Jialin Wu

Anthony Brohan, Noah Brown, Justice Carbajal et al. · stanford

We study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning. Our goal is to enable a single end-to-end trained model to both learn to map robot observations to actions and enjoy the benefits of large-scale pretraining on language and vision-language data from the web. To this end, we propose to co-fine-tune state-of-the-art vision-language models on both robotic trajectory data and Internet-scale vision-language tasks, such as visual question answering. In contrast to other approaches, we propose a simple, general recipe to achieve this goal: in order to fit both natural language responses and robotic actions into the same format, we express the actions as text tokens and incorporate them directly into the training set of the model in the same way as natural language tokens. We refer to such category of models as vision-language-action models (VLA) and instantiate an example of such a model, which we call RT-2. Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data (such as placing an object onto a particular number or icon), and the ability to perform rudimentary reasoning in response to user commands (such as picking up the smallest or largest object, or the one closest to another object). We further show that incorporating chain of thought reasoning allows RT-2 to perform multi-stage semantic reasoning, for example figuring out which object to pick up for use as an improvised hammer (a rock), or which type of drink is best suited for someone who is tired (an energy drink).

Xi Chen, Xiao Wang, Lucas Beyer et al. · deepmind

This paper presents PaLI-3, a smaller, faster, and stronger vision language model (VLM) that compares favorably to similar models that are 10x larger. As part of arriving at this strong performance, we compare Vision Transformer (ViT) models pretrained using classification objectives to contrastively (SigLIP) pretrained ones. We find that, while slightly underperforming on standard image classification benchmarks, SigLIP-based PaLI shows superior performance across various multimodal benchmarks, especially on localization and visually-situated text understanding. We scale the SigLIP image encoder up to 2 billion parameters, and achieves a new state-of-the-art on multilingual cross-modal retrieval. We hope that PaLI-3, at only 5B parameters, rekindles research on fundamental pieces of complex VLMs, and could fuel a new generation of scaled-up models.

Jialin Wu, Raymond J. Mooney

Most Outside-Knowledge Visual Question Answering (OK-VQA) systems employ a two-stage framework that first retrieves external knowledge given the visual question and then predicts the answer based on the retrieved content. However, the retrieved knowledge is often inadequate. Retrievals are frequently too general and fail to cover specific knowledge needed to answer the question. Also, the naturally available supervision (whether the passage contains the correct answer) is weak and does not guarantee question relevancy. To address these issues, we propose an Entity-Focused Retrieval (EnFoRe) model that provides stronger supervision during training and recognizes question-relevant entities to help retrieve more specific knowledge. Experiments show that our EnFoRe model achieves superior retrieval performance on OK-VQA, the currently largest outside-knowledge VQA dataset. We also combine the retrieved knowledge with state-of-the-art VQA models, and achieve a new state-of-the-art performance on OK-VQA.

Peigui Qi, Kunsheng Tang, Yanpu Yu et al.

Vision-Language Models (VLMs) face significant safety vulnerabilities from malicious prompt attacks due to weakened alignment during visual integration. Existing defenses suffer from efficiency and robustness. To address these challenges, we first propose the Multimodal Aggregated Feature Extraction (MAFE) framework that enables CLIP to handle long text and fuse multimodal information into unified representations. Through empirical analysis of MAFE-extracted features, we discover distinct distributional patterns between benign and malicious prompts. Building upon this finding, we develop VLMShield, a lightweight safety detector that efficiently identifies multimodal malicious attacks as a plug-and-play solution. Extensive experiments demonstrate superior performance across multiple dimensions, including robustness, efficiency, and utility. Through our work, we hope to pave the way for more secure multimodal AI deployment. Code is available at [this https URL](https://github.com/pgqihere/VLMShield).

Nan Ding, Tomer Levinboim, Jialin Wu et al.

