Shijie Geng

Peng Gao, Jiaming Han, Renrui Zhang et al. · berkeley, stanford

How to efficiently transform large language models (LLMs) into instruction followers is recently a popular research direction, while training LLM for multi-modal reasoning remains less explored. Although the recent LLaMA-Adapter demonstrates the potential to handle visual inputs with LLMs, it still cannot generalize well to open-ended visual instructions and lags behind GPT-4. In this paper, we present LLaMA-Adapter V2, a parameter-efficient visual instruction model. Specifically, we first augment LLaMA-Adapter by unlocking more learnable parameters (e.g., norm, bias and scale), which distribute the instruction-following ability across the entire LLaMA model besides adapters. Secondly, we propose an early fusion strategy to feed visual tokens only into the early LLM layers, contributing to better visual knowledge incorporation. Thirdly, a joint training paradigm of image-text pairs and instruction-following data is introduced by optimizing disjoint groups of learnable parameters. This strategy effectively alleviates the interference between the two tasks of image-text alignment and instruction following and achieves strong multi-modal reasoning with only a small-scale image-text and instruction dataset. During inference, we incorporate additional expert models (e.g. captioning/OCR systems) into LLaMA-Adapter to further enhance its image understanding capability without incurring training costs. Compared to the original LLaMA-Adapter, our LLaMA-Adapter V2 can perform open-ended multi-modal instructions by merely introducing 14M parameters over LLaMA. The newly designed framework also exhibits stronger language-only instruction-following capabilities and even excels in chat interactions. Our code and models are available at https://github.com/ZrrSkywalker/LLaMA-Adapter.

Ziyi Lin, Shijie Geng, Renrui Zhang et al.

Video recognition has been dominated by the end-to-end learning paradigm -- first initializing a video recognition model with weights of a pretrained image model and then conducting end-to-end training on videos. This enables the video network to benefit from the pretrained image model. However, this requires substantial computation and memory resources for finetuning on videos and the alternative of directly using pretrained image features without finetuning the image backbone leads to subpar results. Fortunately, recent advances in Contrastive Vision-Language Pre-training (CLIP) pave the way for a new route for visual recognition tasks. Pretrained on large open-vocabulary image-text pair data, these models learn powerful visual representations with rich semantics. In this paper, we present Efficient Video Learning (EVL) -- an efficient framework for directly training high-quality video recognition models with frozen CLIP features. Specifically, we employ a lightweight Transformer decoder and learn a query token to dynamically collect frame-level spatial features from the CLIP image encoder. Furthermore, we adopt a local temporal module in each decoder layer to discover temporal clues from adjacent frames and their attention maps. We show that despite being efficient to train with a frozen backbone, our models learn high quality video representations on a variety of video recognition datasets. Code is available at https://github.com/OpenGVLab/efficient-video-recognition.

Yuxiao Chen, Long Zhao, Jianbo Yuan et al. · deepmind

Despite the success of fully-supervised human skeleton sequence modeling, utilizing self-supervised pre-training for skeleton sequence representation learning has been an active field because acquiring task-specific skeleton annotations at large scales is difficult. Recent studies focus on learning video-level temporal and discriminative information using contrastive learning, but overlook the hierarchical spatial-temporal nature of human skeletons. Different from such superficial supervision at the video level, we propose a self-supervised hierarchical pre-training scheme incorporated into a hierarchical Transformer-based skeleton sequence encoder (Hi-TRS), to explicitly capture spatial, short-term, and long-term temporal dependencies at frame, clip, and video levels, respectively. To evaluate the proposed self-supervised pre-training scheme with Hi-TRS, we conduct extensive experiments covering three skeleton-based downstream tasks including action recognition, action detection, and motion prediction. Under both supervised and semi-supervised evaluation protocols, our method achieves the state-of-the-art performance. Additionally, we demonstrate that the prior knowledge learned by our model in the pre-training stage has strong transfer capability for different downstream tasks.

Shijie Geng, Shuchang Liu, Zuohui Fu et al.

For a long time, different recommendation tasks typically require designing task-specific architectures and training objectives. As a result, it is hard to transfer the learned knowledge and representations from one task to another, thus restricting the generalization ability of existing recommendation approaches, e.g., a sequential recommendation model can hardly be applied or transferred to a review generation method. To deal with such issues, considering that language can describe almost anything and language grounding is a powerful medium to represent various problems or tasks, we present a flexible and unified text-to-text paradigm called "Pretrain, Personalized Prompt, and Predict Paradigm" (P5) for recommendation, which unifies various recommendation tasks in a shared framework. In P5, all data such as user-item interactions, user descriptions, item metadata, and user reviews are converted to a common format -- natural language sequences. The rich information from natural language assists P5 to capture deeper semantics for personalization and recommendation. Specifically, P5 learns different tasks with the same language modeling objective during pretraining. Thus, it serves as the foundation model for various downstream recommendation tasks, allows easy integration with other modalities, and enables instruction-based recommendation based on prompts. P5 advances recommender systems from shallow model to deep model to big model, and will revolutionize the technical form of recommender systems towards universal recommendation engine. With adaptive personalized prompt for different users, P5 is able to make predictions in a zero-shot or few-shot manner and largely reduces the necessity for extensive fine-tuning. On several recommendation benchmarks, we conduct experiments to show the effectiveness of P5. We release the source code at https://github.com/jeykigung/P5.

