ICLR 2025 Papers

Multi-modal Large Language Models have demonstrated remarkable capabilities in executing instructions for a variety of single-image tasks. Despite this progress, significant challenges remain in modeling long image sequences. In this work, we introduce the versatile multi-modal large language model, mPLUG-Owl3, which enhances the capability for long image-sequence understanding in scenarios that incorporate retrieved image-text knowledge, multimodal in-context examples, and lengthy videos. Specifically, we propose novel hyper attention blocks to efficiently integrate vision and language into a common language-guided semantic space, thereby facilitating the processing of extended multi-image scenarios. We conduct evaluations on 21 benchmarks that cover single/multi-image, and short/long video understanding. mPLUG-Owl3 achieves competitive performance with the state-of-the-art methods while reducing inference time and memory usage by 87.8\% and 48.5\% in average. Moreover, we propose a Distractor Resistance evaluation to assess the ability of models to maintain focus amidst distractions. mPLUG-Owl3 also demonstrates outstanding performance in distractor resistance on ultra-long visual sequence inputs. We hope that mPLUG-Owl3 can contribute to the development of more efficient and powerful multimodal large language models.

Recent progress in large-scale text-to-video (T2V) and image-to-video (I2V) diffusion models has greatly enhanced video generation, especially in terms of keyframe interpolation. However, current image-to-video diffusion models, while powerful in generating videos from a single conditioning frame, need adaptation for two-frame (start \& end) conditioned generation, which is essential for effective bounded interpolation. Unfortunately, existing approaches that fuse temporally forward and backward paths in parallel often suffer from off-manifold issues, leading to artifacts or requiring multiple iterative re-noising steps. In this work, we introduce a novel, bidirectional sampling strategy to address these off-manifold issues without requiring extensive re-noising or fine-tuning. Our method employs sequential sampling along both forward and backward paths, conditioned on the start and end frames, respectively, ensuring more coherent and on-manifold generation of intermediate frames. Additionally, we incorporate advanced guidance techniques, CFG++ and DDS, to further enhance the interpolation process. By integrating these, our method achieves state-of-the-art performance, efficiently generating high-quality, smooth videos between keyframes. On a single 3090 GPU, our method can interpolate 25 frames at 1024$\times$576 resolution in just 195 seconds, establishing it as a leading solution for keyframe interpolation. Project page: https://vibidsampler.github.io/

Multimodal Large Language Models (MLLMs) have demonstrated impressive performance in short video understanding. However, understanding long-form videos still remains challenging for MLLMs. This paper proposes TimeSuite, a collection of new designs to adapt the existing short-form video MLLMs for long video understanding, including a simple yet efficient framework to process long video sequence, a high-quality video dataset for grounded tuning of MLLMs, and a carefully-designed instruction tuning task to explicitly incorporate the grounding supervision in the traditional QA format. Specifically, based on VideoChat, we propose our long-video MLLM, coined as VideoChat-T, by implementing a token shuffling to compress long video tokens and introducing Temporal Adaptive Position Encoding (TAPE) to enhance the temporal awareness of visual representation. Meanwhile, we introduce the TimePro, a comprehensive grounding-centric instruction tuning dataset composed of 9 tasks and 349k high-quality grounded annotations. Notably, we design a new instruction tuning task type, called Temporal Grounded Caption, to peform detailed video descriptions with the corresponding time stamps prediction. This explicit temporal location prediction will guide MLLM to correctly attend on the visual content when generating description, and thus reduce the hallucination risk caused by the LLMs. Experimental results demonstrate that our TimeSuite provides a successful solution to enhance the long video understanding capability of short-form MLLM, achieving improvement of 5.6% and 6.8% on the benchmarks of Egoschema and VideoMME, respectively. In addition, VideoChat-T exhibits robust zero-shot temporal grounding capabilities, significantly outperforming the existing state-of-the-art MLLMs. After fine-tuning, it performs on par with the traditional supervised expert models.

