Vision language models (VLMs) demonstrate strong capabilities in jointly processing visual and textual data. However, they often incur substantial computational overhead due to redundant visual information, particularly in long-form video scenarios. Existing approaches predominantly focus on either vision token pruning, which may overlook spatio-temporal dependencies, or keyframe selection, which identifies informative frames but discards others, thus disrupting contextual continuity. In this work, we propose KVTP (Keyframe-oriented Vision Token Pruning), a novel framework that overcomes the drawbacks of token pruning and keyframe selection. By adaptively assigning pruning rates based on frame relevance to the query, KVTP effectively retains essential contextual information while significantly reducing redundant computation. To thoroughly evaluate the long-form video understanding capacities of VLMs, we curated and reorganized subsets from VideoMME, EgoSchema, and NextQA into a unified benchmark named SparseKV-QA that highlights real-world scenarios with sparse but crucial events. Our experiments with VLMs of various scales show that KVTP can reduce token usage by 80% without compromising spatiotemporal and contextual consistency, significantly cutting computation while maintaining the performance. These results demonstrate our approach's effectiveness in efficient long-video processing, facilitating more scalable VLM deployment.

With the rise of large language models (LLMs), increasing research has recognized their risk of leaking personally identifiable information (PII) under malicious attacks. Although efforts have been made to protect PII in LLMs, existing methods struggle to balance privacy protection with maintaining model utility. In this paper, inspired by studies of amnesia in cognitive science, we propose a novel approach, Proactive Privacy Amnesia (PPA), to safeguard PII in LLMs while preserving their utility. This mechanism works by actively identifying and forgetting key memories most closely associated with PII in sequences, followed by a memory implanting using suitable substitute memories to maintain the LLM's functionality. We conduct evaluations across multiple models to protect common PII, such as phone numbers and physical addresses, against prevalent PII-targeted attacks, demonstrating the superiority of our method compared with other existing defensive techniques. The results show that our PPA method completely eliminates the risk of phone number exposure by 100% and significantly reduces the risk of physical address exposure by 9.8% - 87.6%, all while maintaining comparable model utility performance. Large Language Models (LLMs) (Touvron et al., 2023; Achiam et al., 2023; Team et al., 2023; Dubey et al., 2024) have achieved remarkable success in recent years, with their wide adoption either as general-purpose models or, after fine-tuning, as specialized and personal assistants. Despite their success, LLMs with huge parameter counts and great capacity in the meantime exhibit the concerning "memorization" phenomenons (Carlini et al., 2019; 2021), i.e., they can precisely memorize some training data. Such memorization is vulnerable to various attacks (e.g., membership inference attacks and data extraction attacks) and risks severe privacy breaches. One of the most serious concerns comes from the attacks that aim to extract personal identifiable information (PII) memorized by the models, which compromise users' privacy and are likely to cause real-world harm consequently. To defend against such PII or data extraction attacks, several machine unlearning techniques have been applied to LLMs. However, existing methods typically fall short in terms of the trade-off between the defense performance and model utility. For example, most unlearning approaches are based on gradient ascent (Jang et al., 2022; Wang et al., 2024) and often adversely affect model functionalities to an extent where the model cannot handle their original tasks anymore and thus becomes no longer useful.

Recently, Text-to-Image (T2I) generation models have achieved significant advancements. Correspondingly, many automated metrics have emerged to evaluate the image-text alignment capabilities of generative models. However, the performance comparison among these automated metrics is limited by existing small datasets. Additionally, these datasets lack the capacity to assess the performance of automated metrics at a fine-grained level. In this study, we contribute an EvalMuse-40K benchmark, gathering 40K image-text pairs with fine-grained human annotations for image-text alignment-related tasks. In the construction process, we employ various strategies such as balanced prompt sampling and data re-annotation to ensure the diversity and reliability of our benchmark. This allows us to comprehensively evaluate the effectiveness of image-text alignment metrics for T2I models. Meanwhile, we introduce two new methods to evaluate the image-text alignment capabilities of T2I models: FGA-BLIP2 which involves end-to-end fine-tuning of a vision-language model to produce fine-grained image-text alignment scores and PN-VQA which adopts a novel positive-negative VQA manner in VQA models for zero-shot fine-grained evaluation. Both methods achieve impressive performance in image-text alignment evaluations. We also use our methods to rank current AIGC models, in which the results can serve as a reference source for future study and promote the development of T2I generation. The data and code will be made publicly available.