Bytedance-Seed-Foundation-Code-Team, Yao Cheng, Jianfeng Chen et al. · bytedance

As the capabilities of code large language models (LLMs) continue to expand, their applications across diverse code intelligence domains are rapidly increasing. However, most existing datasets only evaluate limited application domains. To address this gap, we have developed a comprehensive code evaluation dataset FullStack Bench focusing on full-stack programming, which encompasses a wide range of application domains (e.g., basic programming, data analysis, software engineering, mathematics, and machine learning). Besides, to assess multilingual programming capabilities, in FullStack Bench, we design real-world instructions and corresponding unit test cases from 16 widely-used programming languages to reflect real-world usage scenarios rather than simple translations. Moreover, we also release an effective code sandbox execution tool (i.e., SandboxFusion) supporting various programming languages and packages to evaluate the performance of our FullStack Bench efficiently. Comprehensive experimental results on our FullStack Bench demonstrate the necessity and effectiveness of our FullStack Bench and SandboxFusion.

Zijian An, Augustus Sroka, Ran Yang et al.

Orientation-aware manipulation is essential in post-harvest agricultural processing, where produce must be grasped and placed in consistent configurations. This paper presents ROG-Grasp, a geometry-based robotic grasping and placement framework that estimates the produce orientation from root surface geometry using RGB-D perception. A YOLO-based root detector and point cloud plane fitting are used to infer the root normal, enabling stable grasp pose generation and orientation-constrained Cartesian motion planning. Experiments on tomatoes and onions demonstrate high success rates and stable execution time in both isolated and cluttered scenarios. Compared with vision-language-action (VLA) policies, the proposed method achieves more reliable and accurate grasp completion with faster execution. These results highlight the effectiveness of geometry-driven perception for practical orientation-controlled manipulation tasks. A video of our paper is available online https://youtu.be/Ir2UtGODdMo.

Yuxiao Chen, Jianbo Yuan, Yu Tian et al.

Contrastive learning-based vision-language pre-training approaches, such as CLIP, have demonstrated great success in many vision-language tasks. These methods achieve cross-modal alignment by encoding a matched image-text pair with similar feature embeddings, which are generated by aggregating information from visual patches and language tokens. However, direct aligning cross-modal information using such representations is challenging, as visual patches and text tokens differ in semantic levels and granularities. To alleviate this issue, we propose a Finite Discrete Tokens (FDT) based multimodal representation. FDT is a set of learnable tokens representing certain visual-semantic concepts. Both images and texts are embedded using shared FDT by first grounding multimodal inputs to FDT space and then aggregating the activated FDT representations. The matched visual and semantic concepts are enforced to be represented by the same set of discrete tokens by a sparse activation constraint. As a result, the granularity gap between the two modalities is reduced. Through both quantitative and qualitative analyses, we demonstrate that using FDT representations in CLIP-style models improves cross-modal alignment and performance in visual recognition and vision-language downstream tasks. Furthermore, we show that our method can learn more comprehensive representations, and the learned FDT capture meaningful cross-modal correspondence, ranging from objects to actions and attributes.

Haonan Chang, Kowndinya Boyalakuntla, Shiyang Lu et al.

We present an Open-Vocabulary 3D Scene Graph (OVSG), a formal framework for grounding a variety of entities, such as object instances, agents, and regions, with free-form text-based queries. Unlike conventional semantic-based object localization approaches, our system facilitates context-aware entity localization, allowing for queries such as ``pick up a cup on a kitchen table" or ``navigate to a sofa on which someone is sitting". In contrast to existing research on 3D scene graphs, OVSG supports free-form text input and open-vocabulary querying. Through a series of comparative experiments using the ScanNet dataset and a self-collected dataset, we demonstrate that our proposed approach significantly surpasses the performance of previous semantic-based localization techniques. Moreover, we highlight the practical application of OVSG in real-world robot navigation and manipulation experiments.

Shijie Geng, Jianbo Yuan, Yu Tian et al.