As video language models (VLMs) gain more applications in various scenarios, the need for robust and scalable evaluation of their performance becomes increasingly critical. The traditional human expert-based evaluation of VLMs has limitations in consistency and scalability, which sparked interest in automatic methods such as employing VLMs to evaluate VLMs. However, the reliability of VLMs as judges remains underexplored. Existing methods often rely on a single VLM as the evaluator. However, this approach can be unreliable or biased because such a model may lack the ability to fully understand the content and may have inherent biases, ultimately compromising evaluation reliability. A remedy is to apply the principle of collective thoughts, aggregating evaluations from multiple VLMs to enhance reliability. This study investigates the efficacy of such approaches, particularly when the pool of judges includes both reliable and unreliable models. Our findings reveal that incorporating collective judgments from such a mixed pool does not necessarily improve the accuracy of the final evaluation. The inclusion of less reliable judges can introduce noise, undermining the overall reliability of the outcomes. To explore the factors that impact evaluation reliability, we fine-tune an underperforming VLM judge, Video-LLaVA, and observe that improved understanding ability alone is insufficient to make VLM judges more reliable. These findings stress the limitations of collective thought approaches and highlight the need for more advanced methods that can account for the reliability of individual models. Our study promotes the development of more reliable evaluation methods for VLMs

Training language models currently requires pre-determining a fixed compute budget because the typical cosine learning rate schedule depends on the total number of steps. In contrast, the Warmup-Stable-Decay (WSD) schedule uses a constant learning rate to produce a main branch of iterates that can in principle continue indefinitely without a pre-specified compute budget. Then, given any compute budget, one can branch out from the main branch at a proper time with a rapidly decaying learning rate to produce a strong model. Empirically, WSD generates an intriguing, non-traditional loss curve: the loss remains elevated during the stable phase but sharply declines during the decay phase. Towards explaining this phenomenon, we conjecture that pretraining loss exhibits a river valley landscape, which resembles a deep valley with a river at its bottom. Under this assumption, we show that during the stable phase, the iterate undergoes large oscillations due to the high learning rate, yet it progresses swiftly along the river. During the decay phase, the rapidly dropping learning rate minimizes the iterate’s oscillations, moving it closer to the river and revealing true optimization progress. Therefore, the sustained high learning rate phase and fast decaying phase are responsible for progress in the river and the mountain directions, respectively, and are both critical. Our analysis predicts phenomenons consistent with empirical observations and shows that this landscape can naturally emerge from pretraining on a simple bi-gram dataset. Inspired by the theory, we introduce WSD-S, a variant of WSD that reuses previous checkpoints’ decay phases and keeps only one main branch, where we resume from a decayed checkpoint. WSD-S empirically outperforms WSD and Cyclic-Cosine in obtaining multiple pretrained language model checkpoints across various compute budgets in a single run for parameters scaling from 0.1B to 1.2B.

The existing video understanding benchmarks for multimodal large language models (MLLMs) mainly focus on short videos. The few benchmarks for long video understanding often rely on multiple-choice questions (MCQs). Due to the limitations of MCQ evaluations and the advanced reasoning abilities of MLLMs, models can often answer correctly by combining short video insights with elimination, without truly understanding the content. To bridge this gap, we introduce CG-Bench, a benchmark for clue-grounded question answering in long videos. CG-Bench emphasizes the model's ability to retrieve relevant clues, enhancing evaluation credibility. It includes 1,219 manually curated videos organized into 14 primary, 171 secondary, and 638 tertiary categories, making it the largest benchmark for long video analysis. The dataset features 12,129 QA pairs in three question types: perception, reasoning, and hallucination. To address the limitations of MCQ-based evaluation, we develop two novel clue-based methods: clue-grounded white box and black box evaluations, assessing whether models generate answers based on accurate video understanding. We evaluated multiple closed-source and open-source MLLMs on CG-Bench. The results show that current models struggle significantly with long videos compared to short ones, and there is a notable gap between open-source and commercial models. We hope CG-Bench will drive the development of more reliable and capable MLLMs for long video comprehension.