These benchmarks contain only short audio clips and thus do not reflect the complexity of achieving long-context Speech Large Language Models (Speech LLMs) represent a understanding and extracting precise information from lengthy significant advancement in speech language understanding and audio sequences. To systematically assess the unique challenges processing, as they leverage contextual reasoning capabilities of posed by SIR, we present SPIRAL (Speech Informational large language models to process audio inputs [1]. Unlike traditional Retrieval and Lookup), a 1,012-sample benchmark specifically cascaded pipelines, where automatic speech recognition crafted to evaluate Speech LLM performance on long-form (ASR) and language modeling are handled by separate modules, audio sequences (around 90 seconds in duration). On a high Speech LLMs unify audio processing, cross-modal fusion, and level, SPIRAL constructs SIR questions by embedding a critical language modeling in a single architecture [2]. These unified piece of information within lengthy and potentially distracting models can perform multiple tasks like speech recognition, dialogues, thereby assessing the model ability to pinpoint and speech translation, speaker identification and emotion recognition, retrieve essential content from long-form inputs.

Model merging is an emerging technique that integrates multiple models fine-tuned on different tasks to create a versatile model that excels in multiple domains. This scheme, in the meantime, may open up backdoor attack opportunities where one single malicious model can jeopardize the integrity of the merged model. Existing works try to demonstrate the risk of such attacks by assuming substantial computational resources, focusing on cases where the attacker can fully fine-tune the pre-trained model. Such an assumption, however, may not be feasible given the increasing size of machine learning models. In practice where resources are limited and the attacker can only employ techniques like Low-Rank Adaptation (LoRA) to produce the malicious model, it remains unclear whether the attack can still work and pose threats. In this work, we first identify that the attack efficacy is significantly diminished when using LoRA for fine-tuning. Then, we propose LoBAM, a method that yields high attack success rate with minimal training resources. The key idea of LoBAM is to amplify the malicious weights in an intelligent way that effectively enhances the attack efficacy. We demonstrate that our design can lead to improved attack success rate through both theoretical proof and extensive empirical experiments across various model merging scenarios. Moreover, we show that our method has strong stealthiness and is difficult to detect.

The rapid development of large language models (LLMs) has significantly transformed the field of artificial intelligence, demonstrating remarkable capabilities in natural language processing and moving towards multi-modal functionality. These models are increasingly integrated into diverse applications, impacting both research and industry. However, their development and deployment present substantial challenges, including the need for extensive computational resources, high energy consumption, and complex software optimizations. Unlike traditional deep learning systems, LLMs require unique optimization strategies for training and inference, focusing on system-level efficiency. This paper surveys hardware and software co-design approaches specifically tailored to address the unique characteristics and constraints of large language models. This survey analyzes the challenges and impacts of LLMs on hardware and algorithm research, exploring algorithm optimization, hardware design, and system-level innovations. It aims to provide a comprehensive understanding of the trade-offs and considerations in LLM-centric computing systems, guiding future advancements in AI. Finally, we summarize the existing efforts in this space and outline future directions toward realizing production-grade co-design methodologies for the next generation of large language models and AI systems.