The success of large-scale contrastive vision-language pretraining (CLIP) has benefited both visual recognition and multimodal content understanding. The concise design brings CLIP the advantage in inference efficiency against other vision-language models with heavier cross-attention fusion layers, making it a popular choice for a wide spectrum of downstream tasks. However, CLIP does not explicitly capture the hierarchical nature of high-level and fine-grained semantics conveyed in images and texts, which is arguably critical to vision-language understanding and reasoning. To this end, we equip both the visual and language branches in CLIP with hierarchy-aware attentions, namely Hierarchy-aware CLIP (HiCLIP), to progressively discover semantic hierarchies layer-by-layer from both images and texts in an unsupervised manner. As a result, such hierarchical aggregation significantly improves the cross-modal alignment. To demonstrate the advantages of HiCLIP, we conduct qualitative analysis on its unsupervised hierarchy induction during inference, as well as extensive quantitative experiments on both visual recognition and vision-language downstream tasks.

Haonan Chang, Dhruv Metha Ramesh, Shijie Geng et al.

We present Mono-STAR, the first real-time 3D reconstruction system that simultaneously supports semantic fusion, fast motion tracking, non-rigid object deformation, and topological change under a unified framework. The proposed system solves a new optimization problem incorporating optical-flow-based 2D constraints to deal with fast motion and a novel semantic-aware deformation graph (SAD-graph) for handling topology change. We test the proposed system under various challenging scenes and demonstrate that it significantly outperforms existing state-of-the-art methods.

Dongyang Liu, Renrui Zhang, Longtian Qiu et al. · stanford, tsinghua

We propose SPHINX-X, an extensive Multimodality Large Language Model (MLLM) series developed upon SPHINX. To improve the architecture and training efficiency, we modify the SPHINX framework by removing redundant visual encoders, bypassing fully-padded sub-images with skip tokens, and simplifying multi-stage training into a one-stage all-in-one paradigm. To fully unleash the potential of MLLMs, we assemble a comprehensive multi-domain and multimodal dataset covering publicly available resources in language, vision, and vision-language tasks. We further enrich this collection with our curated OCR intensive and Set-of-Mark datasets, extending the diversity and generality. By training over different base LLMs including TinyLlama1.1B, InternLM2-7B, LLaMA2-13B, and Mixtral8x7B, we obtain a spectrum of MLLMs that vary in parameter size and multilingual capabilities. Comprehensive benchmarking reveals a strong correlation between the multi-modal performance with the data and parameter scales. Code and models are released at https://github.com/Alpha-VLLM/LLaMA2-Accessory

Haixin Wang, Hejie Cui, Chenwei Zhang et al.

Recent progress in multi-turn reinforcement learning (RL) has significantly improved reasoning LLMs' performances on complex interactive tasks. Despite advances in stabilization techniques such as fine-grained credit assignment and trajectory filtering, instability remains pervasive and often leads to training collapse. We argue that this instability stems from inefficient exploration in multi-turn settings, where policies continue to generate low-information actions that neither reduce uncertainty nor advance task progress. To address this issue, we propose Token- and Turn-level Policy Optimization (T$^2$PO), an uncertainty-aware framework that explicitly controls exploration at fine-grained levels. At the token level, T$^2$PO monitors uncertainty dynamics and triggers a thinking intervention once the marginal uncertainty change falls below a threshold. At the turn level, T$^2$PO identifies interactions with negligible exploration progress and dynamically resamples such turns to avoid wasted rollouts. We evaluate T$^2$PO in diverse environments, including WebShop, ALFWorld, and Search QA, demonstrating substantial gains in training stability and performance improvements with better exploration efficiency. Code is available at: https://github.com/WillDreamer/T2PO.

Linyi Li, Shijie Geng, Zhenwen Li et al.

Large Language Models for code (code LLMs) have witnessed tremendous progress in recent years. With the rapid development of code LLMs, many popular evaluation benchmarks, such as HumanEval, DS-1000, and MBPP, have emerged to measure the performance of code LLMs with a particular focus on code generation tasks. However, they are insufficient to cover the full range of expected capabilities of code LLMs, which span beyond code generation to answering diverse coding-related questions. To fill this gap, we propose InfiBench, the first large-scale freeform question-answering (QA) benchmark for code to our knowledge, comprising 234 carefully selected high-quality Stack Overflow questions that span across 15 programming languages. InfiBench uses four types of model-free automatic metrics to evaluate response correctness where domain experts carefully concretize the criterion for each question. We conduct a systematic evaluation for over 100 latest code LLMs on InfiBench, leading to a series of novel and insightful findings. Our detailed analyses showcase potential directions for further advancement of code LLMs. InfiBench is fully open source at https://infi-coder.github.io/infibench and continuously expanding to foster more scientific and systematic practices for code LLM evaluation.

Shijie Geng, Juntao Tan, Shuchang Liu et al.

