provides a more substantial performance boost than training on multimodal image-text data. Our best open 7B model, trained using the RLVR recipe on text-only data, establishes a new state-of-the-art on existing public VQA benchmarks, surpassing all previous open-source medical LMMs. Furthermore, scaling our model to 32B achieves performance on par with the proprietary GPT-4o. We release all curated data, models, and code to provide the community with a strong, open foundation for future research in multimodal medical reasoning.

Xiaoke Huang, Bhavul Gauri, Kam Woh Ng et al.

Vector glyphs are the atomic units of digital typography, yet most learning-based pipelines still depend on carefully curated exemplar sheets and raster-to-vector postprocessing, which limits accessibility and editability. We introduce VecGlypher, a single multimodal language model that generates high-fidelity vector glyphs directly from text descriptions or image exemplars. Given a style prompt, optional reference glyph images, and a target character, VecGlypher autoregressively emits SVG path tokens, avoiding raster intermediates and producing editable, watertight outlines in one pass. A typography-aware data and training recipe makes this possible: (i) a large-scale continuation stage on 39K noisy Envato fonts to master SVG syntax and long-horizon geometry, followed by (ii) post-training on 2.5K expert-annotated Google Fonts with descriptive tags and exemplars to align language and imagery with geometry; preprocessing normalizes coordinate frames, canonicalizes paths, de-duplicates families, and quantizes coordinates for stable long-sequence decoding. On cross-family OOD evaluation, VecGlypher substantially outperforms both general-purpose LLMs and specialized vector-font baselines for text-only generation, while image-referenced generation reaches a state-of-the-art performance, with marked gains over DeepVecFont-v2 and DualVector. Ablations show that model scale and the two-stage recipe are critical and that absolute-coordinate serialization yields the best geometry. VecGlypher lowers the barrier to font creation by letting users design with words or exemplars, and provides a scalable foundation for future multimodal design tools.

Zhaofang Qian, Hardy Chen, Zeyu Wang et al.

Vision-language models (VLMs) have advanced rapidly, yet their capacity for image-grounded geolocation in open-world conditions, a task that is challenging and of demand in real life, has not been comprehensively evaluated. We present EarthWhere, a comprehensive benchmark for VLM image geolocation that evaluates visual recognition, step-by-step reasoning, and evidence use. EarthWhere comprises 810 globally distributed images across two complementary geolocation scales: WhereCountry (i.e., 500 multiple-choice question-answering, with country-level answer and panoramas) and WhereStreet (i.e., 310 fine-grained street-level identification tasks requiring multi-step reasoning with optional web search). For evaluation, we adopt the final-prediction metrics: location accuracies within k km (Acc@k) for coordinates and hierarchical path scores for textual localization. Beyond this, we propose to explicitly score intermediate reasoning chains using human-verified key visual clues and a Shapley-reweighted thinking score that attributes credit to each clue's marginal contribution. We benchmark 13 state-of-the-art VLMs with web searching tools on our EarthWhere and report different types of final answer accuracies as well as the calibrated model thinking scores. Overall, Gemini-2.5-Pro achieves the best average accuracy at 56.32%, while the strongest open-weight model, GLM-4.5V, reaches 34.71%. We reveal that web search and reasoning do not guarantee improved performance when visual clues are limited, and models exhibit regional biases, achieving up to 42.7% higher scores in certain areas than others. These findings highlight not only the promise but also the persistent challenges of models to mitigate bias and achieve robust, fine-grained localization. We open-source our benchmark at https://github.com/UCSC-VLAA/EarthWhere.

Sucheng Ren, Xiaoke Huang, Xianhang Li et al.

This study presents Medical Vision Generalist (MVG), the first foundation model capable of handling various medical imaging tasks -- such as cross-modal synthesis, image segmentation, denoising, and inpainting -- within a unified image-to-image generation framework. Specifically, MVG employs an in-context generation strategy that standardizes the handling of inputs and outputs as images. By treating these tasks as an image generation process conditioned on prompt image-label pairs and input images, this approach enables a flexible unification of various tasks, even those spanning different modalities and datasets. To capitalize on both local and global context, we design a hybrid method combining masked image modeling with autoregressive training for conditional image generation. This hybrid approach yields the most robust performance across all involved medical imaging tasks. To rigorously evaluate MVG's capabilities, we curated the first comprehensive generalist medical vision benchmark, comprising 13 datasets and spanning four imaging modalities (CT, MRI, X-ray, and micro-ultrasound). Our results consistently establish MVG's superior performance, outperforming existing vision generalists, such as Painter and LVM. Furthermore, MVG exhibits strong scalability, with its performance demonstrably improving when trained on a more diverse set of tasks, and can be effectively adapted to unseen datasets with only minimal task-specific samples. The code is available at \url{https://github.com/OliverRensu/MVG}.