String processing, which mainly involves the analysis and manipulation of strings, is a fundamental component of modern computing. Despite the significant advancements of large language models (LLMs) in various natural language processing (NLP) tasks, their capability in string processing remains underexplored and underdeveloped. To bridge this gap, we present a comprehensive study of LLMs' string processing capability. In particular, we first propose StringLLM, a method to construct datasets for benchmarking string processing capability of LLMs. We use StringLLM to build a series of datasets, referred to as StringBench. It encompasses a wide range of string processing tasks, allowing us to systematically evaluate LLMs' performance in this area. Our evaluations indicate that LLMs struggle with accurately processing strings compared to humans. To uncover the underlying reasons for this limitation, we conduct an in-depth analysis and subsequently propose an effective approach that significantly enhances LLMs' string processing capability via fine-tuning. This work provides a foundation for future research to understand LLMs' string processing capability. Our code and data are available at https://github.com/wxl-lxw/StringLLM.

The performance differential of large language models (LLM) between languages hinders their effective deployment in many regions, inhibiting the potential economic and societal value of generative AI tools in many communities. However, the development of functional LLMs in many languages (i.e., multilingual LLMs) is bottlenecked by the lack of high-quality evaluation resources in languages other than English. Moreover, current practices in multilingual benchmark construction often translate English resources, ignoring the regional and cultural knowledge of the environments in which multilingual systems would be used. In this work, we construct an evaluation suite of 197,243 QA pairs from local exam sources to measure the capabilities of multilingual LLMs in a variety of regional contexts. Our novel resource, INCLUDE, is a comprehensive knowledge- and reasoning-centric benchmark across 44 written languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed.

We introduce Duoduo CLIP, a model for 3D representation learning that learns shape encodings from multi-view images instead of point-clouds. The choice of multi-view images allows us to leverage 2D priors from off-the-shelf CLIP models to facilitate fine-tuning with 3D data. Our approach not only shows better generalization compared to existing point cloud methods, but also reduces GPU requirements and training time. In addition, the model is modified with cross-view attention to leverage information across multiple frames of the object which further boosts performance. Notably, our model is permutation invariant to the order of multi-view images while being pose-free. Compared to the current SOTA point cloud method that requires 480 A100 hours to train 1 billion model parameters we only require 57 A5000 hours and 87 million parameters. Multi-view images also provide more flexibility including being able to encode objects with a variable number of images, and performance scales when more views are used. In contrast, point cloud based methods require an entire scan or model of the object. We showcase this flexibility with benchmarks from images of real-world objects. Our model also achieves better performance in more fine-grained text to shape retrieval, demonstrating better text-and-shape alignment than point cloud based models.

Large language models have demonstrated an impressive ability to perform factual recall. Prior work has found that transformers trained on factual recall tasks can store information at a rate proportional to their parameter count. In our work, we show that shallow transformers can use a combination of associative memories to obtain such near optimal storage capacity. We begin by proving that the storage capacities of both linear and MLP associative memories scale linearly with parameter count. We next introduce a synthetic factual recall task, and prove that a transformer with a single layer of self-attention followed by an MLP can obtain 100\% accuracy on the task whenever either the total number of self-attention parameters or MLP parameters scales (up to log factors) linearly with the number of facts. In particular, the transformer can trade off between using the value matrices or the MLP as an associative memory to store the dataset of facts. We complement these expressivity results with an analysis of the gradient flow trajectory of a simplified linear attention model trained on our factual recall task, where we show that the model exhibits sequential learning behavior.

In this paper, we study the problem of video-conditioned policy learning. While previous works mostly focus on learning policies that perform a single skill specified by the given video, we take a step further and aim to learn a policy that can perform multiple skills according to the given video, and generalize to unseen videos by recombining these skills. To solve this problem, we propose our algorithm, Watch-Less-Do-More, an information bottleneck-based imitation learning framework for implicit skill discovery and video-conditioned policy learning. In our method, an information bottleneck objective is employed to control the information contained in the video representation, ensuring that it only encodes information relevant to the current skill (Watch-Less). By discovering potential skills from training videos, the learned policy is able to recombine them and generalize to unseen videos to achieve compositional generalization (Do-More). To evaluate our method, we perform extensive experiments in various environments and show that our algorithm substantially outperforms baselines (up to 2x) in terms of compositional generalization ability.