Computer Vision (CV), Natural Language Processing (NLP), and Recommender Systems (RecSys) are three prominent AI applications that have traditionally developed independently, resulting in disparate modeling and engineering methodologies. This has impeded the ability for these fields to directly benefit from each other's advancements. With the recent development of foundation models, large language models have emerged as a potential general-purpose interface for unifying different modalities and problem formulations. In light of this, we propose the development of a multimodal foundation model (MFM) considering visual, textual, and personalization modalities under the P5 recommendation paradigm, thus named VIP5 (Visual P5), to unify various modalities and recommendation tasks. This will enable the processing of multiple modalities in a shared architecture for improved recommendations. To achieve this, we introduce multimodal personalized prompts to accommodate multiple modalities under a shared format. Additionally, we propose a parameter-efficient training method for foundation models, which involves freezing the P5 backbone and fine-tuning lightweight adapters, resulting in improved recommendation performance and increased efficiency in terms of training time and memory usage. Code and data of VIP5 are available at https://github.com/jeykigung/VIP5.

Juntao Tan, Shijie Geng, Zuohui Fu et al.

Structural data well exists in Web applications, such as social networks in social media, citation networks in academic websites, and threads data in online forums. Due to the complex topology, it is difficult to process and make use of the rich information within such data. Graph Neural Networks (GNNs) have shown great advantages on learning representations for structural data. However, the non-transparency of the deep learning models makes it non-trivial to explain and interpret the predictions made by GNNs. Meanwhile, it is also a big challenge to evaluate the GNN explanations, since in many cases, the ground-truth explanations are unavailable. In this paper, we take insights of Counterfactual and Factual (CF^2) reasoning from causal inference theory, to solve both the learning and evaluation problems in explainable GNNs. For generating explanations, we propose a model-agnostic framework by formulating an optimization problem based on both of the two casual perspectives. This distinguishes CF^2 from previous explainable GNNs that only consider one of them. Another contribution of the work is the evaluation of GNN explanations. For quantitatively evaluating the generated explanations without the requirement of ground-truth, we design metrics based on Counterfactual and Factual reasoning to evaluate the necessity and sufficiency of the explanations. Experiments show that no matter ground-truth explanations are available or not, CF^2 generates better explanations than previous state-of-the-art methods on real-world datasets. Moreover, the statistic analysis justifies the correlation between the performance on ground-truth evaluation and our proposed metrics. Source code is available at https://github.com/chrisjtan/gnn_cff.

Honglu Zhou, Asim Kadav, Aviv Shamsian et al.

Group Activity Recognition detects the activity collectively performed by a group of actors, which requires compositional reasoning of actors and objects. We approach the task by modeling the video as tokens that represent the multi-scale semantic concepts in the video. We propose COMPOSER, a Multiscale Transformer based architecture that performs attention-based reasoning over tokens at each scale and learns group activity compositionally. In addition, prior works suffer from scene biases with privacy and ethical concerns. We only use the keypoint modality which reduces scene biases and prevents acquiring detailed visual data that may contain private or biased information of users. We improve the multiscale representations in COMPOSER by clustering the intermediate scale representations, while maintaining consistent cluster assignments between scales. Finally, we use techniques such as auxiliary prediction and data augmentations tailored to the keypoint signals to aid model training. We demonstrate the model's strength and interpretability on two widely-used datasets (Volleyball and Collective Activity). COMPOSER achieves up to +5.4% improvement with just the keypoint modality. Code is available at https://github.com/hongluzhou/composer

Teli Ma, Shijie Geng, Mengmeng Wang et al.

The visual world naturally exhibits a long-tailed distribution of open classes, which poses great challenges to modern visual systems. Existing approaches either perform class re-balancing strategies or directly improve network modules to address the problem. However, they still train models with a finite set of predefined labels, limiting their supervision information and restricting their transferability to novel instances. Recent advances in large-scale contrastive visual-language pretraining shed light on a new pathway for visual recognition. With open-vocabulary supervisions, pretrained contrastive vision-language models learn powerful multimodal representations that are promising to handle data deficiency and unseen concepts. By calculating the semantic similarity between visual and text inputs, visual recognition is converted to a vision-language matching problem. Inspired by this, we propose BALLAD to leverage contrastive vision-language models for long-tailed recognition. We first continue pretraining the vision-language backbone through contrastive learning on a specific long-tailed target dataset. Afterward, we freeze the backbone and further employ an additional adapter layer to enhance the representations of tail classes on balanced training samples built with re-sampling strategies. Extensive experiments have been conducted on three popular long-tailed recognition benchmarks. As a result, our simple and effective approach sets the new state-of-the-art performances and outperforms competitive baselines with a large margin. Code is released at https://github.com/gaopengcuhk/BALLAD.

Peng Gao, Shijie Geng, Renrui Zhang et al.