Wanhua Li, Xiaoke Huang, Jiwen Lu et al.

Uncertainty is the only certainty there is. Modeling data uncertainty is essential for regression, especially in unconstrained settings. Traditionally the direct regression formulation is considered and the uncertainty is modeled by modifying the output space to a certain family of probabilistic distributions. On the other hand, classification based regression and ranking based solutions are more popular in practice while the direct regression methods suffer from the limited performance. How to model the uncertainty within the present-day technologies for regression remains an open issue. In this paper, we propose to learn probabilistic ordinal embeddings which represent each data as a multivariate Gaussian distribution rather than a deterministic point in the latent space. An ordinal distribution constraint is proposed to exploit the ordinal nature of regression. Our probabilistic ordinal embeddings can be integrated into popular regression approaches and empower them with the ability of uncertainty estimation. Experimental results show that our approach achieves competitive performance. Code is available at https://github.com/Li-Wanhua/POEs.

Zhiheng Liu, Weiming Ren, Xiaoke Huang et al.

Unified multimodal models typically rely on pretrained vision encoders and use separate visual representations for understanding and generation, creating misalignment between the two tasks and preventing fully end-to-end optimization from raw pixels. We introduce Tuna-2, a native unified multimodal model that performs visual understanding and generation directly based on pixel embeddings. Tuna-2 drastically simplifies the model architecture by employing simple patch embedding layers to encode visual input, completely discarding the modular vision encoder designs such as the VAE or the representation encoder. Experiments show that Tuna-2 achieves state-of-the-art performance in multimodal benchmarks, demonstrating that unified pixel-space modelling can fully compete with latent-space approaches for high-quality image generation. Moreover, while the encoder-based variant converges faster in early pretraining, Tuna-2's encoder-free design achieves stronger multimodal understanding at scale, particularly on tasks requiring fine-grained visual perception. These results show that pretrained vision encoders are not necessary for multimodal modelling, and end-to-end pixel-space learning offers a scalable path toward stronger visual representations for both generation and perception.

Xiaoke Huang, Juncheng Wu, Hui Liu et al.

Test-time scaling has emerged as a powerful technique for enhancing the reasoning capabilities of large language models. However, its effectiveness in medical reasoning remains uncertain, as the medical domain fundamentally differs from mathematical tasks in terms of knowledge representation and decision-making processes. In this paper, we provide the first comprehensive investigation of test-time scaling for medical reasoning and present m1, a simple yet effective approach that increases a model's medical reasoning capability at inference. Our evaluation across diverse medical tasks demonstrates that test-time scaling consistently enhances medical reasoning, enabling lightweight fine-tuned models under 10B parameters to establish new state-of-the-art performance, while our 32B model rivals previous 70B-scale medical LLMs. However, we identify an optimal reasoning token budget of approximately 4K, beyond which performance may degrade due to overthinking. Budget forcing, which extends test-time computation through iterative prompts, helps models double-check answers but does not necessarily improve the overall medical QA performance and, in some cases, even introduces errors into previously correct responses. Our case-by-case analysis identifies insufficient medical knowledge as a key bottleneck that prevents further performance gains through test-time scaling. We find that increasing data scale, improving data quality, and expanding model capacity consistently enhance medical knowledge grounding, enabling continued performance improvements, particularly on challenging medical benchmarks where smaller models reach saturation. These findings underscore fundamental differences between medical and mathematical reasoning in LLMs, highlighting that enriched medical knowledge, other than increased reasoning depth alone, is essential for realizing the benefits of test-time scaling.

Juncheng Wu, Sheng Liu, Haoqin Tu et al.

Recent advances in reasoning-enhanced Large Language Models such as OpenAI-o1/3 and DeepSeek-R1 have significantly improved performance on complex tasks. However, the quality and transparency of their internal reasoning processes remain underexplored. This work moves beyond the final-answer accuracy and investigates step-by-step reasoning in the medical and mathematical domains by explicitly decomposing the thinking trajectories into two parts: knowledge and reasoning. Specifically, we introduce a fine-grained evaluation framework that judges: (1) the correctness of knowledge used (measured by Knowledge Index (KI)) and (2) the quality of reasoning (measured by Information Gain (InfoGain)). Using this framework, we study R1-distilled and base Qwen models trained with supervised fine-tuning (SFT) and/or reinforcement learning (RL) in the medical and math domains. Three intriguing findings emerge: (1) The general reasoning abilities in R1-distilled models do not transfer effectively to the medical domain through either SFT or RL. (2) SFT raises final-answer accuracy in both domains, but often at the cost of reasoning quality: InfoGain drops by 38.9% on average compared with untrained models; In the medical domain, however, SFT remains crucial because domain knowledge is indispensable. (3) RL enhances medical reasoning by pruning inaccurate or irrelevant knowledge from reasoning paths, thereby improving both reasoning accuracy and knowledge correctness.