Recent empirical evidence indicates that transformer based in-context learning performs better when using a prefix language model (prefixLM), in which in-context samples can all attend to each other, compared to causal language models (causalLM), which use auto-regressive attention that prohibits in-context samples to attend to future samples. While this result is intuitive, it is not understood from a theoretical perspective. In this paper we take a theoretical approach and analyze the convergence behavior of prefixLM and causalLM under a certain parameter construction. Our analysis shows that both LM types converge to their stationary points at a linear rate, but that while prefixLM converges to the optimal solution of linear regression, causalLM convergence dynamics follows that of an online gradient descent algorithm, which is not guaranteed to be optimal even as the number of samples grows infinitely. We supplement our theoretical claims with empirical experiments over synthetic and real tasks and using various types of transformers. Our experiments verify that causalLM consistently underperforms prefixLM in all settings.

Sicheng Shen, Mingyang Lv, Han Shen et al.

The safety of large language models (LLMs) has increasingly emerged as a fundamental aspect of their development. Existing safety alignment for LLMs is predominantly achieved through post-training methods, which are computationally expensive and often fail to generalize well across different models. A small number of lightweight alignment approaches either rely heavily on prior-computed safety injections or depend excessively on the model's own capabilities, resulting in limited generalization and degraded efficiency and usability during generation. In this work, we propose a safety-aware decoding method that requires only low-cost training of an expert model and employs a single neuron as a gating mechanism. By effectively balancing the model's intrinsic capabilities with external guidance, our approach simultaneously preserves utility and enhances output safety. It demonstrates clear advantages in training overhead and generalization across model scales, offering a new perspective on lightweight alignment for the safe and practical deployment of large language models. Code: https://github.com/Beijing-AISI/NGSD.

Yaqing Wang, Jialin Wu, Tanmaya Dabral et al.

Large language models (LLMs) and vision language models (VLMs) demonstrate excellent performance on a wide range of tasks by scaling up parameter counts from O(10^9) to O(10^{12}) levels and further beyond. These large scales make it impossible to adapt and deploy fully specialized models given a task of interest. Parameter-efficient fine-tuning (PEFT) emerges as a promising direction to tackle the adaptation and serving challenges for such large models. We categorize PEFT techniques into two types: intrusive and non-intrusive. Intrusive PEFT techniques directly change a model's internal architecture. Though more flexible, they introduce significant complexities for training and serving. Non-intrusive PEFT techniques leave the internal architecture unchanged and only adapt model-external parameters, such as embeddings for input. In this work, we describe AdaLink as a non-intrusive PEFT technique that achieves competitive performance compared to SoTA intrusive PEFT (LoRA) and full model fine-tuning (FT) on various tasks. We evaluate using both text-only and multimodal tasks, with experiments that account for both parameter-count scaling and training regime (with and without instruction tuning).

Jialin Wu, Jiangyi Deng, Shengyuan Pang et al.

Given the societal impact of unsafe content generated by large language models (LLMs), ensuring that LLM services comply with safety standards is a crucial concern for LLM service providers. Common content moderation methods are limited by an effectiveness-and-efficiency dilemma, where simple models are fragile while sophisticated models consume excessive computational resources. In this paper, we reveal for the first time that effective and efficient content moderation can be achieved by extracting conceptual features from chat-oriented LLMs, despite their initial fine-tuning for conversation rather than content moderation. We propose a practical and unified content moderation framework for LLM services, named Legilimens, which features both effectiveness and efficiency. Our red-team model-based data augmentation enhances the robustness of Legilimens against state-of-the-art jailbreaking. Additionally, we develop a framework to theoretically analyze the cost-effectiveness of Legilimens compared to other methods. We have conducted extensive experiments on five host LLMs, seventeen datasets, and nine jailbreaking methods to verify the effectiveness, efficiency, and robustness of Legilimens against normal and adaptive adversaries. A comparison of Legilimens with both commercial and academic baselines demonstrates the superior performance of Legilimens. Furthermore, we confirm that Legilimens can be applied to few-shot scenarios and extended to multi-label classification tasks.