Large-scale contrastive vision-language pre-training has shown significant progress in visual representation learning. Unlike traditional visual systems trained by a fixed set of discrete labels, a new paradigm was introduced in \cite{radford2021learning} to directly learn to align images with raw texts in an open-vocabulary setting. On downstream tasks, a carefully chosen text prompt is employed to make zero-shot predictions.~To avoid non-trivial prompt engineering, context optimization \cite{zhou2021coop} has been proposed to learn continuous vectors as task-specific prompts with few-shot training examples.~In this paper, we show that there is an alternative path to achieve better vision-language models other than prompt tuning.~While prompt tuning is for the textual inputs, we propose CLIP-Adapter to conduct fine-tuning with feature adapters on either visual or language branch. Specifically, CLIP-Adapter adopts an additional bottleneck layer to learn new features and performs residual-style feature blending with the original pre-trained features.~As a consequence, CLIP-Adapter is able to outperform context optimization while maintains a simple design. Experiments and extensive ablation studies on various visual classification tasks demonstrate the effectiveness of our approach. Code is released at t https://github.com/gaopengcuhk/CLIP-Adapter.

Shijie Geng, Peng Gao, Zuohui Fu et al.

BERT has achieved superior performances on Natural Language Understanding (NLU) tasks. However, BERT possesses a large number of parameters and demands certain resources to deploy. For acceleration, Dynamic Early Exiting for BERT (DeeBERT) has been proposed recently, which incorporates multiple exits and adopts a dynamic early-exit mechanism to ensure efficient inference. While obtaining an efficiency-performance tradeoff, the performances of early exits in multi-exit BERT are significantly worse than late exits. In this paper, we leverage gradient regularized self-distillation for RObust training of Multi-Exit BERT (RomeBERT), which can effectively solve the performance imbalance problem between early and late exits. Moreover, the proposed RomeBERT adopts a one-stage joint training strategy for multi-exits and the BERT backbone while DeeBERT needs two stages that require more training time. Extensive experiments on GLUE datasets are performed to demonstrate the superiority of our approach. Our code is available at https://github.com/romebert/RomeBERT.

Lei Shi, Kai Shuang, Shijie Geng et al.

Several multi-modality representation learning approaches such as LXMERT and ViLBERT have been proposed recently. Such approaches can achieve superior performance due to the high-level semantic information captured during large-scale multimodal pretraining. However, as ViLBERT and LXMERT adopt visual region regression and classification loss, they often suffer from domain gap and noisy label problems, based on the visual features having been pretrained on the Visual Genome dataset. To overcome these issues, we propose unbiased Contrastive Visual-Linguistic Pretraining (CVLP), which constructs a visual self-supervised loss built upon contrastive learning. We evaluate CVLP on several down-stream tasks, including VQA, GQA and NLVR2 to validate the superiority of contrastive learning on multi-modality representation learning. Our code is available at: https://github.com/ArcherYunDong/CVLP-.

Zhiqiang Tang, Xi Peng, Shijie Geng et al.

In this paper, we propose quantized densely connected U-Nets for efficient visual landmark localization. The idea is that features of the same semantic meanings are globally reused across the stacked U-Nets. This dense connectivity largely improves the information flow, yielding improved localization accuracy. However, a vanilla dense design would suffer from critical efficiency issue in both training and testing. To solve this problem, we first propose order-K dense connectivity to trim off long-distance shortcuts; then, we use a memory-efficient implementation to significantly boost the training efficiency and investigate an iterative refinement that may slice the model size in half. Finally, to reduce the memory consumption and high precision operations both in training and testing, we further quantize weights, inputs, and gradients of our localization network to low bit-width numbers. We validate our approach in two tasks: human pose estimation and face alignment. The results show that our approach achieves state-of-the-art localization accuracy, but using ~70% fewer parameters, ~98% less model size and saving ~75% training memory compared with other benchmark localizers. The code is available at https://github.com/zhiqiangdon/CU-Net.

Zijian An, Hadi Khezam, Bill Cai et al.

We present VILAS, a fully low-cost, modular robotic manipulation platform designed to support end-to-end vision-language-action (VLA) policy learning and deployment on accessible hardware. The system integrates a Fairino FR5 collaborative arm, a Jodell RG52-50 electric gripper, and a dual-camera perception module, unified through a ZMQ-based communication architecture that seamlessly coordinates teleoperation, data collection, and policy deployment within a single framework. To enable safe manipulation of fragile objects without relying on explicit force sensing, we design a kirigami-based soft compliant gripper extension that induces predictable deformation under compressive loading, providing gentle and repeatable contact with delicate targets. We deploy and evaluate three state-of-the-art VLA models on the VILAS platform: pi_0, pi_0.5, and GR00T N1.6. All models are fine-tuned from publicly released pretrained checkpoints using an identical demonstration dataset collected via our teleoperation pipeline. Experiments on a grape grasping task validate the effectiveness of the proposed system, confirming that capable manipulation policies can be successfully trained and deployed on low-cost modular hardware. Our results further provide practical insights into the deployment characteristics of current VLA models in real-world settings.

