Large Language Models (LLMs) with long context capabilities are integral to complex tasks in natural language processing and computational biology, such as text generation and protein sequence analysis. However, training LLMs directly on extremely long contexts demands considerable GPU resources and increased memory, leading to higher costs and greater complexity. Alternative approaches that introduce long context capabilities via downstream finetuning or adaptations impose significant design limitations. In this paper, we propose Fully Pipelined Distributed Transformer (FPDT) for efficiently training long-context LLMs with extreme hardware efficiency. For GPT and Llama models, we achieve a 16x increase in sequence length that can be trained on the same hardware compared to current state-of-the-art solutions. With our dedicated sequence chunk pipeline design, we can now train 8B LLM with 2 million sequence length on only 4 GPUs, while also maintaining over 55% of MFU. Our proposed FPDT is agnostic to existing training techniques and is proven to work efficiently across different LLM models.
Recent advancements in Large Language Models (LLMs) have yielded remarkable success across diverse fields. However, handling long contexts remains a significant challenge for LLMs due to the quadratic time and space complexity of attention mechanisms and the growing memory consumption of the key-value cache during generation. This work introduces MemLong: Memory-Augmented Retrieval for Long Text Generation, a method designed to enhance the capabilities of long-context language modeling by utilizing an external retriever for historical information retrieval. MemLong combines a non-differentiable ``ret-mem'' module with a partially trainable decoder-only language model and introduces a fine-grained, controllable retrieval attention mechanism that leverages semantic-level relevant chunks. Comprehensive evaluations on multiple long-context language modeling benchmarks demonstrate that MemLong consistently outperforms other state-of-the-art LLMs. More importantly, MemLong can extend the context length on a single 3090 GPU from 4k up to 80k. Our code is available at https://github.com/Bui1dMySea/MemLong
Recent progress in Text-to-Image (T2I) generative models has enabled high-quality image generation. As performance and accessibility increase, these models are gaining significant attraction and popularity: ensuring their fairness and safety is a priority to prevent the dissemination and perpetuation of biases. However, existing studies in bias detection focus on closed sets of predefined biases (e.g., gender, ethnicity). In this paper, we propose a general framework to identify, quantify, and explain biases in an open set setting, i.e. without requiring a predefined set. This pipeline leverages a Large Language Model (LLM) to propose biases starting from a set of captions. Next, these captions are used by the target generative model for generating a set of images. Finally, Vision Question Answering (VQA) is leveraged for bias evaluation. We show two variations of this framework: OpenBias and GradBias. OpenBias detects and quantifies biases, while GradBias determines the contribution of individual prompt words on biases. OpenBias effectively detects both well-known and novel biases related to people, objects, and animals and highly aligns with existing closed-set bias detection methods and human judgment. GradBias shows that neutral words can significantly influence biases and it outperforms several baselines, including state-of-the-art foundation models. Code available here: https://github.com/Moreno98/GradBias.
We present the "Law of Vision Representation" in multimodal large language models (MLLMs). It reveals a strong correlation between the combination of cross-modal alignment, correspondence in vision representation, and MLLM performance. We quantify the two factors using the cross-modal Alignment and Correspondence score (AC score). Through extensive experiments involving thirteen different vision representation settings and evaluations across eight benchmarks, we find that the AC score is linearly correlated to model performance. By leveraging this relationship, we are able to identify and train the optimal vision representation only, which does not require finetuning the language model every time, resulting in a 99.7% reduction in computational cost.
