We apply language modeling techniques to detect ransomware activity in NVMe command sequences. We design and train two types of transformer-based models: the Command-Level Transformer (CLT) performs in-context token classification to determine whether individual commands are initiated by ransomware, and the Patch-Level Transformer (PLT) predicts the volume of data accessed by ransomware within a patch of commands. We present both model designs and the corresponding tokenization and embedding schemes and show that they improve over state-of-the-art tabular methods by up to 24% in missed-detection rate, 66% in data loss prevention, and 84% in identifying data accessed by ransomware.
Disaggregated Large Language Model (LLM) inference has gained popularity as it separates the computation-intensive prefill stage from the memory-intensive decode stage, avoiding the prefill-decode interference and improving resource utilization. However, transmitting Key-Value (KV) data between the two stages can be a bottleneck, especially for long prompts. Additionally, the computation time overhead for prefill and decode is key for optimizing Job Completion Time (JCT), and KV data size can become prohibitive for long prompts and sequences. Existing KV quantization methods can alleviate the transmission bottleneck and reduce memory requirements, but they introduce significant dequantization overhead, exacerbating the computation time. We propose Homomorphic Acceleration via Compression of the KV cache (HACK) for disaggregated LLM inference. HACK eliminates the heavy KV dequantization step, and directly performs computations on quantized KV data to approximate and reduce the cost of the expensive matrix-multiplication step. Extensive trace-driven experiments show that HACK reduces JCT by up to 70.9% compared to disaggregated LLM inference baseline and by up to 52.3% compared to state-of-the-art KV quantization methods.
This paper examines the ethical question, 'What is a good search engine?' Since search engines are gatekeepers of global online information, it is vital they do their job ethically well. While the Internet is now several decades old, the topic remains under-explored from interdisciplinary perspectives. This paper presents a novel role-based approach involving four ethical models of types of search engine behavior: Customer Servant, Librarian, Journalist, and Teacher. It explores these ethical models with reference to the research field of information retrieval, and by means of a case study involving the COVID-19 global pandemic. It also reflects on the four ethical models in terms of the history of search engine development, from earlier crude efforts in the 1990s, to the very recent prospect of Large Language Model-based conversational information seeking systems taking on the roles of established web search engines like Google. Finally, the paper outlines considerations that inform present and future regulation and accountability for search engines as they continue to evolve. The paper should interest information retrieval researchers and others interested in the ethics of search engines.
Emerging AI accelerators increasingly adopt wafer-scale manufacturing technologies, integrating hundreds of thousands of AI cores in a mesh architecture with large distributed on-chip memory (tens of GB in total) and ultra-high on-chip memory bandwidth (tens of PB/s). However, current LLM inference systems, optimized for shared memory architectures like GPUs, fail to exploit these accelerators fully. We introduce WaferLLM, the first wafer-scale LLM inference system. WaferLLM is guided by a novel PLMR model (pronounced as "Plummer") that captures the unique hardware characteristics of wafer-scale architectures. Leveraging this model, WaferLLM pioneers wafer-scale LLM parallelism, optimizing the utilization of hundreds of thousands of on-chip cores. It also introduces MeshGEMM and MeshGEMV, the first GEMM and GEMV implementations designed to scale effectively on wafer-scale accelerators. Evaluations show that WaferLLM achieves up to 200$\times$ higher accelerator utilization than state-of-the-art methods. Leveraging a wafer-scale accelerator (Cerebras WSE2), WaferLLM delivers GEMV operations 606$\times$ faster and 16$\times$ more energy-efficient than on an NVIDIA A100 GPU. For full LLM inference, WaferLLM achieves 10-20$\times$ speedups over A100 GPU clusters running SGLang and vLLM. These advantages are expected to grow as wafer-scale AI models, software, and hardware continue to mature. WaferLLM is open-sourced at https://github.com/MeshInfra/WaferLLM.
As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as "AI Oversight". We study how model similarity affects both aspects of AI oversight by proposing a probabilistic metric for LM similarity based on overlap in model mistakes. Using this metric, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from "weak-to-strong generalization". As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.