Genglin Liu, Shijie Geng, Sha Li et al.

Multimodal LLM-powered agents have recently demonstrated impressive capabilities in web navigation, enabling agents to complete complex browsing tasks across diverse domains. However, current agents struggle with repetitive errors and lack the ability to learn from past experiences across sessions, limiting their long-term robustness and sample efficiency. We introduce WebCoach, a model-agnostic self-evolving framework that equips web browsing agents with persistent cross-session memory, enabling improved long-term planning, reflection, and continual learning without retraining. WebCoach consists of three key components: (1) a WebCondenser, which standardizes raw navigation logs into concise summaries; (2) an External Memory Store, which organizes complete trajectories as episodic experiences; and (3) a Coach, which retrieves relevant experiences based on similarity and recency, and decides whether to inject task-specific advice into the agent via runtime hooks. This design empowers web agents to access long-term memory beyond their native context window, improving robustness in complex browsing tasks. Moreover, WebCoach achieves self-evolution by continuously curating episodic memory from new navigation trajectories, enabling agents to improve over time without retraining. Evaluations on the WebVoyager benchmark demonstrate that WebCoach consistently improves the performance of browser-use agents across three different LLM backbones. With a 38B model, it increases task success rates from 47% to 61% while reducing or maintaining the average number of steps. Notably, smaller base models with WebCoach achieve performance comparable to the same web agent using GPT-4o.

Peng Gao, Le Zhuo, Dongyang Liu et al.

Sora unveils the potential of scaling Diffusion Transformer for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this technical report, we introduce the Lumina-T2X family - a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a unified framework designed to transform noise into images, videos, multi-view 3D objects, and audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as [nextline] and [nextframe] tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. This unified approach enables training within a single framework for different modalities and allows for flexible generation of multimodal data at any resolution, aspect ratio, and length during inference. Advanced techniques like RoPE, RMSNorm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. We expect that the open-sourcing of Lumina-T2X will further foster creativity, transparency, and diversity in the generative AI community.

Ankit P. Shah, Shijie Geng, Peng Gao et al.

In previous work, we have proposed the Audio-Visual Scene-Aware Dialog (AVSD) task, collected an AVSD dataset, developed AVSD technologies, and hosted an AVSD challenge track at both the 7th and 8th Dialog System Technology Challenges (DSTC7, DSTC8). In these challenges, the best-performing systems relied heavily on human-generated descriptions of the video content, which were available in the datasets but would be unavailable in real-world applications. To promote further advancements for real-world applications, we proposed a third AVSD challenge, at DSTC10, with two modifications: 1) the human-created description is unavailable at inference time, and 2) systems must demonstrate temporal reasoning by finding evidence from the video to support each answer. This paper introduces the new task that includes temporal reasoning and our new extension of the AVSD dataset for DSTC10, for which we collected human-generated temporal reasoning data. We also introduce a baseline system built using an AV-transformer, which we released along with the new dataset. Finally, this paper introduces a new system that extends our baseline system with attentional multimodal fusion, joint student-teacher learning (JSTL), and model combination techniques, achieving state-of-the-art performances on the AVSD datasets for DSTC7, DSTC8, and DSTC10. We also propose two temporal reasoning methods for AVSD: one attention-based, and one based on a time-domain region proposal network.

Lei Shi, Kai Shuang, Shijie Geng et al.

Inspired by the success of BERT, several multimodal representation learning approaches have been proposed that jointly represent image and text. These approaches achieve superior performance by capturing high-level semantic information from large-scale multimodal pretraining. In particular, LXMERT and UNITER adopt visual region feature regression and label classification as pretext tasks. However, they tend to suffer from the problems of noisy labels and sparse semantic annotations, based on the visual features having been pretrained on a crowdsourced dataset with limited and inconsistent semantic labeling. To overcome these issues, we propose unbiased Dense Contrastive Visual-Linguistic Pretraining (DCVLP), which replaces the region regression and classification with cross-modality region contrastive learning that requires no annotations. Two data augmentation strategies (Mask Perturbation and Intra-/Inter-Adversarial Perturbation) are developed to improve the quality of negative samples used in contrastive learning. Overall, DCVLP allows cross-modality dense region contrastive learning in a self-supervised setting independent of any object annotations. We compare our method against prior visual-linguistic pretraining frameworks to validate the superiority of dense contrastive learning on multimodal representation learning.

Yingqiang Ge, Shuchang Liu, Zelong Li et al.

