Apollo: An Exploration of Video Understanding in Large Multimodal Models

Despite the rapid integration of video perception capabilities into Large Multimodal Models (LMMs), the underlying mechanisms driving their video understanding remain poorly understood. Consequently, many design decisions in this domain are made without proper justification or analysis. The high computational cost of training and evaluating such models, coupled with limited open research, hinders the development of video-LMMs. To address this, we present a comprehensive study that helps uncover what effectively drives video understanding in LMMs. We begin by critically examining the primary contributors to the high computational requirements associated with video-LMM research and discover Scaling Consistency, wherein design and training decisions made on smaller models and datasets (up to a critical size) effectively transfer to larger models. Leveraging these insights, we explored many video-specific aspects of video-LMMs, including video sampling, architectures, data composition, training schedules, and more. For example, we demonstrated that fps sampling during training is vastly preferable to uniform frame sampling and which vision encoders are the best for video representation. Guided by these findings, we introduce Apollo, a state-of-the-art family of LMMs that achieve superior performance across different model sizes. Our models can perceive hour-long videos efficiently, with Apollo-3B outperforming most existing $7$B models with an impressive 55.1 on LongVideoBench. Apollo-7B is state-of-the-art compared to 7B LMMs with a 70.9 on MLVU, and 63.3 on Video-MME.

Targeted Learning for Variable Importance

Variable importance is one of the most widely used measures for interpreting machine learning with significant interest from both statistics and machine learning communities. Recently, increasing attention has been directed toward uncertainty quantification in these metrics. Current approaches largely rely on one-step procedures, which, while asymptotically efficient, can present higher sensitivity and instability in finite sample settings. To address these limitations, we propose a novel method by employing the targeted learning (TL) framework, designed to enhance robustness in inference for variable importance metrics. Our approach is particularly suited for conditional permutation variable importance. We show that it (i) retains the asymptotic efficiency of traditional methods, (ii) maintains comparable computational complexity, and (iii) delivers improved accuracy, especially in finite sample contexts. We further support these findings with numerical experiments that illustrate the practical advantages of our method and validate the theoretical results.

Zero-shot Action Localization via the Confidence of Large Vision-Language Models

Precise action localization in untrimmed video is vital for fields such as professional sports and minimally invasive surgery, where the delineation of particular motions in recordings can dramatically enhance analysis. But in many cases, large scale datasets with video-label pairs for localization are unavailable, limiting the opportunity to fine-tune video-understanding models. Recent developments in large vision-language models (LVLM) address this need with impressive zero-shot capabilities in a variety of video understanding tasks. However, the adaptation of image-based LVLMs, with their powerful visual question answering capabilities, to action localization in long-form video is still relatively unexplored. To this end, we introduce a true ZEro-shot Action Localization method (ZEAL). Specifically, we leverage the built-in action knowledge of a large language model (LLM) to inflate actions into highly-detailed descriptions of the archetypal start and end of the action. These descriptions serve as queries to LVLM for generating frame-level confidence scores which can be aggregated to produce localization outputs. The simplicity and flexibility of our method lends it amenable to more capable LVLMs as they are developed, and we demonstrate remarkable results in zero-shot action localization on a challenging benchmark, without any training.

Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps

Humor is a culturally nuanced aspect of human language that presents challenges for understanding and generation, requiring participants to possess good creativity and strong associative thinking. Similar to reasoning tasks like solving math problems, humor generation requires continuous reflection and revision to foster creative thinking, rather than relying on a sudden flash of inspiration like Creative Leap-of-Thought (CLoT) paradigm. Although CLoT can realize the ability of remote association generation, this paradigm fails to generate humor content. Therefore, in this paper, we propose a systematic way of thinking about generating humor and based on it, we built Creative Leap of Structured Thought (CLoST) frame. First, a reward model is necessary achieve the purpose of being able to correct errors, since there is currently no expert model of humor and a usable rule to determine whether a piece of content is humorous. Judgement-oriented instructions are designed to improve the capability of a model, and we also propose an open-domain instruction evolutionary method to fully unleash the potential. Then, through reinforcement learning, the model learns to hone its rationales of the thought chain and refine the strategies it uses. Thus, it learns to recognize and correct its mistakes, and finally generate the most humorous and creative answer. These findings deepen our understanding of the creative capabilities of LLMs and provide ways to enhance LLMs' creative abilities for cross-domain innovative applications.

RuleAlign: Making Large Language Models Better Physicians with Diagnostic Rule Alignment

Large Language Models (LLMs) like GPT-4, MedPaLM-2, and Med-Gemini achieve performance competitively with human experts across various medical benchmarks. However, they still face challenges in making professional diagnoses akin to physicians, particularly in efficiently gathering patient information and reasoning the final diagnosis. To this end, we introduce the RuleAlign framework, designed to align LLMs with specific diagnostic rules. We develop a medical dialogue dataset comprising rule-based communications between patients and physicians and design an alignment learning approach through preference learning. Experimental results demonstrate the effectiveness of the proposed approach. We hope that our work can serve as an inspiration for exploring the potential of LLMs as AI physicians.

