KV cache has become a de facto technique for the inference of large language models (LLMs), where tensors of shape (layer number, head number, sequence length, feature dimension) are introduced to cache historical information for self-attention. As the size of the model and data grows, the KV cache can quickly become a bottleneck within the system in both storage and memory transfer. To address this, prior studies usually focus on the first three axes of the cache tensors for compression. This paper supplements them, focusing on the feature dimension axis, by utilizing low-rank projection matrices to transform the cache features into spaces with reduced dimensions. We begin by investigating the canonical orthogonal projection method for data compression through principal component analysis (PCA). We observe the issue with PCA projection where significant performance degradation is observed at low compression rates. To bridge the gap, we propose to directly tune the orthogonal projection matrices with a distillation objective using an elaborate Matryoshka training strategy. After training, we adaptively search for the optimal compression rates for various layers and heads given varying compression budgets. Compared to previous works, our method can easily embrace pre-trained LLMs and hold a smooth tradeoff between performance and compression rate. We empirically witness the high data efficiency of our training procedure and find that our method can sustain over 90% performance with an average KV cache compression rate of 60% (and up to 75% in certain extreme scenarios) for popular LLMs like LLaMA2-7B-base and Mistral-7B-v0.3-base.

Proprietary large language models (LLMs) demonstrate exceptional generalization ability across various tasks. Additionally, deploying LLMs on edge devices is trending for efficiency and privacy reasons. However, edge deployment of proprietary LLMs introduces new security threats: attackers who obtain an edge-deployed LLM can easily use it as a base model for various tasks due to its high generalization ability, which we call foundational capability stealing. Unfortunately, existing model protection mechanisms are often task-specific and fail to protect general-purpose LLMs, as they mainly focus on protecting task-related parameters using trusted execution environments (TEEs). Although some recent TEE-based methods are able to protect the overall model parameters in a computation-efficient way, they still suffer from prohibitive communication costs between TEE and CPU/GPU, making it impractical to deploy for edge LLMs. To protect the foundational capabilities of edge LLMs, we propose CoreGuard, a computation- and communication-efficient model protection approach against model stealing on edge devices. The core component of CoreGuard is a lightweight and propagative authorization module residing in TEE. Extensive experiments show that CoreGuard achieves the same security protection as the black-box security guarantees with negligible overhead.

Online reviews play a pivotal role in influencing consumer decisions across various domains, from purchasing products to selecting hotels or restaurants. However, the sheer volume of reviews -- often containing repetitive or irrelevant content -- leads to information overload, making it challenging for users to extract meaningful insights. Traditional opinion summarization models face challenges in handling long inputs and large volumes of reviews, while newer Large Language Model (LLM) approaches often fail to generate accurate and faithful summaries. To address those challenges, this paper introduces (1) a new dataset of long-form user reviews, each entity comprising over a thousand reviews, (2) two training-free LLM-based summarization approaches that scale to long inputs, and (3) automatic evaluation metrics. Our dataset of user reviews is paired with in-depth and unbiased critical summaries by domain experts, serving as a reference for evaluation. Additionally, our novel reference-free evaluation metrics provide a more granular, context-sensitive assessment of summary faithfulness. We benchmark several open-source and closed-source LLMs using our methods. Our evaluation reveals that LLMs still face challenges in balancing sentiment and format adherence in long-form summaries, though open-source models can narrow the gap when relevant information is retrieved in a focused manner.

Large Language Models (LLMs) have shown remarkable capabilities in natural language processing but exhibit significant performance gaps among different languages. Most existing approaches to address these disparities rely on pretraining or fine-tuning, which are resource-intensive. To overcome these limitations without incurring significant costs, we propose Inference-Time Cross-Lingual Intervention (INCLINE), a novel framework that enhances LLM performance on low-performing (source) languages by aligning their internal representations with those of high-performing (target) languages during inference. INCLINE initially learns alignment matrices using parallel sentences from source and target languages through a Least-Squares optimization, and then applies these matrices during inference to transform the low-performing language representations toward the high-performing language space. Extensive experiments on nine benchmarks with five LLMs demonstrate that INCLINE significantly improves performance across diverse tasks and languages, compared to recent strong baselines. Our analysis demonstrates that INCLINE is highly cost-effective and applicable to a wide range of applications. In addition, we release the code to foster research along this line: https://github.com/weixuan-wang123/INCLINE.

Prompt learning for pre-trained Vision-Language Models (VLMs) like CLIP has demonstrated potent applicability across diverse downstream tasks. This lightweight approach has quickly gained traction from federated learning (FL) researchers who seek to efficiently adapt VLMs to heterogeneous scenarios. However, current federated prompt learning methods are habitually restricted to the traditional FL paradigm, where the participating clients are generally only allowed to download a single globally aggregated model from the server. While justifiable for training full-sized models under federated settings, in this work, we argue that this paradigm is ill-suited for lightweight prompts. By facilitating the clients to download multiple pre-aggregated prompts as fixed non-local experts, we propose Personalized Federated Mixture of Adaptive Prompts (pFedMoAP), a novel FL framework that personalizes the prompt learning process through the lens of Mixture of Experts (MoE). pFedMoAP implements a local attention-based gating network that learns to generate enhanced text features for better alignment with local image data on the client, benefiting from both local and downloaded non-local adaptive prompt experts. The non-local experts are sparsely selected from a server-maintained pool, fostering collaborative learning across clients. To evaluate the proposed algorithm, we conduct extensive experiments across 9 datasets under various heterogeneous federated settings. The results show that pFedMoAP consistently outperforms the state-of-the-art alternatives, underscoring its efficacy in personalizing prompt learning for CLIP within the federated learning paradigm.