Large language models (LLMs) have demonstrated remarkable proficiency in a range of natural language processing tasks. Once deployed, LLMs encounter users with personalized factual knowledge, and such personalized knowledge is consistently reflected through users' interactions with the LLMs. To enhance user experience, real-time model personalization is essential, allowing LLMs to adapt user-specific knowledge based on user feedback during human-LLM interactions. Existing methods mostly require back-propagation to finetune the model parameters, which incurs high computational and memory costs. In addition, these methods suffer from low interpretability, which will cause unforeseen impacts on model performance during long-term use, where the user's personalized knowledge is accumulated extensively.To address these challenges, we propose Knowledge Graph Tuning (KGT), a novel approach that leverages knowledge graphs (KGs) to personalize LLMs. KGT extracts personalized factual knowledge triples from users' queries and feedback and optimizes KGs without modifying the LLM parameters. Our method improves computational and memory efficiency by avoiding back-propagation and ensures interpretability by making the KG adjustments comprehensible to humans.Experiments with state-of-the-art LLMs, including GPT-2, Llama2, and Llama3, show that KGT significantly improves personalization performance while reducing latency and GPU memory costs. Ultimately, KGT offers a promising solution of effective, efficient, and interpretable real-time LLM personalization during user interactions with the LLMs.

Despite improved performance, existing methods (including the state-of-the-art, Min-K%) are mostly developed upon simple heuristics and lack solid, reasonable foundations. In this work, we propose a novel and theoretically motivated methodology for pre-training data detection, named Min-K%++. Specifically, we present a key insight that training samples tend to be local maxima of the modeled distribution along each input dimension through maximum likelihood training, which in turn allow us to insightfully translate the problem into identification of local maxima. Then, we design our method accordingly that works under the discrete distribution modeled by LLMs, whose core idea is to determine whether the input forms a mode or has relatively high probability under the conditional categorical distribution. Empirically, the proposed method achieves new SOTA performance across multiple settings. On the WikiMIA benchmark, Min-K%++ outperforms the runner-up by 6.2% to 10.5% in detection AUROC averaged over five models. On the more challenging MIMIR benchmark, it consistently improves upon reference-free methods while performing on par with reference-based method that requires an extra reference model.

Data heterogeneity presents significant challenges for federated learning (FL). Recently, dataset distillation techniques have been introduced, and performed at the client level, to attempt to mitigate some of these challenges. In this paper, we propose a highly efficient FL dataset distillation framework on the server side, significantly reducing both the computational and communication demands on local devices while enhancing the clients' privacy. Unlike previous strategies that perform dataset distillation on local devices and upload synthetic data to the server, our technique enables the server to leverage prior knowledge from pre-trained deep generative models to synthesize essential data representations from a heterogeneous model architecture. This process allows local devices to train smaller surrogate models while enabling the training of a larger global model on the server, effectively minimizing resource utilization. We substantiate our claim with a theoretical analysis, demonstrating the asymptotic resemblance of the process to the hypothetical ideal of completely centralized training on a heterogeneous dataset. Empirical evidence from our comprehensive experiments indicates our method's superiority, delivering an accuracy enhancement of up to 40% over non-dataset-distillation techniques in highly heterogeneous FL contexts, and surpassing existing dataset-distillation methods by 18%. In addition to the high accuracy, our framework converges faster than the baselines because rather than the server trains on several sets of heterogeneous data distributions, it trains on a multi-modal distribution. Our code is available at https://github.com/FedDG23/FedDG-main.git

Mixture-of-Experts (MoE) has emerged as a favorable architecture in the era of large models due to its inherent advantage, i.e., enlarging model capacity without incurring notable computational overhead. Yet, the realization of such benefits often results in ineffective GPU memory utilization, as large portions of the model parameters remain dormant during inference. Moreover, the memory demands of large models consistently outpace the memory capacity of contemporary GPUs. Addressing this, we introduce SiDA (Sparsity-inspired Data-Aware), an efficient inference approach tailored for large MoE models. SiDA judiciously exploits both the system's main memory, which is now abundant and readily scalable, and GPU memory by capitalizing on the inherent sparsity on expert activation in MoE models. By adopting a data-aware perspective, SiDA achieves enhanced model efficiency with a neglectable performance drop. Specifically, SiDA attains a remarkable speedup in MoE inference with up to 3.93X throughput increasing, up to 75% latency reduction, and up to 80% GPU memory saving with down to 1% performance drop. This work paves the way for scalable and efficient deployment of large MoE models, even in memory-constrained systems.