MetaTool: Facilitating Large Language Models to Master Tools with Meta-task Augmentation

Utilizing complex tools with Large Language Models (LLMs) is a critical component for grounding AI agents in various real-world scenarios. The core challenge of manipulating tools lies in understanding their usage and functionality. The prevailing approach involves few-shot prompting with demonstrations or fine-tuning on expert trajectories. However, for complex tools and tasks, mere in-context demonstrations may fail to cover sufficient knowledge. Training-based methods are also constrained by the high cost of dataset construction and limited generalizability. In this paper, we introduce a new tool learning methodology (MetaTool) that is generalizable for mastering any reusable toolset. Our approach includes a self-supervised data augmentation technique that enables LLMs to gain a comprehensive understanding of various tools, thereby improving their ability to complete tasks effectively. We develop a series of meta-tasks that involve predicting masked factors of tool execution. These self-supervised tasks enable the automatic generation of high-quality QA data concerning tool comprehension. By incorporating meta-task data into the instruction tuning process, the proposed MetaTool model achieves significant superiority to open-source models and is comparable to GPT-4/GPT-3.5 on multiple tool-oriented tasks.

Video-STaR: Self-Training Enables Video Instruction Tuning with Any Supervision

The performance of Large Vision Language Models (LVLMs) is dependent on the size and quality of their training datasets. Existing video instruction tuning datasets lack diversity as they are derived by prompting large language models with video captions to generate question-answer pairs, and are therefore mostly descriptive. Meanwhile, many labeled video datasets with diverse labels and supervision exist - however, we find that their integration into LVLMs is non-trivial. Herein, we present Video Self-Training with augmented Reasoning (Video-STaR), the first video self-training approach. Video-STaR allows the utilization of any labeled video dataset for video instruction tuning. In Video-STaR, an LVLM cycles between instruction generation and finetuning, which we show (I) improves general video understanding and (II) adapts LVLMs to novel downstream tasks with existing supervision. During generation, an LVLM is prompted to propose an answer. The answers are then filtered only to those that contain the original video labels, and the LVLM is then re-trained on the generated dataset. By only training on generated answers that contain the correct video labels, Video-STaR utilizes these existing video labels as weak supervision for video instruction tuning. Our results demonstrate that Video-STaR-enhanced LVLMs exhibit improved performance in (I) general video QA, where TempCompass performance improved by 10%, and (II) on downstream tasks, where Video-STaR improved Kinetics700-QA accuracy by 20% and action quality assessment on FineDiving by 15%.

Why are Visually-Grounded Language Models Bad at Image Classification?

Image classification is one of the most fundamental capabilities of machine vision intelligence. In this work, we revisit the image classification task using visually-grounded language models (VLMs) such as GPT-4V and LLaVA. We find that existing proprietary and public VLMs, despite often using CLIP as a vision encoder and having many more parameters, significantly underperform CLIP on standard image classification benchmarks like ImageNet. To understand the reason, we explore several hypotheses concerning the inference algorithms, training objectives, and data processing in VLMs. Our analysis reveals that the primary cause is data-related: critical information for image classification is encoded in the VLM's latent space but can only be effectively decoded with enough training data. Specifically, there is a strong correlation between the frequency of class exposure during VLM training and instruction-tuning and the VLM's performance in those classes; when trained with sufficient data, VLMs can match the accuracy of state-of-the-art classification models. Based on these findings, we enhance a VLM by integrating classification-focused datasets into its training, and demonstrate that the enhanced classification performance of the VLM transfers to its general capabilities, resulting in an improvement of 11.8% on the newly collected ImageWikiQA dataset.

Just Shift It: Test-Time Prototype Shifting for Zero-Shot Generalization with Vision-Language Models

Advancements in vision-language models (VLMs) have propelled the field of computer vision, particularly in the zero-shot learning setting. Despite their promise, the effectiveness of these models often diminishes due to domain shifts in test environments. To address this, we introduce the Test-Time Prototype Shifting (TPS) framework, a pioneering approach designed to adapt VLMs to test datasets using unlabeled test inputs. Our method is based on the notion of modulating per-class prototypes in the shared embedding space. By pre-computing and caching prototypes generated with the pre-trained text encoder, TPS not only facilitates optimization-free prototype reuse for subsequent predictions but also enables seamless integration with current advancements in prompt engineering. At test-time, TPS dynamically learns shift vectors for each prototype based solely on the given test sample, effectively bridging the domain gap and enhancing classification accuracy. A notable aspect of our framework is its significantly reduced memory and computational demands when compared to conventional text-prompt tuning methods. Extensive evaluations across 15 datasets involving natural distribution shifts and cross-dataset generalization demonstrate TPS's superior performance, achieving state-of-the-art results while reducing resource requirements.

VideoAgent: Long-form Video Understanding with Large Language Model as Agent

Long-form video understanding represents a significant challenge within computer vision, demanding a model capable of reasoning over long multi-modal sequences. Motivated by the human cognitive process for long-form video understanding, we emphasize interactive reasoning and planning over the ability to process lengthy visual inputs. We introduce a novel agent-based system, VideoAgent, that employs a large language model as a central agent to iteratively identify and compile crucial information to answer a question, with vision-language foundation models serving as tools to translate and retrieve visual information. Evaluated on the challenging EgoSchema and NExT-QA benchmarks, VideoAgent achieves 54.1% and 71.3% zero-shot accuracy with only 8.4 and 8.2 frames used on average. These results demonstrate superior effectiveness and efficiency of our method over the current state-of-the-art methods, highlighting the potential of agent-based approaches in advancing long-form video understanding.