Large language models (LLMs) are increasingly explored as general-purpose reasoners, particularly in agentic contexts. However, their outputs remain prone to mathematical and logical errors. This is especially challenging in open-ended tasks, where unstructured outputs lack explicit ground truth and may contain subtle inconsistencies. To address this issue, we propose Logic-Enhanced Language Model Agents (LELMA), a framework that integrates LLMs with formal logic to enable validation and refinement of natural language reasoning. LELMA comprises three components: an LLM-Reasoner, an LLM-Translator, and a Solver, and employs autoformalization to translate reasoning into logic representations, which are then used to assess logical validity. Using game-theoretic scenarios such as the Prisoner's Dilemma as testbeds, we highlight the limitations of both less capable (Gemini 1.0 Pro) and advanced (GPT-4o) models in generating logically sound reasoning. LELMA achieves high accuracy in error detection and improves reasoning correctness via self-refinement, particularly in GPT-4o. The study also highlights challenges in autoformalization accuracy and in evaluation of inherently ambiguous open-ended reasoning tasks.
In the rapidly evolving field of cybersecurity, the integration of flow-level and packet-level information for real-time intrusion detection remains a largely untapped area of research. This paper introduces "XG-NID," a novel framework that, to the best of our knowledge, is the first to fuse flow-level and packet-level data within a heterogeneous graph structure, offering a comprehensive analysis of network traffic. Leveraging a heterogeneous graph neural network (GNN) with graph-level classification, XG-NID uniquely enables real-time inference while effectively capturing the intricate relationships between flow and packet payload data. Unlike traditional GNN-based methodologies that predominantly analyze historical data, XG-NID is designed to accommodate the heterogeneous nature of network traffic, providing a robust and real-time defense mechanism. Our framework extends beyond mere classification; it integrates Large Language Models (LLMs) to generate detailed, human-readable explanations and suggest potential remedial actions, ensuring that the insights produced are both actionable and comprehensible. Additionally, we introduce a new set of flow features based on temporal information, further enhancing the contextual and explainable inferences provided by our model. To facilitate practical application and accessibility, we developed "GNN4ID," an open-source tool that enables the extraction and transformation of raw network traffic into the proposed heterogeneous graph structure, seamlessly integrating flow and packet-level data. Our comprehensive quantitative comparative analysis demonstrates that XG-NID achieves an F1 score of 97\% in multi-class classification, outperforming existing baseline and state-of-the-art methods. This sets a new standard in Network Intrusion Detection Systems by combining innovative data fusion with enhanced interpretability and real-time capabilities.
Language models have been effectively applied to modeling natural signals, such as images, video, speech, and audio. A crucial component of these models is the codec tokenizer, which compresses high-dimensional natural signals into lower-dimensional discrete tokens. In this paper, we introduce WavTokenizer, which offers several advantages over previous SOTA acoustic codec models in the audio domain: 1)extreme compression. By compressing the layers of quantizers and the temporal dimension of the discrete codec, one-second audio of 24kHz sampling rate requires only a single quantizer with 40 or 75 tokens. 2)improved subjective quality. Despite the reduced number of tokens, WavTokenizer achieves state-of-the-art reconstruction quality with outstanding UTMOS scores and inherently contains richer semantic information. Specifically, we achieve these results by designing a broader VQ space, extended contextual windows, and improved attention networks, as well as introducing a powerful multi-scale discriminator and an inverse Fourier transform structure. We conducted extensive reconstruction experiments in the domains of speech, audio, and music. WavTokenizer exhibited strong performance across various objective and subjective metrics compared to state-of-the-art models. We also tested semantic information, VQ utilization, and adaptability to generative models. Comprehensive ablation studies confirm the necessity of each module in WavTokenizer. The related code, demos, and pre-trained models are available at https://github.com/jishengpeng/WavTokenizer.