While recent years have witnessed the emergence of various explainable methods in machine learning, to what degree the explanations really represent the reasoning process behind the model prediction -- namely, the faithfulness of explanation -- is still an open problem. One commonly used way to measure faithfulness is \textit{erasure-based} criteria. Though conceptually simple, erasure-based criterion could inevitably introduce biases and artifacts. We propose a new methodology to evaluate the faithfulness of explanations from the \textit{counterfactual reasoning} perspective: the model should produce substantially different outputs for the original input and its corresponding counterfactual edited on a faithful feature. Specially, we introduce two algorithms to find the proper counterfactuals in both discrete and continuous scenarios and then use the acquired counterfactuals to measure faithfulness. Empirical results on several datasets show that compared with existing metrics, our proposed counterfactual evaluation method can achieve top correlation with the ground truth under diffe

Peng Gao, Shijie Geng, Yu Qiao et al.

Transformer has been widely adopted in Neural Machine Translation (NMT) because of its large capacity and parallel training of sequence generation. However, the deployment of Transformer is challenging because different scenarios require models of different complexities and scales. Naively training multiple Transformers is redundant in terms of both computation and memory. In this paper, we propose a novel Scalable Transformers, which naturally contains sub-Transformers of different scales and have shared parameters. Each sub-Transformer can be easily obtained by cropping the parameters of the largest Transformer. A three-stage training scheme is proposed to tackle the difficulty of training the Scalable Transformers, which introduces additional supervisions from word-level and sequence-level self-distillation. Extensive experiments were conducted on WMT EN-De and En-Fr to validate our proposed Scalable Transformers.

Yikun Xian, Zuohui Fu, Handong Zhao et al.

Recent research explores incorporating knowledge graphs (KG) into e-commerce recommender systems, not only to achieve better recommendation performance, but more importantly to generate explanations of why particular decisions are made. This can be achieved by explicit KG reasoning, where a model starts from a user node, sequentially determines the next step, and walks towards an item node of potential interest to the user. However, this is challenging due to the huge search space, unknown destination, and sparse signals over the KG, so informative and effective guidance is needed to achieve a satisfactory recommendation quality. To this end, we propose a CoArse-to-FinE neural symbolic reasoning approach (CAFE). It first generates user profiles as coarse sketches of user behaviors, which subsequently guide a path-finding process to derive reasoning paths for recommendations as fine-grained predictions. User profiles can capture prominent user behaviors from the history, and provide valuable signals about which kinds of path patterns are more likely to lead to potential items of interest for the user. To better exploit the user profiles, an improved path-finding algorithm called Profile-guided Path Reasoning (PPR) is also developed, which leverages an inventory of neural symbolic reasoning modules to effectively and efficiently find a batch of paths over a large-scale KG. We extensively experiment on four real-world benchmarks and observe substantial gains in the recommendation performance compared with state-of-the-art methods.

Peng Gao, Chiori Hori, Shijie Geng et al.

In contrast with previous approaches where information flows only towards deeper layers of a stack, we consider a multi-pass transformer (MPT) architecture in which earlier layers are allowed to process information in light of the output of later layers. To maintain a directed acyclic graph structure, the encoder stack of a transformer is repeated along a new multi-pass dimension, keeping the parameters tied, and information is allowed to proceed unidirectionally both towards deeper layers within an encoder stack and towards any layer of subsequent stacks. We consider both soft (i.e., continuous) and hard (i.e., discrete) connections between parallel encoder stacks, relying on a neural architecture search to find the best connection pattern in the hard case. We perform an extensive ablation study of the proposed MPT architecture and compare it with other state-of-the-art transformer architectures. Surprisingly, Base Transformer equipped with MPT can surpass the performance of Large Transformer on the challenging machine translation En-De and En-Fr datasets. In the hard connection case, the optimal connection pattern found for En-De also leads to improved performance for En-Fr.

Shijie Geng, Peng Gao, Moitreya Chatterjee et al.

Given an input video, its associated audio, and a brief caption, the audio-visual scene aware dialog (AVSD) task requires an agent to indulge in a question-answer dialog with a human about the audio-visual content. This task thus poses a challenging multi-modal representation learning and reasoning scenario, advancements into which could influence several human-machine interaction applications. To solve this task, we introduce a semantics-controlled multi-modal shuffled Transformer reasoning framework, consisting of a sequence of Transformer modules, each taking a modality as input and producing representations conditioned on the input question. Our proposed Transformer variant uses a shuffling scheme on their multi-head outputs, demonstrating better regularization. To encode fine-grained visual information, we present a novel dynamic scene graph representation learning pipeline that consists of an intra-frame reasoning layer producing spatio-semantic graph representations for every frame, and an inter-frame aggregation module capturing temporal cues. Our entire pipeline is trained end-to-end. We present experiments on the benchmark AVSD dataset, both on answer generation and selection tasks. Our results demonstrate state-of-the-art performances on all evaluation metrics.

Zuohui Fu, Yikun Xian, Ruoyuan Gao et al.