We present Pyramid Attention Broadcast (PAB), a real-time, high quality and training-free approach for DiT-based video generation. Our method is founded on the observation that attention difference in the diffusion process exhibits a U-shaped pattern, indicating significant redundancy. We mitigate this by broadcasting attention outputs to subsequent steps in a pyramid style. It applies different broadcast strategies to each attention based on their variance for best efficiency. We further introduce broadcast sequence parallel for more efficient distributed inference. PAB demonstrates up to 10.5x speedup across three models compared to baselines, achieving real-time generation for up to 720p videos. We anticipate that our simple yet effective method will serve as a robust baseline and facilitate future research and application for video generation.

Communication is defined as "*Who* says *what* to *whom* with *what* effect." A message from a communicator generates downstream receiver effects, also known as behavior. Receiver behavior, being a downstream effect of the message, carries rich signals about it. Even after carrying signals about the message, the behavior signal is often ignored while training vision-language models. We show that training VLMs on receiver behavior can actually help improve their content-understanding abilities. We demonstrate that training VLMs to predict receiver behaviors, such as likes, comments, and replay graphs, which are available at scale, enhances the VLM's performance across a broad range of downstream content understanding tasks. We show this performance increase over 6 types of behavior, 46 different tasks covering image, video, text, and audio over 26 benchmark datasets across both zero-shot and fine-tuning settings, outperforming many supervised baselines on diverse tasks ranging from emotion recognition to captioning by up to 150%. We note that since receiver behavior, such as likes, comments, and replay graphs, is collected by default on the internet and does not need any human annotations to be useful, the performance improvement we get after training on this data is essentially free lunch. We also release **BLIFT**, our **Behaviour-LLaVA IFT** dataset comprising 730k images and videos with their receiver behavior collected from multiple platforms on which we train our models to achieve this. The dataset and code are available at [behavior-in-the-wild.github.io/behavior-llava](https://behavior-in-the-wild.github.io/behavior-llava).

Deep Generative Models are frequently used to learn continuous representations of complex data distributions by training on a finite number of samples. For any generative model, including pre-trained foundation models with Diffusion or Transformer architectures, generation performance can significantly vary across the learned data manifold. In this paper, we study the local geometry of the learned manifold and its relationship to generation outcomes for a wide range of generative models, including DDPM, Diffusion Transformer (DiT), and Stable Diffusion 1.4. Building on the theory of continuous piecewise-linear (CPWL) generators, we characterize the local geometry in terms of three geometric descriptors - scaling ($\psi$), rank ($\nu$), and complexity/un-smoothness ($\delta$). We provide quantitative and qualitative evidence showing that for a given latent vector, the local descriptors are indicative of post-generation aesthetics, generation diversity, and memorization by the generative model. Finally, we demonstrate that by training a reward model on the 'local scaling' for Stable Diffusion, we can self-improve both generation aesthetics and diversity using geometry sensitive guidance during denoising. Website: https://imtiazhumayun.github.io/generative_geometry.

Self-supervised learning has driven significant progress in learning from single-subject, _iconic_ images. However, there are still unanswered questions about the use of minimally-curated, naturalistic video data, which contain _dense_ scenes with many independent objects, imbalanced class distributions, and varying object sizes. In this paper, we propose PooDLe, a self-supervised learning method that combines an invariance-based objective on pooled representations with a dense SSL objective that enforces equivariance to optical flow warping. Our results show that a unified objective applied at multiple feature scales is essential for learning effective image representations from naturalistic videos. We validate our method with experiments on the BDD100K driving video dataset and the Walking Tours first-person video dataset, demonstrating its ability to capture spatial understanding from a dense objective and semantic understanding via a pooled representation objective.