Large Language Model (LLM)-empowered multi-agent systems extend the cognitive boundaries of individual agents through disciplined collaboration and interaction, while constructing these systems often requires labor-intensive manual designs. Despite the availability of methods to automate the design of agentic workflows, they typically seek to identify a static, complex, one-size-fits-all system, which, however, fails to dynamically allocate inference resources based on the difficulty and domain of each query. To address this challenge, we shift away from the pursuit of a monolithic agentic system, instead optimizing the \textbf{agentic supernet}, a probabilistic and continuous distribution of agentic architectures. We introduce MaAS, an automated framework that samples query-dependent agentic systems from the supernet, delivering high-quality solutions and tailored resource allocation (\textit{e.g.}, LLM calls, tool calls, token cost). Comprehensive evaluation across six benchmarks demonstrates that MaAS \textbf{(I)} requires only $6\sim45\%$ of the inference costs of existing handcrafted or automated multi-agent systems, \textbf{(II)} surpasses them by $0.54\%\sim11.82\%$, and \textbf{(III)} enjoys superior cross-dataset and cross-LLM-backbone transferability.
The challenge of formal proof generation has a rich history, but with modern techniques, we may finally be at the stage of making actual progress in real-life mathematical problems. This paper explores the integration of ChatGPT and basic searching techniques to simplify generating formal proofs, with a particular focus on the miniF2F dataset. We demonstrate how combining a large language model like ChatGPT with a formal language such as Lean, which has the added advantage of being verifiable, enhances the efficiency and accessibility of formal proof generation. Despite its simplicity, our best-performing Lean-based model surpasses all known benchmarks with a 31.15% pass rate. We extend our experiments to include other datasets and employ alternative language models, showcasing our models' comparable performance in diverse settings and allowing for a more nuanced analysis of our results. Our findings offer insights into AI-assisted formal proof generation, suggesting a promising direction for future research in formal mathematical proof.
The design of analog circuits is a cornerstone of integrated circuit (IC) development, requiring the optimization of complex, interconnected sub-structures such as amplifiers, comparators, and buffers. Traditionally, this process relies heavily on expert human knowledge to refine design objectives by carefully tuning sub-components while accounting for their interdependencies. Existing methods, such as Bayesian Optimization (BO), offer a mathematically driven approach for efficiently navigating large design spaces. However, these methods fall short in two critical areas compared to human expertise: (i) they lack the semantic understanding of the sizing solution space and its direct correlation with design objectives before optimization, and (ii) they fail to reuse knowledge gained from optimizing similar sub-structures across different circuits. To overcome these limitations, we propose the Large Language Model-based Universal Sizing Optimizer (LLM-USO), which introduces a novel method for knowledge representation to encode circuit design knowledge in a structured text format. This representation enables the systematic reuse of optimization insights for circuits with similar sub-structures. LLM-USO employs a hybrid framework that integrates BO with large language models (LLMs) and a learning summary module. This approach serves to: (i) infuse domain-specific knowledge into the BO process and (ii) facilitate knowledge transfer across circuits, mirroring the cognitive strategies of expert designers. Specifically, LLM-USO constructs a knowledge summary mechanism to distill and apply design insights from one circuit to related ones. It also incorporates a knowledge summary critiquing mechanism to ensure the accuracy and quality of the summaries and employs BO-guided suggestion filtering to identify optimal design points efficiently.
Long Context Language Models have drawn great attention in the past few years. There has been work discussing the impact of long context on Language Model performance: some find that long irrelevant context could harm performance, while some experimentally summarize loss reduction by relevant long context as Scaling Laws. This calls for a more thorough understanding on how long context impacts Language Modeling. In this work, we (1) propose a clean and effective theoretical framework for explaining the impact of context length on Language Modeling, from an Intrinsic Space perspective; and (2) conduct experiments on natural language and synthetic data, validating our proposed theoretical assumptions and deductions. Our theoretical framework can provide practical insights such as establishing that training dataset size dictates an optimal context length and bounds context length scaling for certain cases. We hope our work may inspire new long context Language Models, as well as future work studying Physics for Language Models. Code for our experiments is available at: https://github.com/JingzheShi/NLPCtlScalingAndBounds.
Large language models have been widely applied, but can inadvertently encode sensitive or harmful information, raising significant safety concerns. Machine unlearning has emerged to alleviate this concern; however, existing training-time unlearning approaches, relying on coarse-grained loss combinations, have limitations in precisely separating knowledge and balancing removal effectiveness with model utility. In contrast, we propose Fine-grained Activation manipuLation by Contrastive Orthogonal uNalignment (FALCON), a novel representation-guided unlearning approach that leverages information-theoretic guidance for efficient parameter selection, employs contrastive mechanisms to enhance representation separation, and projects conflict gradients onto orthogonal subspaces to resolve conflicts between forgetting and retention objectives. Extensive experiments demonstrate that FALCON achieves superior unlearning effectiveness while maintaining model utility, exhibiting robust resistance against knowledge recovery attempts.