There has been growing attention on fairness considerations recently, especially in the context of intelligent decision making systems. Explainable recommendation systems, in particular, may suffer from both explanation bias and performance disparity. In this paper, we analyze different groups of users according to their level of activity, and find that bias exists in recommendation performance between different groups. We show that inactive users may be more susceptible to receiving unsatisfactory recommendations, due to insufficient training data for the inactive users, and that their recommendations may be biased by the training records of more active users, due to the nature of collaborative filtering, which leads to an unfair treatment by the system. We propose a fairness constrained approach via heuristic re-ranking to mitigate this unfairness problem in the context of explainable recommendation over knowledge graphs. We experiment on several real-world datasets with state-of-the-art knowledge graph-based explainable recommendation algorithms. The promising results show that our algorithm is not only able to provide high-quality explainable recommendations, but also reduces the recommendation unfairness in several respects.

Shijie Geng, Ji Zhang, Zuohui Fu et al.

Different from short videos and GIFs, video stories contain clear plots and lists of principal characters. Without identifying the connection between appearing people and character names, a model is not able to obtain a genuine understanding of the plots. Video Story Question Answering (VSQA) offers an effective way to benchmark higher-level comprehension abilities of a model. However, current VSQA methods merely extract generic visual features from a scene. With such an approach, they remain prone to learning just superficial correlations. In order to attain a genuine understanding of who did what to whom, we propose a novel model that continuously refines character-aware relations. This model specifically considers the characters in a video story, as well as the relations connecting different characters and objects. Based on these signals, our framework enables weakly-supervised face naming through multi-instance co-occurrence matching and supports high-level reasoning utilizing Transformer structures. We train and test our model on the six diverse TV shows in the TVQA dataset, which is by far the largest and only publicly available dataset for VSQA. We validate our proposed approach over TVQA dataset through extensive ablation study.

Zuohui Fu, Yikun Xian, Shijie Geng et al.

A number of cross-lingual transfer learning approaches based on neural networks have been proposed for the case when large amounts of parallel text are at our disposal. However, in many real-world settings, the size of parallel annotated training data is restricted. Additionally, prior cross-lingual mapping research has mainly focused on the word level. This raises the question of whether such techniques can also be applied to effortlessly obtain cross-lingually aligned sentence representations. To this end, we propose an Adversarial Bi-directional Sentence Embedding Mapping (ABSent) framework, which learns mappings of cross-lingual sentence representations from limited quantities of parallel data.

Lei Shi, Shijie Geng, Kai Shuang et al.

Multi-modality fusion technologies have greatly improved the performance of neural network-based Video Description/Caption, Visual Question Answering (VQA) and Audio Visual Scene-aware Dialog (AVSD) over the recent years. Most previous approaches only explore the last layers of multiple layer feature fusion while omitting the importance of intermediate layers. To solve the issue for the intermediate layers, we propose an efficient Quaternion Block Network (QBN) to learn interaction not only for the last layer but also for all intermediate layers simultaneously. In our proposed QBN, we use the holistic text features to guide the update of visual features. In the meantime, Hamilton quaternion products can efficiently perform information flow from higher layers to lower layers for both visual and text modalities. The evaluation results show our QBN improved the performance on VQA 2.0, even though using surpass large scale BERT or visual BERT pre-trained models. Extensive ablation study has been carried out to testify the influence of each proposed module in this study.

Shijie Geng, Ji Zhang, Hang Zhang et al.

We present a simple method that achieves unexpectedly superior performance for Complex Reasoning involved Visual Question Answering. Our solution collects statistical features from high-frequency words of all the questions asked about an image and use them as accurate knowledge for answering further questions of the same image. We are fully aware that this setting is not ubiquitously applicable, and in a more common setting one should assume the questions are asked separately and they cannot be gathered to obtain a knowledge base. Nonetheless, we use this method as an evidence to demonstrate our observation that the bottleneck effect is more severe on the feature extraction part than it is on the knowledge reasoning part. We show significant gaps when using the same reasoning model with 1) ground-truth features; 2) statistical features; 3) detected features from completely learned detectors, and analyze what these gaps mean to researches on visual reasoning topics. Our model with the statistical features achieves the 2nd place in the GQA Challenge 2019.

Zhiqiang Tang, Xi Peng, Shijie Geng et al.

We design a new connectivity pattern for the U-Net architecture. Given several stacked U-Nets, we couple each U-Net pair through the connections of their semantic blocks, resulting in the coupled U-Nets (CU-Net). The coupling connections could make the information flow more efficiently across U-Nets. The feature reuse across U-Nets makes each U-Net very parameter efficient. We evaluate the coupled U-Nets on two benchmark datasets of human pose estimation. Both the accuracy and model parameter number are compared. The CU-Net obtains comparable accuracy as state-of-the-art methods. However, it only has at least 60% fewer parameters than other approaches.