Xiaoke Huang, Ningsen Wang, Hui Liu et al.

Large Multimodal Models (LMMs) are increasingly capable of answering medical questions that require joint reasoning over images and text, yet training general medical VQA systems is impeded by the lack of large, openly usable, high-quality corpora. We present MedVLSynther, a rubric-guided generator-verifier framework that synthesizes high-quality multiple-choice VQA items directly from open biomedical literature by conditioning on figures, captions, and in-text references. The generator produces self-contained stems and parallel, mutually exclusive options under a machine-checkable JSON schema; a multi-stage verifier enforces essential gates (self-containment, single correct answer, clinical validity, image-text consistency), awards fine-grained positive points, and penalizes common failure modes before acceptance. Applying this pipeline to PubMed Central yields MedSynVQA: 13,087 audited questions over 14,803 images spanning 13 imaging modalities and 28 anatomical regions. Training open-weight LMMs with reinforcement learning using verifiable rewards improves accuracy across six medical VQA benchmarks, achieving averages of 55.85 (3B) and 58.15 (7B), with up to 77.57 on VQA-RAD and 67.76 on PathVQA, outperforming strong medical LMMs. A Ablations verify that both generation and verification are necessary and that more verified data consistently helps, and a targeted contamination analysis detects no leakage from evaluation suites. By operating entirely on open literature and open-weight models, MedVLSynther offers an auditable, reproducible, and privacy-preserving path to scalable medical VQA training data.

Xubing Ye, Yukang Gan, Xiaoke Huang et al.

Vision-Language Models (VLMs) have achieved remarkable success in various multi-modal tasks, but they are often bottlenecked by the limited context window and high computational cost of processing high-resolution image inputs and videos. Vision compression can alleviate this problem by reducing the vision token count. Previous approaches compress vision tokens with external modules and force LLMs to understand the compressed ones, leading to visual information loss. However, the LLMs' understanding paradigm of vision tokens is not fully utilised in the compression learning process. We propose VoCo-LLaMA, the first approach to compress vision tokens using LLMs. By introducing Vision Compression tokens during the vision instruction tuning phase and leveraging attention distillation, our method distill how LLMs comprehend vision tokens into their processing of VoCo tokens. VoCo-LLaMA facilitates effective vision compression and improves the computational efficiency during the inference stage. Specifically, our method achieves minimal performance loss with a compression ratio of 576$\times$, resulting in up to 94.8$\%$ fewer FLOPs and 69.6$\%$ acceleration in inference time. Furthermore, through continuous training using time-series compressed token sequences of video frames, VoCo-LLaMA demonstrates the ability to understand temporal correlations, outperforming previous methods on popular video question-answering benchmarks. Our approach presents a promising way to unlock the full potential of VLMs' contextual window, enabling more scalable multi-modal applications. The project page, along with the associated code, can be accessed via https://yxxxb.github.io/VoCo-LLaMA-page/.

Hong Wang, Su Yang, Xiaoke Huang et al.

Token filtering to reduce irrelevant tokens prior to self-attention is a straightforward way to enable efficient vision Transformer. This is the first work to view token filtering from a feature selection perspective, where we weigh the importance of a token according to how much it can change the loss once masked. If the loss changes greatly after masking a token of interest, it means that such a token has a significant impact on the final decision and is thus relevant. Otherwise, the token is less important for the final decision, so it can be filtered out. After applying the token filtering module generalized from the whole training data, the token number fed to the self-attention module can be obviously reduced in the inference phase, leading to much fewer computations in all the subsequent self-attention layers. The token filter can be realized using a very simple network, where we utilize multi-layer perceptron. Except for the uniqueness of performing token filtering only once from the very beginning prior to self-attention, the other core feature making our method different from the other token filters lies in the predictability of token impact from a feature selection point of view. The experiments show that the proposed method provides an efficient way to approach a light weighted model after optimized with a backbone by means of fine tune, which is easy to be deployed in comparison with the existing methods based on training from scratch.