Similarity between training examples is used to curate pretraining datasets for language models by many methods -- for diversification and to select examples similar to high-quality data. However, similarity is typically measured with off-the-shelf embedding models that are generic or trained for tasks such as retrieval. This paper introduces a framework to analyze the suitability of embedding models specifically for data curation in the language model pretraining setting. We quantify the correlation between similarity in the embedding space to similarity in pretraining loss between different training examples, and how diversifying in the embedding space affects pretraining quality. We analyze a variety of embedding models in our framework, with experiments using the Pile dataset for pretraining a 1.7B parameter decoder-only language model. We find that the embedding models we consider are all useful for pretraining data curation. Moreover, a simple approach of averaging per-token embeddings proves to be surprisingly competitive with more sophisticated embedding models -- likely because the latter are not designed specifically for pretraining data curation. Indeed, we believe our analysis and evaluation framework can serve as a foundation for the design of embedding models that specifically reason about similarity in pretraining datasets.
Recent advancements in deep learning optimization have introduced new algorithms, such as Schedule-Free optimizers, AdEMAMix, MARS and Lion which modify traditional momentum mechanisms. In a separate line of work, theoretical acceleration of stochastic gradient descent (SGD) in noise-dominated regime has been achieved by decoupling the momentum coefficient from the current gradient's weight. In this paper, we establish explicit connections between these two lines of work. We substantiate our theoretical findings with preliminary experiments on a 150m language modeling task. We find that AdEMAMix, which most closely resembles accelerated versions of stochastic gradient descent, exhibits superior performance. Building on these insights, we introduce a modification to AdEMAMix, termed Simplified-AdEMAMix, which maintains the same performance as AdEMAMix across both large and small batch-size settings while eliminating the need for two different momentum terms. The code for Simplified-AdEMAMix is available on the repository: https://github.com/DepenM/Simplified-AdEMAMix/.
The context caching technique is employed to accelerate the Multimodal Large Language Model (MLLM) inference by prevailing serving platforms currently. However, this approach merely reuses the Key-Value (KV) cache of the initial sequence of prompt, resulting in full KV cache recomputation even if the prefix differs slightly. This becomes particularly inefficient in the context of interleaved text and images, as well as multimodal retrieval-augmented generation. This paper proposes position-independent caching as a more effective approach for multimodal information management. We have designed and implemented a caching system, named MPIC, to address both system-level and algorithm-level challenges. MPIC stores the KV cache on local or remote disks when receiving multimodal data, and calculates and loads the KV cache in parallel during inference. To mitigate accuracy degradation, we have incorporated integrated reuse and recompute mechanisms within the system. The experimental results demonstrate that MPIC can achieve up to 54% reduction in response time compared to existing context caching systems, while maintaining negligible or no accuracy loss.
Recent advancements in Large Language Model (LLM) compression, such as BitNet and BitNet b1.58, have marked significant strides in reducing the computational demands of LLMs through innovative one-bit quantization techniques. We extend this frontier by looking at Large Inference Models (LIMs) that have become indispensable across various applications. However, their scale and complexity often come at a significant computational cost. We introduce a novel approach that leverages one-bit algorithm unrolling, effectively integrating information from the physical world in the model architecture. Our method achieves a bit-per-link rate significantly lower than the 1.58 bits reported in prior work, thanks to the natural sparsity that emerges in our network architectures. We numerically demonstrate that the proposed one-bit algorithm unrolling scheme can improve both training and test outcomes by effortlessly increasing the number of layers while substantially compressing the network. Additionally, we provide theoretical results on the generalization gap, convergence rate, stability, and sensitivity of our proposed one-bit algorithm unrolling.
Machine learning models are routinely trained on a mixture of different data domains. Different domain weights yield very different downstream performances. We propose the Soup-of-Experts, a novel architecture that can instantiate a model at test time for any domain weights with minimal computational cost and without re-training the model. Our architecture consists of a bank of expert parameters, which are linearly combined to instantiate one model. We learn the linear combination coefficients as a function of the input domain weights. To train this architecture, we sample random domain weights, instantiate the corresponding model, and backprop through one batch of data sampled with these domain weights. We demonstrate how our approach obtains small specialized models on several language modeling tasks quickly. Soup-of-Experts are particularly appealing when one needs to ship many different specialist models quickly under a model size constraint.