Jialin Wu

The possibility of super-AIs taking over the world has been intensively studied by numerous scholars. This paper focuses on the multi-AI competition scenario under the premise of super-AIs in power. Firstly, the article points out the defects of existing arguments supporting single-AI domination and presents arguments in favour of multi-AI competition. Then the article concludes that the multi-AI competition situation is a non-negligible possibility. Attention then turns to whether multi-AI competition is better for the overall good of humanity than a situation where a single AI is in power. After analysing the best, worst, and intermediate scenarios, the article concludes that multi-AI competition is better for humanity. Finally, considering the factors related to the formation of the best-case scenario of multiple AIs, the article gives some suggestions for current initiatives in AI development.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Xinnan Zhang, Jialin Wu, Junyi Xie et al.

The prediction modeling of drug-target interactions is crucial to drug discovery and design, which has seen rapid advancements owing to deep learning technologies. Recently developed methods, such as those based on graph neural networks (GNNs) and Transformers, demonstrate exceptional performance across various datasets by effectively extracting structural information. However, the benchmarking of these novel methods often varies significantly in terms of hyperparameter settings and datasets, which limits algorithmic progress. In view of these, we conducted a comprehensive survey and benchmark for drug-target interaction modeling from a structural perspective via integrating tens of explicit (i.e., GNN-based) and implicit (i.e., Transformer-based) structure learning algorithms. We conducted a macroscopical comparison between these two classes of encoding strategies as well as the different featurization techniques that inform molecules' chemical and physical properties. We then carry out the microscopical comparison between all the integrated models across the six datasets via comprehensively benchmarking their effectiveness and efficiency. To ensure fairness, we investigate model performance under individually optimized configuration. Remarkably, the summarized insights from the benchmark studies lead to the design of model combos. We demonstrate that our combos can achieve new state-of-the-art performance on various datasets associated with cost-effective memory and computation.

Qianshan Wei, Jiaqi Li, Zihan You et al.

Differential Privacy (DP) is a widely adopted technique, valued for its effectiveness in protecting the privacy of task-specific datasets, making it a critical tool for large language models. However, its effectiveness in Multimodal Large Language Models (MLLMs) remains uncertain. Applying Differential Privacy (DP) inherently introduces substantial computation overhead, a concern particularly relevant for MLLMs which process extensive textual and visual data. Furthermore, a critical challenge of DP is that the injected noise, necessary for privacy, scales with parameter dimensionality, leading to pronounced model degradation; This trade-off between privacy and utility complicates the application of Differential Privacy (DP) to complex architectures like MLLMs. To address these, we propose Dual-Priv Pruning, a framework that employs two complementary pruning mechanisms for DP fine-tuning in MLLMs: (i) visual token pruning to reduce input dimensionality by removing redundant visual information, and (ii) gradient-update pruning during the DP optimization process. This second mechanism selectively prunes parameter updates based on the magnitude of noisy gradients, aiming to mitigate noise impact and improve utility. Experiments demonstrate that our approach achieves competitive results with minimal performance degradation. In terms of computational efficiency, our approach consistently utilizes less memory than standard DP-SGD. While requiring only 1.74% more memory than zeroth-order methods which suffer from severe performance issues on A100 GPUs, our method demonstrates leading memory efficiency on H20 GPUs. To the best of our knowledge, we are the first to explore DP fine-tuning in MLLMs. Our code is coming soon.

Jialin Wu, Jiasen Lu, Ashish Sabharwal et al.

The problem of knowledge-based visual question answering involves answering questions that require external knowledge in addition to the content of the image. Such knowledge typically comes in various forms, including visual, textual, and commonsense knowledge. Using more knowledge sources increases the chance of retrieving more irrelevant or noisy facts, making it challenging to comprehend the facts and find the answer. To address this challenge, we propose Multi-modal Answer Validation using External knowledge (MAVEx), where the idea is to validate a set of promising answer candidates based on answer-specific knowledge retrieval. Instead of searching for the answer in a vast collection of often irrelevant facts as most existing approaches do, MAVEx aims to learn how to extract relevant knowledge from noisy sources, which knowledge source to trust for each answer candidate, and how to validate the candidate using that source. Our multi-modal setting is the first to leverage external visual knowledge (images searched using Google), in addition to textual knowledge in the form of Wikipedia sentences and ConceptNet concepts. Our experiments with OK-VQA, a challenging knowledge-based VQA dataset, demonstrate that MAVEx achieves new state-of-the-art results. Our code is available at https://github.com/jialinwu17/MAVEX

Mehran Kazemi, Hamidreza Alvari, Ankit Anand et al.