This work explores the integration of ontology-based reasoning and Machine Learning techniques for explainable value classification. By relying on an ontological formalization of moral values as in the Moral Foundations Theory, relying on the DnS Ontology Design Pattern, the \textit{sandra} neuro-symbolic reasoner is used to infer values (fomalized as descriptions) that are \emph{satisfied by} a certain sentence. Sentences, alongside their structured representation, are automatically generated using an open-source Large Language Model. The inferred descriptions are used to automatically detect the value associated with a sentence. We show that only relying on the reasoner's inference results in explainable classification comparable to other more complex approaches. We show that combining the reasoner's inferences with distributional semantics methods largely outperforms all the baselines, including complex models based on neural network architectures. Finally, we build a visualization tool to explore the potential of theory-based values classification, which is publicly available at http://xmv.geomeaning.com/.

This paper presents an analysis of biases in open-source Large Language Models (LLMs) across various genders, religions, and races. We introduce a methodology for generating a bias detection dataset using seven bias triggers: General Debate, Positioned Debate, Career Advice, Story Generation, Problem-Solving, Cover-Letter Writing, and CV Generation. We use GPT-4o to generate a diverse set of prompts for each trigger across various genders, religious and racial groups. We evaluate models from Llama and Gemma family on the generated dataset. We anonymise the LLM-generated text associated with each group using GPT-4o-mini and do a pairwise comparison using GPT-4o-as-a-Judge. To quantify bias in the LLM-generated text we use the number of wins and losses in the pairwise comparison. Our analysis spans three languages, English, German, and Arabic to explore how language influences bias manifestation. Our findings reveal that LLMs exhibit strong polarization toward certain groups across each category, with a notable consistency observed across models. However, when switching languages, variations and anomalies emerge, often attributable to cultural cues and contextual differences.

Transformer-based Large Language Models (LLMs) traditionally rely on final-layer loss for training and final-layer representations for predictions, potentially overlooking the predictive power embedded in intermediate layers. Surprisingly, we find that focusing training efforts on these intermediate layers can yield training losses comparable to those of final layers, with complementary test-time performance. We introduce a novel tuning framework, Mixture-of-Depths (MoD), which trains late layers as ensembles contributing to the final logits through learned routing weights. With the auxiliary distillation loss and additional normalization modules, we ensure that the outputs of the late layers adapt to language modeling. Our MoD framework, which can be integrated with any existing tuning method, shows consistent improvement on various language modelling tasks. Furthermore, by replacing traditional trainable modules with MoD, our approach achieves similar performance with significantly fewer trainable parameters, demonstrating the potential of leveraging predictive power from intermediate representations during training.

A lot of recent progress in deep learning, particularly in the form of large language models (LLMs) has been driven by the transformer architecture [Vaswani et al., 2017]. While these models have great quality, it comes at a computation cost which scales quadratically in sequence length - both during training and inference. This can become prohibitive for very long contexts and as such a number of alternative architectures with better computational scaling in context length have been proposed [Gu and Dao, 2023, Poli et al., 2023, Fu et al., 2024]. While most of these works have improved computational efficiency for training, some still scale quadratically in sequence length when it comes to inference, thus not improving asymptotically over transformers. In this work, we propose a framework for optimizing inference efficiency for a general class of such models. As a case study, which inspired the method, we focus on long convolution sequence models (LCSMs) [Poli et al., 2023, Fu et al., 2022, Romero et al., 2021, Li et al., 2022, Karami and Ghodsi, 2024, Fu et al., 2023a]. However, our approach is not limited to LCSMs alone and we identify the properties that allow for such inference speedups in hope to guide the design of future architectures. In the particular case of LCSMs (including Hyena), the building block of the architecture is that of convolving the input sequence with a sequence-length long, (potentially underparameterized) filter. If we let L be the sequence length (e.g.

Reasoning, the ability to logically draw conclusions from existing knowledge, is a hallmark of human. Together with perception, they constitute the two major themes of artificial intelligence. While deep learning has pushed the limit of perception beyond human-level performance, the progress in reasoning domains is way behind. One fundamental reason is that reasoning problems usually have flexible structures for both knowledge and queries, and many existing models only perform well on structures seen during training. Here we aim to push the boundary of reasoning models by devising algorithms that generalize across knowledge and query structures, as well as systems that accelerate development on structured data. This thesis consists of three parts. In Part I, we study models that can inductively generalize to unseen knowledge graphs with new entity and relation vocabularies. For new entities, we propose a framework that learns neural operators in a dynamic programming algorithm computing path representations. For relations, we construct a relation graph to capture the interactions between relations, thereby converting new relations into new entities. In Part II, we propose two solutions for generalizing across multi-step queries on knowledge graphs and text respectively. For knowledge graphs, we show that multi-step queries can be solved by multiple calls of graph neural networks and fuzzy logic operations. For text, we devise an algorithm to learn explicit knowledge as textual rules to improve large language models on multi-step queries. In Part III, we propose two systems to facilitate machine learning development on structured data. Our library treats structured data as first-class citizens and removes the barrier for developing algorithms on structured data. Our node embedding system solves the GPU memory bottleneck of embedding matrices and scales to graphs with billion nodes.