Fine-tuning large language models (LLMs) is often constrained by the computational costs of processing massive datasets. We propose \textbf{QLESS} (Quantized Low-rank Gradient Similarity Search), which integrates gradient quantization with the LESS framework to enable memory-efficient data valuation and selection. QLESS employs a two-step compression process: first, it obtains low-dimensional gradient representations through LoRA-based random projection; then, it quantizes these gradients to low-bitwidth representations. Experiments on multiple LLM architectures (LLaMA, Mistral, Qwen) and benchmarks (MMLU, BBH, TyDiQA) show that QLESS achieves comparable data selection performance to LESS while reducing memory usage by up to 16x. Even 1-bit gradient quantization preserves data valuation quality. These findings underscore QLESS as a practical, scalable approach to identifying informative examples within strict memory constraints.
This paper presents a comparative study aimed at optimizing Llama2 inference, a critical aspect of machine learning and natural language processing (NLP). We evaluate various programming languages and frameworks, including TensorFlow, PyTorch, Python, Mojo, C++, and Java, analyzing their performance in terms of speed, memory consumption, and ease of implementation through extensive benchmarking. Strengths and limitations of each approach are highlighted, along with proposed optimization strategies for parallel processing and hardware utilization. Furthermore, we investigate the Mojo SDK, a novel framework designed for large language model (LLM) inference on Apple Silicon, benchmarking its performance against implementations in C, C++, Rust, Zig, Go, and Julia. Our experiments, conducted on an Apple M1 Max, demonstrate Mojo SDK's competitive performance, ease of use, and seamless Python compatibility, positioning it as a strong alternative for LLM inference on Apple Silicon. We also discuss broader implications for LLM deployment on resource-constrained hardware and identify potential directions for future research.
Large Language Models (LLM)-based systems, i.e. interconnected elements that include an LLM as a central component (e.g., conversational agents), are typically monolithic static architectures that rely on a single LLM for all user queries. However, they often require different preprocessing strategies, levels of reasoning, or knowledge. Generalist LLMs (e.g. GPT-4) trained on very large multi-topic corpora can perform well in a variety of tasks. They require significant financial, energy, and hardware resources that may not be justified for basic tasks. This implies potentially investing in unnecessary costs for a given query. To overcome this problem, a routing mechanism routes user queries to the most suitable components, such as smaller LLMs or experts in specific topics. This approach may improve response quality while minimising costs. Routing can be expanded to other components of the conversational agent architecture, such as the selection of optimal embedding strategies. This paper explores key considerations for integrating routing into LLM-based systems, focusing on resource management, cost definition, and strategy selection. Our main contributions include a formalisation of the problem, a novel taxonomy of existing approaches emphasising relevance and resource efficiency, and a comparative analysis of these strategies in relation to industry practices. Finally, we identify critical challenges and directions for future research.
We introduce AutoGraph, a novel autoregressive framework for generating large attributed graphs using decoder-only transformers. At the core of our approach is a reversible "flattening" process that transforms graphs into random sequences. By sampling and learning from these sequences, AutoGraph enables transformers to model and generate complex graph structures in a manner akin to natural language. In contrast to diffusion models that rely on computationally intensive node features, our approach operates exclusively on these sequences. The sampling complexity and sequence length scale linearly with the number of edges, making AutoGraph highly scalable for generating large sparse graphs. Empirically, AutoGraph achieves state-of-the-art performance across diverse synthetic and molecular graph generation benchmarks, while delivering a 100-fold generation and a 3-fold training speedup compared to leading diffusion models. Additionally, it demonstrates promising transfer capabilities and supports substructure-conditioned generation without additional fine-tuning. By extending language modeling techniques to graph generation, this work paves the way for developing graph foundation models.
Generating semantically coherent text requires a robust internal representation of linguistic structures, which traditional embedding techniques often fail to capture adequately. A novel approach, Latent Lexical Projection (LLP), is introduced to refine lexical representations through a structured transformation into a latent space, thereby enhancing the alignment between input embeddings and their contextual meanings. The method integrates an optimized projection mechanism within an existing language model architecture, enabling more accurate token selection while maintaining syntactic integrity. Evaluations across multiple benchmarks indicate a reduction in perplexity and an increase in BLEU scores, suggesting improvements in predictive accuracy and fluency. The analysis of lexical diversity reveals a more varied vocabulary in generated text, addressing common issues of redundancy and repetitive phrase structures. Further assessments of entropy distributions demonstrate a decline in uncertainty during decoding, reflecting enhanced confidence in word selection. Additionally, long-range dependency retention exhibits measurable gains, with increased classification accuracy at extended token distances. Computational efficiency remains within manageable constraints, despite the added projection mechanism, highlighting the practicality of LLP for integration into existing architectures.