Large language models have shown impressive results for multi-hop mathematical reasoning when the input question is only textual. Many mathematical reasoning problems, however, contain both text and image. With the ever-increasing adoption of vision language models (VLMs), understanding their reasoning abilities for such problems is crucial. In this paper, we evaluate the reasoning capabilities of VLMs along various axes through the lens of geometry problems. We procedurally create a synthetic dataset of geometry questions with controllable difficulty levels along multiple axes, thus enabling a systematic evaluation. The empirical results obtained using our benchmark for state-of-the-art VLMs indicate that these models are not as capable in subjects like geometry (and, by generalization, other topics requiring similar reasoning) as suggested by previous benchmarks. This is made especially clear by the construction of our benchmark at various depth levels, since solving higher-depth problems requires long chains of reasoning rather than additional memorized knowledge. We release the dataset for further research in this area.

Jialin Wu, Jian Yang, Handing Wang et al.

Model merging combines expert models for multitask performance but faces challenges from parameter interference. This has sparked recent interest in controllable model merging, giving users the ability to explicitly balance performance trade-offs. Existing approaches employ a compile-then-query paradigm, performing a costly offline multi-objective optimization to enable fast, preference-aware model generation. This offline stage typically involves iterative search or dedicated training, with complexity that grows exponentially with the number of tasks. To overcome these limitations, we shift the perspective from parameter-space optimization to a direct correction of the model's final representation. Our approach models this correction as an optimal linear transformation, yielding a closed-form solution that replaces the entire offline optimization process with a single-step, architecture-agnostic computation. This solution directly incorporates user preferences, allowing a Pareto-optimal model to be generated on-the-fly with complexity that scales linearly with the number of tasks. Experimental results show our method generates a superior Pareto front with more precise preference alignment and drastically reduced computational cost.

Jialin Wu, Kecen Li, Zhicong Huang et al.

Many machine learning models are fine-tuned from large language models (LLMs) to achieve high performance in specialized domains like code generation, biomedical analysis, and mathematical problem solving. However, this fine-tuning process often introduces a critical vulnerability: the systematic degradation of safety alignment, undermining ethical guidelines and increasing the risk of harmful outputs. Addressing this challenge, we introduce EnchTable, a novel framework designed to transfer and maintain safety alignment in downstream LLMs without requiring extensive retraining. EnchTable leverages a Neural Tangent Kernel (NTK)-based safety vector distillation method to decouple safety constraints from task-specific reasoning, ensuring compatibility across diverse model architectures and sizes. Additionally, our interference-aware merging technique effectively balances safety and utility, minimizing performance compromises across various task domains. We implemented a fully functional prototype of EnchTable on three different task domains and three distinct LLM architectures, and evaluated its performance through extensive experiments on eleven diverse datasets, assessing both utility and model safety. Our evaluations include LLMs from different vendors, demonstrating EnchTable's generalization capability. Furthermore, EnchTable exhibits robust resistance to static and dynamic jailbreaking attacks, outperforming vendor-released safety models in mitigating adversarial prompts. Comparative analyses with six parameter modification methods and two inference-time alignment baselines reveal that EnchTable achieves a significantly lower unsafe rate, higher utility score, and universal applicability across different task domains. Additionally, we validate EnchTable can be seamlessly integrated into various deployment pipelines without significant overhead.