Given the societal impact of unsafe content generated by large language models (LLMs), ensuring that LLM services comply with safety standards is a crucial concern for LLM service providers. Common content moderation methods are limited by an effectiveness-and-efficiency dilemma, where simple models are fragile while sophisticated models consume excessive computational resources. In this paper, we reveal for the first time that effective and efficient content moderation can be achieved by extracting conceptual features from chat-oriented LLMs, despite their initial fine-tuning for conversation rather than content moderation. We propose a practical and unified content moderation framework for LLM services, named Legilimens, which features both effectiveness and efficiency. Our red-team model-based data augmentation enhances the robustness of Legilimens against state-of-the-art jailbreaking. Additionally, we develop a framework to theoretically analyze the cost-effectiveness of Legilimens compared to other methods. We have conducted extensive experiments on five host LLMs, seventeen datasets, and nine jailbreaking methods to verify the effectiveness, efficiency, and robustness of Legilimens against normal and adaptive adversaries. A comparison of Legilimens with both commercial and academic baselines demonstrates the superior performance of Legilimens. Furthermore, we confirm that Legilimens can be applied to few-shot scenarios and extended to multi-label classification tasks.
In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git
Understanding the learning process and the embedded computation in transformers is becoming a central goal for the development of interpretable AI. In the present study, we introduce a hierarchical filtering procedure for data models of sequences on trees, allowing us to hand-tune the range of positional correlations in the data. Leveraging this controlled setting, we provide evidence that vanilla encoder-only transformers can approximate the exact inference algorithm when trained on root classification and masked language modeling tasks, and study how this computation is discovered and implemented. We find that correlations at larger distances, corresponding to increasing layers of the hierarchy, are sequentially included by the network during training. By comparing attention maps from models trained with varying degrees of filtering and by probing the different encoder levels, we find clear evidence of a reconstruction of correlations on successive length scales corresponding to the various levels of the hierarchy, which we relate to a plausible implementation of the exact inference algorithm within the same architecture.
Known as low energy consumption networks, spiking neural networks (SNNs) have gained a lot of attention within the past decades. While SNNs are increasing competitive with artificial neural networks (ANNs) for vision tasks, they are rarely used for long sequence tasks, despite their intrinsic temporal dynamics. In this work, we develop spiking state space models (SpikingSSMs) for long sequence learning by leveraging on the sequence learning abilities of state space models (SSMs). Inspired by dendritic neuron structure, we hierarchically integrate neuronal dynamics with the original SSM block, meanwhile realizing sparse synaptic computation. Furthermore, to solve the conflict of event-driven neuronal dynamics with parallel computing, we propose a light-weight surrogate dynamic network which accurately predicts the after-reset membrane potential and compatible to learnable thresholds, enabling orders of acceleration in training speed compared with conventional iterative methods. On the long range arena benchmark task, SpikingSSM achieves competitive performance to state-of-the-art SSMs meanwhile realizing on average 90\% of network sparsity. On language modeling, our network significantly surpasses existing spiking large language models (spikingLLMs) on the WikiText-103 dataset with only a third of the model size, demonstrating its potential as backbone architecture for low computation cost LLMs.
We introduce SHADOW, a fine-tuned language model trained on an intermediate task using associative deductive reasoning, and measure its performance on a knowledge base construction task using Wikidata triple completion. We evaluate SHADOW on the LM-KBC 2024 challenge and show that it outperforms the baseline solution by 20% with a F1 score of 68.72%.
Abundant, well-annotated multimodal data in remote sensing are pivotal for aligning complex visual remote sensing (RS) scenes with human language, enabling the development of specialized vision language models across diverse RS interpretation tasks. However, annotating RS images with rich linguistic semantics at scale demands expertise in RS and substantial human labor, making it costly and often impractical. In this study, we propose a workflow that leverages large language models (LLMs) to generate multimodal datasets with semantically rich captions at scale from plain OpenStreetMap (OSM) data for images sourced from the Google Earth Engine (GEE) platform. This approach facilitates the generation of paired remote sensing data and can be readily scaled up using openly available data. Within this framework, we present RSTeller, a multimodal dataset comprising over 1.3 million RS images, each accompanied by two descriptive captions. Extensive experiments demonstrate that RSTeller enhances the performance of multiple existing vision language models for RS scene understanding through continual pre-training. Our methodology significantly reduces the manual effort and expertise needed for annotating remote sensing imagery while democratizing access to high-quality annotated data. This advancement fosters progress in visual language modeling and encourages broader participation in remote sensing research and applications. The RSTeller dataset is available at https://github.com/SlytherinGe/RSTeller.