We propose a new task and model for dense video object captioning -- detecting, tracking and captioning trajectories of objects in a video. This task unifies spatial and temporal localization in video, whilst also requiring fine-grained visual understanding that is best described by natural language. We propose a unified model, and demonstrate how our end-to-end approach is more accurate and temporally coherent than a multi-stage pipeline combining state-of-the-art detection, tracking, and captioning models. Moreover, we propose a training strategy based on a mixture of disjoint tasks, which allows us to leverage diverse, large-scale datasets which supervise different parts of our model. Although each pretraining task only provides weak supervision, they are complementary and, when combined, result in noteworthy zero-shot ability and serve as strong initialization for additional finetuning to further improve accuracy. We carefully design new metrics capturing all components of our task, and show how we can repurpose existing video grounding datasets (e.g. VidSTG and VLN) for our new task. We show that our model improves upon a number of strong baselines for this new task. Furthermore, we can apply our model to the task of spatial grounding, outperforming prior state-of-the-art on VidSTG and VLN, without explicitly training for it. Our code is available at https://github.com/google-research/scenic.

Jailbreaking attacks can effectively manipulate open-source large language models (LLMs) to produce harmful responses. However, these attacks exhibit limited transferability, failing to disrupt proprietary LLMs consistently. To reliably identify vulnerabilities in proprietary LLMs, this work investigates the transferability of jailbreaking attacks by analysing their impact on the model's intent perception. By incorporating adversarial sequences, these attacks can redirect the source LLM's focus away from malicious-intent tokens in the original input, thereby obstructing the model's intent recognition and eliciting harmful responses. Nevertheless, these adversarial sequences fail to mislead the target LLM's intent perception, allowing the target LLM to refocus on malicious-intent tokens and abstain from responding. Our analysis further reveals the inherent $\textit{distributional dependency}$ within the generated adversarial sequences, whose effectiveness stems from overfitting the source LLM's parameters, resulting in limited transferability to target LLMs. To this end, we propose the Perceived-importance Flatten (PiF) method, which uniformly disperses the model's focus across neutral-intent tokens in the original input, thus obscuring malicious-intent tokens without relying on overfitted adversarial sequences. Extensive experiments demonstrate that PiF provides an effective and efficient red-teaming evaluation for proprietary LLMs.

As deep generative models have progressed, recent work has shown them to be capable of memorizing and reproducing training datapoints when deployed. These findings call into question the usability of generative models, especially in light of the legal and privacy risks brought about by memorization. To better understand this phenomenon, we propose the *manifold memorization hypothesis* (MMH), a geometric framework which leverages the manifold hypothesis into a clear language in which to reason about memorization. We propose to analyze memorization in terms of the relationship between the dimensionalities of $(i)$ the ground truth data manifold and $(ii)$ the manifold learned by the model. This framework provides a formal standard for "how memorized" a datapoint is and systematically categorizes memorized data into two types: memorization driven by overfitting and memorization driven by the underlying data distribution. By analyzing prior work in the context of the MMH, we explain and unify assorted observations in the literature. We empirically validate the MMH using synthetic data and image datasets up to the scale of Stable Diffusion, developing new tools for detecting and preventing generation of memorized samples in the process.

It has recently been demonstrated empirically that in-context learning emerges in transformers when certain distributional properties are present in the training data, but this ability can also diminish upon further training. We provide a new theoretical understanding of these phenomena by identifying simplified distributional properties that give rise to the emergence and eventual disappearance of in-context learning. We do so by first analyzing a simplified model that uses a gating mechanism to choose between an in-weight and an in-context predictor. Through a combination of a generalization error and regret analysis we identify conditions where in-context and in-weight learning emerge. These theoretical findings are then corroborated experimentally by comparing the behaviour of a full transformer on the simplified distributions to that of the stylized model, demonstrating aligned results. We then extend the study to a full large language model, showing how fine-tuning on various collections of natural language prompts can elicit similar in-context and in-weight learning behaviour.