Yue Zhao, Long Zhao, Xingyi Zhou et al. · deepmind

The recent advance in vision-language models is largely attributed to the abundance of image-text data. We aim to replicate this success for video-language models, but there simply is not enough human-curated video-text data available. We thus resort to fine-tuning a video-language model from a strong image-language baseline with synthesized instructional data. The resulting video model by video-instruction-tuning (VIIT) is then used to auto-label millions of videos to generate high-quality captions. We show the adapted video-language model performs well on a wide range of video-language benchmarks. For instance, it surpasses the best prior result on open-ended NExT-QA by 2.8%. Besides, our model generates detailed descriptions for previously unseen videos, which provide better textual supervision than existing methods. Experiments show that a video-language dual-encoder model contrastively trained on these auto-generated captions is 3.8% better than the strongest baseline that also leverages vision-language models. Our best model outperforms state-of-the-art methods on MSR-VTT zero-shot text-to-video retrieval by 6%. As a side product, we generate the largest video caption dataset to date.

Jialin Wu, Wei Shi, Han Shen et al.

Despite the advanced capabilities of Large Vision-Language Models (LVLMs), they frequently suffer from object hallucination. One reason is that visual features and pretrained textual representations often become intertwined in the deeper network layers. To address this, we propose REVIS, a training-free framework designed to explicitly re-activate this suppressed visual information. Rooted in latent space geometry, REVIS extracts the pure visual information vector via orthogonal projection and employs a calibrated strategy to perform sparse intervention only at the precise depth where suppression occurs. This surgical approach effectively restores visual information with minimal computational cost. Empirical evaluations on standard benchmarks demonstrate that REVIS reduces object hallucination rates by approximately 19% compared to state-of-the-art baselines, while preserving general reasoning capabilities.

Jialin Wu, Kaikai Pan, Yanjiao Chen et al.

Transformer models have excelled in natural language tasks, prompting the vision community to explore their implementation in computer vision problems. However, these models are still influenced by adversarial examples. In this paper, we investigate the attack capabilities of six common adversarial attacks on three pretrained ViT models to reveal the vulnerability of ViT models. To understand and analyse the bias in neural network decisions when the input is adversarial, we use two visualisation techniques that are attention rollout and grad attention rollout. To prevent ViT models from adversarial attack, we propose Protego, a detection framework that leverages the transformer intrinsic capabilities to detection adversarial examples of ViT models. Nonetheless, this is challenging due to a diversity of attack strategies that may be adopted by adversaries. Inspired by the attention mechanism, we know that the token of prediction contains all the information from the input sample. Additionally, the attention region for adversarial examples differs from that of normal examples. Given these points, we can train a detector that achieves superior performance than existing detection methods to identify adversarial examples. Our experiments have demonstrated the high effectiveness of our detection method. For these six adversarial attack methods, our detector's AUC scores all exceed 0.95. Protego may advance investigations in metaverse security.

Jialin Wu, Shreya Saha, Yiqing Bo et al.

The extent to which different neural or artificial neural networks (models) rely on equivalent representations to support similar tasks remains a central question in neuroscience and machine learning. Prior work has typically compared systems using a single representational similarity metric, yet each captures only one facet of representational structure. To address this, we leverage a suite of representational similarity metrics-each capturing a distinct facet of representational correspondence, such as geometry, unit-level tuning, or linear decodability-and assess brain region or model separability using multiple complementary measures. Metrics that preserve geometric or tuning structure (e.g., RSA, Soft Matching) yield stronger region-based discrimination, whereas more flexible mappings such as Linear Predictivity show weaker separation. These findings suggest that geometry and tuning encode brain-region- or model-family-specific signatures, while linearly decodable information tends to be more globally shared across regions or models. To integrate these complementary representational facets, we adapt Similarity Network Fusion (SNF), a framework originally developed for multi-omics data integration. SNF produces substantially sharper regional and model family-level separation than any single metric and yields robust composite similarity profiles. Moreover, clustering cortical regions using SNF-derived similarity scores reveals a clearer hierarchical organization that aligns closely with established anatomical and functional hierarchies of the visual cortex-surpassing the correspondence achieved by individual metrics.

Xinfeng Li, Shengyuan Pang, Jialin Wu et al.

