Train Small, Infer Large: Memory-Efficient LoRA Training for Large Language Models

Large Language Models (LLMs) have significantly advanced natural language processing with exceptional task generalization capabilities. Low-Rank Adaption (LoRA) offers a cost-effective fine-tuning solution, freezing the original model parameters and training only lightweight, low-rank adapter matrices. However, the memory footprint of LoRA is largely dominated by the original model parameters. To mitigate this, we propose LoRAM, a memory-efficient LoRA training scheme founded on the intuition that many neurons in over-parameterized LLMs have low training utility but are essential for inference. LoRAM presents a unique twist: it trains on a pruned (small) model to obtain pruned low-rank matrices, which are then recovered and utilized with the original (large) model for inference. Additionally, minimal-cost continual pre-training, performed by the model publishers in advance, aligns the knowledge discrepancy between pruned and original models. Our extensive experiments demonstrate the efficacy of LoRAM across various pruning strategies and downstream tasks. For a model with 70 billion parameters, LoRAM enables training on a GPU with only 20G HBM, replacing an A100-80G GPU for LoRA training and 15 GPUs for full fine-tuning. Specifically, QLoRAM implemented by structured pruning combined with 4-bit quantization, for LLaMA-3.1-70B (LLaMA-2-70B), reduces the parameter storage cost that dominates the memory usage in low-rank matrix training by 15.81$\times$ (16.95$\times$), while achieving dominant performance gains over both the original LLaMA-3.1-70B (LLaMA-2-70B) and LoRA-trained LLaMA-3.1-8B (LLaMA-2-13B).

Preventing the Popular Item Embedding Based Attack in Federated Recommendations

Privacy concerns have led to the rise of federated recommender systems (FRS), which can create personalized models across distributed clients. However, FRS is vulnerable to poisoning attacks, where malicious users manipulate gradients to promote their target items intentionally. Existing attacks against FRS have limitations, as they depend on specific models and prior knowledge, restricting their real-world applicability. In our exploration of practical FRS vulnerabilities, we devise a model-agnostic and prior-knowledge-free attack, named PIECK (Popular Item Embedding based Attack). The core module of PIECK is popular item mining, which leverages embedding changes during FRS training to effectively identify the popular items. Built upon the core module, PIECK branches into two diverse solutions: The PIECKIPE solution employs an item popularity enhancement module, which aligns the embeddings of targeted items with the mined popular items to increase item exposure. The PIECKUEA further enhances the robustness of the attack by using a user embedding approximation module, which approximates private user embeddings using mined popular items. Upon identifying PIECK, we evaluate existing federated defense methods and find them ineffective against PIECK, as poisonous gradients inevitably overwhelm the cold target items. We then propose a novel defense method by introducing two regularization terms during user training, which constrain item popularity enhancement and user embedding approximation while preserving FRS performance. We evaluate PIECK and its defense across two base models, three real datasets, four top-tier attacks, and six general defense methods, affirming the efficacy of both PIECK and its defense.

Cognitive-Aligned Document Selection for Retrieval-augmented Generation

Large language models (LLMs) inherently display hallucinations since the precision of generated texts cannot be guaranteed purely by the parametric knowledge they include. Although retrieval-augmented generation (RAG) systems enhance the accuracy and reliability of generative models by incorporating external documents, these retrieved documents often fail to adequately support the model's responses in practical applications. To address this issue, we propose GGatrieval (Fine-\textbf{G}rained \textbf{G}rounded \textbf{A}lignment Re\textbf{trieval} for verifiable generation), which leverages an LLM to dynamically update queries and filter high-quality, reliable retrieval documents. Specifically, we parse the user query into its syntactic components and perform fine-grained grounded alignment with the retrieved documents. For query components that cannot be individually aligned, we propose a dynamic semantic compensation mechanism that iteratively refines and rewrites the query while continuously updating the retrieval results. This iterative process continues until the retrieved documents sufficiently support the query's response. Our approach introduces a novel criterion for filtering retrieved documents, closely emulating human strategies for acquiring targeted information. This ensures that the retrieved content effectively supports and verifies the generated outputs. On the ALCE benchmark, our method significantly surpasses a wide range of baselines, achieving state-of-the-art performance.