Zero-shot graph machine learning, especially with graph neural networks (GNNs), has garnered significant interest due to the challenge of scarce labeled data. While methods like self-supervised learning and graph prompt learning have been extensively explored, they often rely on fine-tuning with task-specific labels, limiting their effectiveness in zero-shot scenarios. Inspired by the zero-shot capabilities of instruction-fine-tuned large language models (LLMs), we introduce a novel framework named Token Embedding-Aligned Graph Language Model (TEA-GLM) that leverages LLMs as cross-dataset and cross-task zero-shot learners for graph machine learning. Concretely, we pretrain a GNN, aligning its representations with token embeddings of an LLM. We then train a linear projector that transforms the GNN's representations into a fixed number of graph token embeddings without tuning the LLM. A unified instruction is designed for various graph tasks at different levels, such as node classification (node-level) and link prediction (edge-level). These design choices collectively enhance our method's effectiveness in zero-shot learning, setting it apart from existing methods. Experiments show that our graph token embeddings help the LLM predictor achieve state-of-the-art performance on unseen datasets and tasks compared to other methods using LLMs as predictors.
Linear RNN architectures, like Mamba, can be competitive with Transformer models in language modeling while having advantageous deployment characteristics. Given the focus on training large-scale Transformer models, we consider the challenge of converting these pretrained models for deployment. We demonstrate that it is feasible to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers with academic GPU resources. The resulting hybrid model, which incorporates a quarter of the attention layers, achieves performance comparable to the original Transformer in chat benchmarks and outperforms open-source hybrid Mamba models trained from scratch with trillions of tokens in both chat benchmarks and general benchmarks. Moreover, we introduce a hardware-aware speculative decoding algorithm that accelerates the inference speed of Mamba and hybrid models. Overall we show how, with limited computation resources, we can remove many of the original attention layers and generate from the resulting model more efficiently. Our top-performing model, distilled from Llama3-8B-Instruct, achieves a 29.61 length-controlled win rate on AlpacaEval 2 against GPT-4 and 7.35 on MT-Bench, surpassing the best 8B scale instruction-tuned linear RNN model. We also find that the distilled model has natural length extrapolation, showing almost perfect accuracy in the needle-in-a-haystack test at 20x the distillation length. Code and pre-trained checkpoints are open-sourced at https://github.com/jxiw/MambaInLlama and https://github.com/itsdaniele/speculative_mamba.
The increasing difficulty to distinguish language-model-generated from human-written text has led to the development of detectors of machine-generated text (MGT). However, in many contexts, a black-box prediction is not sufficient, it is equally important to know on what grounds a detector made that prediction. Explanation methods that estimate feature importance promise to provide indications of which parts of an input are used by classifiers for prediction. However, the quality of different explanation methods has not previously been assessed for detectors of MGT. This study conducts the first systematic evaluation of explanation quality for this task. The dimensions of faithfulness and stability are assessed with five automated experiments, and usefulness is evaluated in a user study. We use a dataset of ChatGPT-generated and human-written documents, and pair predictions of three existing language-model-based detectors with the corresponding SHAP, LIME, and Anchor explanations. We find that SHAP performs best in terms of faithfulness, stability, and in helping users to predict the detector's behavior. In contrast, LIME, perceived as most useful by users, scores the worst in terms of user performance at predicting the detectors' behavior.