Text-to-image (T2I) models, though exhibiting remarkable creativity in image generation, can be exploited to produce unsafe images. Existing safety measures, e.g., content moderation or model alignment, fail in the presence of white-box adversaries who know and can adjust model parameters, e.g., by fine-tuning. This paper presents a novel defensive framework, named Patronus, which equips T2I models with holistic protection to defend against white-box adversaries. Specifically, we design an internal moderator that decodes unsafe input features into zero vectors while ensuring the decoding performance of benign input features. Furthermore, we strengthen the model alignment with a carefully designed non-fine-tunable learning mechanism, ensuring the T2I model will not be compromised by malicious fine-tuning. We conduct extensive experiments to validate the intactness of the performance on safe content generation and the effectiveness of rejecting unsafe content generation. Results also confirm the resilience of Patronus against various fine-tuning attacks by white-box adversaries.

Jialin Wu, Xiaofeng Liu

Longitudinal medical visual question answering (Diff-VQA) requires comparing paired studies from different time points and answering questions about clinically meaningful changes. In this setting, the difference signal and the consistency of visual focus across time are more informative than absolute single-image findings. We propose a saliency-guided encoder-decoder for chest X-ray Diff-VQA that turns post-hoc saliency into actionable supervision. The model first performs a lightweight near-identity affine pre-alignment to reduce nuisance motion between visits. It then executes a within-epoch two-step loop: step 1 extracts a medically relevant keyword from the answer and generates keyword-conditioned Grad-CAM on both images to obtain disease-focused saliency; step 2 applies the shared saliency mask to both time points and generates the final answer. This closes the language-vision loop so that the terms that matter also guide where the model looks, enforcing spatially consistent attention on corresponding anatomy. On Medical-Diff-VQA, the approach attains competitive performance on BLEU, ROUGE-L, CIDEr, and METEOR while providing intrinsic interpretability. Notably, the backbone and decoder are general-domain pretrained without radiology-specific pretraining, highlighting practicality and transferability. These results support saliency-conditioned generation with mild pre-alignment as a principled framework for longitudinal reasoning in medical VQA.

Jialin Wu, Shreya Saha, Yiqing Bo et al.

Large vision models differ widely in architecture and training paradigm, yet we lack principled methods to determine which aspects of their representations are shared across families and which reflect distinctive computational strategies. We leverage a suite of representational similarity metrics, each capturing a different facet-geometry, unit tuning, or linear decodability-and assess family separability using multiple complementary measures. Metrics preserving geometry or tuning (e.g., RSA, Soft Matching) yield strong family discrimination, whereas flexible mappings such as Linear Predictivity show weaker separation. These findings indicate that geometry and tuning carry family-specific signatures, while linearly decodable information is more broadly shared. To integrate these complementary facets, we adapt Similarity Network Fusion (SNF), a method inspired by multi-omics integration. SNF achieves substantially sharper family separation than any individual metric and produces robust composite signatures. Clustering of the fused similarity matrix recovers both expected and surprising patterns: supervised ResNets and ViTs form distinct clusters, yet all self-supervised models group together across architectural boundaries. Hybrid architectures (ConvNeXt, Swin) cluster with masked autoencoders, suggesting convergence between architectural modernization and reconstruction-based training. This biology-inspired framework provides a principled typology of vision models, showing that emergent computational strategies-shaped jointly by architecture and training objective-define representational structure beyond surface design categories.

Jialin Wu, Shreya Saha, Yiqing Bo et al.

Representational similarity metrics are fundamental tools in neuroscience and AI, yet we lack systematic comparisons of their discriminative power across model families. We introduce a quantitative framework to evaluate representational similarity measures based on their ability to separate model families-across architectures (CNNs, Vision Transformers, Swin Transformers, ConvNeXt) and training regimes (supervised vs. self-supervised). Using three complementary separability measures-dprime from signal detection theory, silhouette coefficients and ROC-AUC, we systematically assess the discriminative capacity of commonly used metrics including RSA, linear predictivity, Procrustes, and soft matching. We show that separability systematically increases as metrics impose more stringent alignment constraints. Among mapping-based approaches, soft-matching achieves the highest separability, followed by Procrustes alignment and linear predictivity. Non-fitting methods such as RSA also yield strong separability across families. These results provide the first systematic comparison of similarity metrics through a separability lens, clarifying their relative sensitivity and guiding metric choice for large-scale model and brain comparisons.

Xi Chen, Josip Djolonga, Piotr Padlewski et al.

We present the training recipe and results of scaling up PaLI-X, a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. PaLI-X advances the state-of-the-art on most vision-and-language benchmarks considered (25+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.

Jung-Jun Kim, Dong-Gyu Lee, Jialin Wu et al.

Visual question answering requires a deep understanding of both images and natural language. However, most methods mainly focus on visual concept; such as the relationships between various objects. The limited use of object categories combined with their relationships or simple question embedding is insufficient for representing complex scenes and explaining decisions. To address this limitation, we propose the use of text expressions generated for images, because such expressions have few structural constraints and can provide richer descriptions of images. The generated expressions can be incorporated with visual features and question embedding to obtain the question-relevant answer. A joint-embedding multi-head attention network is also proposed to model three different information modalities with co-attention. We quantitatively and qualitatively evaluated the proposed method on the VQA v2 dataset and compared it with state-of-the-art methods in terms of answer prediction. The quality of the generated expressions was also evaluated on the RefCOCO, RefCOCO+, and RefCOCOg datasets. Experimental results demonstrate the effectiveness of the proposed method and reveal that it outperformed all of the competing methods in terms of both quantitative and qualitative results.

Jialin Wu, Liyan Chen, Raymond J. Mooney

Most recent state-of-the-art Visual Question Answering (VQA) systems are opaque black boxes that are only trained to fit the answer distribution given the question and visual content. As a result, these systems frequently take shortcuts, focusing on simple visual concepts or question priors. This phenomenon becomes more problematic as the questions become complex that requires more reasoning and commonsense knowledge. To address this issue, we present a novel framework that uses explanations for competing answers to help VQA systems select the correct answer. By training on human textual explanations, our framework builds better representations for the questions and visual content, and then reweights confidences in the answer candidates using either generated or retrieved explanations from the training set. We evaluate our framework on the VQA-X dataset, which has more difficult questions with human explanations, achieving new state-of-the-art results on both VQA and its explanations.

Jialin Wu, Raymond J. Mooney

Most RNN-based image captioning models receive supervision on the output words to mimic human captions. Therefore, the hidden states can only receive noisy gradient signals via layers of back-propagation through time, leading to less accurate generated captions. Consequently, we propose a novel framework, Hidden State Guidance (HSG), that matches the hidden states in the caption decoder to those in a teacher decoder trained on an easier task of autoencoding the captions conditioned on the image. During training with the REINFORCE algorithm, the conventional rewards are sentence-based evaluation metrics equally distributed to each generated word, no matter their relevance. HSG provides a word-level reward that helps the model learn better hidden representations. Experimental results demonstrate that HSG clearly outperforms various state-of-the-art caption decoders using either raw images or detected objects as inputs.

Jialin Wu, Zeyuan Hu, Raymond J. Mooney

Visual question answering (VQA) and image captioning require a shared body of general knowledge connecting language and vision. We present a novel approach to improve VQA performance that exploits this connection by jointly generating captions that are targeted to help answer a specific visual question. The model is trained using an existing caption dataset by automatically determining question-relevant captions using an online gradient-based method. Experimental results on the VQA v2 challenge demonstrates that our approach obtains state-of-the-art VQA performance (e.g. 68.4% on the Test-standard set using a single model) by simultaneously generating question-relevant captions.

Jialin Wu, Raymond J. Mooney

Visual Question Answering (VQA) deep-learning systems tend to capture superficial statistical correlations in the training data because of strong language priors and fail to generalize to test data with a significantly different question-answer (QA) distribution. To address this issue, we introduce a self-critical training objective that ensures that visual explanations of correct answers match the most influential image regions more than other competitive answer candidates. The influential regions are either determined from human visual/textual explanations or automatically from just significant words in the question and answer. We evaluate our approach on the VQA generalization task using the VQA-CP dataset, achieving a new state-of-the-art i.e., 49.5% using textual explanations and 48.5% using automatically annotated regions.