This paper constructs question answering system for bridge design specification based on large language model. Three implementation schemes are tried: full fine-tuning of the Bert pretrained model, parameter-efficient fine-tuning of the Bert pretrained model, and self-built language model from scratch. Through the self-built question and answer task dataset, based on the tensorflow and keras deep learning platform framework, the model is constructed and trained to predict the start position and end position of the answer in the bridge design specification given by the user. The experimental results show that full fine-tuning of the Bert pretrained model achieves 100% accuracy in the training-dataset, validation-dataset and test-dataset, and the system can extract the answers from the bridge design specification given by the user to answer various questions of the user; While parameter-efficient fine-tuning of the Bert pretrained model and self-built language model from scratch perform well in the training-dataset, their generalization ability in the test-dataset needs to be improved. The research of this paper provides a useful reference for the development of question answering system in professional field.
Machine learning research, crucial for technological advancements and innovation, often faces significant challenges due to its inherent complexity, slow pace of experimentation, and the necessity for specialized expertise. Motivated by this, we present a new systematic framework, autonomous Machine Learning Research with large language models (MLR-Copilot), designed to enhance machine learning research productivity through the automatic generation and implementation of research ideas using Large Language Model (LLM) agents. The framework consists of three phases: research idea generation, experiment implementation, and implementation execution. First, existing research papers are used to generate hypotheses and experimental plans vis IdeaAgent powered by LLMs. Next, the implementation generation phase translates these plans into executables with ExperimentAgent. This phase leverages retrieved prototype code and optionally retrieves candidate models and data. Finally, the execution phase, also managed by ExperimentAgent, involves running experiments with mechanisms for human feedback and iterative debugging to enhance the likelihood of achieving executable research outcomes. We evaluate our framework on five machine learning research tasks and the experimental results show the framework's potential to facilitate the research progress and innovations.
Human mobility prediction plays a crucial role in various real-world applications. Although deep learning based models have shown promising results over the past decade, their reliance on extensive private mobility data for training and their inability to perform zero-shot predictions, have hindered further advancements. Recently, attempts have been made to apply large language models (LLMs) to mobility prediction task. However, their performance has been constrained by the absence of a systematic design of workflow. They directly generate the final output using LLMs, which limits the potential of LLMs to uncover complex mobility patterns and underestimates their extensive reserve of global geospatial knowledge. In this paper, we introduce AgentMove, a systematic agentic prediction framework to achieve generalized mobility prediction for any cities worldwide. In AgentMove, we first decompose the mobility prediction task into three sub-tasks and then design corresponding modules to complete these subtasks, including spatial-temporal memory for individual mobility pattern mining, world knowledge generator for modeling the effects of urban structure and collective knowledge extractor for capturing the shared patterns among population. Finally, we combine the results of three modules and conduct a reasoning step to generate the final predictions. Extensive experiments on mobility data from two sources in 12 cities demonstrate that AgentMove outperforms the best baseline more than 8% in various metrics and it shows robust predictions with various LLMs as base and also less geographical bias across cities. Codes and data can be found in https://github.com/tsinghua-fib-lab/AgentMove.
With the prevalence of large-scale pretrained vision-language models (VLMs), such as CLIP, soft-prompt tuning has become a popular method for adapting these models to various downstream tasks. However, few works delve into the inherent properties of learnable soft-prompt vectors, specifically the impact of their norms to the performance of VLMs. This motivates us to pose an unexplored research question: ``Do we need to normalize the soft prompts in VLMs?'' To fill this research gap, we first uncover a phenomenon, called the \textbf{Low-Norm Effect} by performing extensive corruption experiments, suggesting that reducing the norms of certain learned prompts occasionally enhances the performance of VLMs, while increasing them often degrades it. To harness this effect, we propose a novel method named \textbf{N}ormalizing th\textbf{e} soft-pro\textbf{m}pt v\textbf{e}ctors of vi\textbf{si}on-language model\textbf{s} (\textbf{Nemesis}) to normalize soft-prompt vectors in VLMs. To the best of our knowledge, our work is the first to systematically investigate the role of norms of soft-prompt vector in VLMs, offering valuable insights for future research in soft-prompt tuning. The code is available at \texttt{\href{https://github.com/ShyFoo/Nemesis}{https://github.com/ShyFoo/Nemesis}}.