Simiao Zuo, Jialin Wu

There has long been debates on how we could interpret neural networks and understand the decisions our models make. Specifically, why deep neural networks tend to be error-prone when dealing with samples that output low softmax scores. We present an efficient approach to measure the confidence of decision-making steps by statistically investigating each unit's contribution to that decision. Instead of focusing on how the models react on datasets, we study the datasets themselves given a pre-trained model. Our approach is capable of assigning a score to each sample within a dataset that measures the frequency of occurrence of that sample's chain of activation. We demonstrate with experiments that our method could select useful samples to improve deep neural networks in a semi-supervised leaning setting.

Jialin Wu, Raymond J. Mooney

AI systems' ability to explain their reasoning is critical to their utility and trustworthiness. Deep neural networks have enabled significant progress on many challenging problems such as visual question answering (VQA). However, most of them are opaque black boxes with limited explanatory capability. This paper presents a novel approach to developing a high-performing VQA system that can elucidate its answers with integrated textual and visual explanations that faithfully reflect important aspects of its underlying reasoning while capturing the style of comprehensible human explanations. Extensive experimental evaluation demonstrates the advantages of this approach compared to competing methods with both automatic evaluation metrics and human evaluation metrics.

Jialin Wu, Zeyuan Hu, Raymond J. Mooney

Answering visual questions need acquire daily common knowledge and model the semantic connection among different parts in images, which is too difficult for VQA systems to learn from images with the only supervision from answers. Meanwhile, image captioning systems with beam search strategy tend to generate similar captions and fail to diversely describe images. To address the aforementioned issues, we present a system to have these two tasks compensate with each other, which is capable of jointly producing image captions and answering visual questions. In particular, we utilize question and image features to generate question-related captions and use the generated captions as additional features to provide new knowledge to the VQA system. For image captioning, our system attains more informative results in term of the relative improvements on VQA tasks as well as competitive results using automated metrics. Applying our system to the VQA tasks, our results on VQA v2 dataset achieve 65.8% using generated captions and 69.1% using annotated captions in validation set and 68.4% in the test-standard set. Further, an ensemble of 10 models results in 69.7% in the test-standard split.

Jialin Wu, Dai Li, Yu Yang et al.

We propose a dynamic filtering strategy with large sampling field for ConvNets (LS-DFN), where the position-specific kernels learn from not only the identical position but also multiple sampled neighbor regions. During sampling, residual learning is introduced to ease training and an attention mechanism is applied to fuse features from different samples. Such multiple samples enlarge the kernels' receptive fields significantly without requiring more parameters. While LS-DFN inherits the advantages of DFN, namely avoiding feature map blurring by position-wise kernels while keeping translation invariance, it also efficiently alleviates the overfitting issue caused by much more parameters than normal CNNs. Our model is efficient and can be trained end-to-end via standard back-propagation. We demonstrate the merits of our LS-DFN on both sparse and dense prediction tasks involving object detection, semantic segmentation, and flow estimation. Our results show LS-DFN enjoys stronger recognition abilities in object detection and semantic segmentation tasks on VOC benchmark and sharper responses in flow estimation on FlyingChairs dataset compared to strong baselines.

Jialin Wu, Gu Wang, Wukui Yang et al.

We propose a novel deep supervised neural network for the task of action recognition in videos, which implicitly takes advantage of visual tracking and shares the robustness of both deep Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). In our method, a multi-branch model is proposed to suppress noise from background jitters. Specifically, we firstly extract multi-level deep features from deep CNNs and feed them into 3d-convolutional network. After that we feed those feature cubes into our novel joint LSTM module to predict labels and to generate attention regularization. We evaluate our model on two challenging datasets: UCF101 and HMDB51. The results show that our model achieves the state-of-art by only using convolutional features.