Enhancing Auto-regressive Chain-of-Thought through Loop-Aligned Reasoning

Chain-of-Thought (CoT) prompting has emerged as a powerful technique for enhancing language model's reasoning capabilities. However, generating long and correct CoT trajectories is challenging. Recent studies have demonstrated that Looped Transformers possess remarkable length generalization capabilities, but their limited generality and adaptability prevent them from serving as an alternative to auto-regressive solutions. To better leverage the strengths of Looped Transformers, we propose RELAY (REasoning through Loop Alignment iterativelY). Specifically, we align the steps of Chain-of-Thought (CoT) reasoning with loop iterations and apply intermediate supervision during the training of Looped Transformers. This additional iteration-wise supervision not only preserves the Looped Transformer's ability for length generalization but also enables it to predict CoT reasoning steps for unseen data. Therefore, we leverage this Looped Transformer to generate accurate reasoning chains for complex problems that exceed the training length, which will then be used to fine-tune an auto-regressive model. We conduct extensive experiments, and the results demonstrate the effectiveness of our approach, with significant improvements in the performance of the auto-regressive model. Code will be released at https://github.com/qifanyu/RELAY.

AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

Understanding human behavior and society is a central focus in social sciences, with the rise of generative social science marking a significant paradigmatic shift. By leveraging bottom-up simulations, it replaces costly and logistically challenging traditional experiments with scalable, replicable, and systematic computational approaches for studying complex social dynamics. Recent advances in large language models (LLMs) have further transformed this research paradigm, enabling the creation of human-like generative social agents and realistic simulacra of society. In this paper, we propose AgentSociety, a large-scale social simulator that integrates LLM-driven agents, a realistic societal environment, and a powerful large-scale simulation engine. Based on the proposed simulator, we generate social lives for over 10k agents, simulating their 5 million interactions both among agents and between agents and their environment. Furthermore, we explore the potential of AgentSociety as a testbed for computational social experiments, focusing on four key social issues: polarization, the spread of inflammatory messages, the effects of universal basic income policies, and the impact of external shocks such as hurricanes. These four issues serve as valuable cases for assessing AgentSociety's support for typical research methods -- such as surveys, interviews, and interventions -- as well as for investigating the patterns, causes, and underlying mechanisms of social issues. The alignment between AgentSociety's outcomes and real-world experimental results not only demonstrates its ability to capture human behaviors and their underlying mechanisms, but also underscores its potential as an important platform for social scientists and policymakers.

Bringing RGB and IR Together: Hierarchical Multi-Modal Enhancement for Robust Transmission Line Detection

Ensuring a stable power supply in rural areas relies heavily on effective inspection of power equipment, particularly transmission lines (TLs). However, detecting TLs from aerial imagery can be challenging when dealing with misalignments between visible light (RGB) and infrared (IR) images, as well as mismatched high- and low-level features in convolutional networks. To address these limitations, we propose a novel Hierarchical Multi-Modal Enhancement Network (HMMEN) that integrates RGB and IR data for robust and accurate TL detection. Our method introduces two key components: (1) a Mutual Multi-Modal Enhanced Block (MMEB), which fuses and enhances hierarchical RGB and IR feature maps in a coarse-to-fine manner, and (2) a Feature Alignment Block (FAB) that corrects misalignments between decoder outputs and IR feature maps by leveraging deformable convolutions. We employ MobileNet-based encoders for both RGB and IR inputs to accommodate edge-computing constraints and reduce computational overhead. Experimental results on diverse weather and lighting conditionsfog, night, snow, and daytimedemonstrate the superiority and robustness of our approach compared to state-of-the-art methods, resulting in fewer false positives, enhanced boundary delineation, and better overall detection performance. This framework thus shows promise for practical large-scale power line inspections with unmanned aerial vehicles.

Joint Automatic Speech Recognition And Structure Learning For Better Speech Understanding

Spoken language understanding (SLU) is a structure prediction task in the field of speech. Recently, many works on SLU that treat it as a sequence-to-sequence task have achieved great success. However, This method is not suitable for simultaneous speech recognition and understanding. In this paper, we propose a joint speech recognition and structure learning framework (JSRSL), an end-to-end SLU model based on span, which can accurately transcribe speech and extract structured content simultaneously. We conduct experiments on name entity recognition and intent classification using the Chinese dataset AISHELL-NER and the English dataset SLURP. The results show that our proposed method not only outperforms the traditional sequence-to-sequence method in both transcription and extraction capabilities but also achieves state-of-the-art performance on the two datasets.

Dual Mutual Learning Network with Global-local Awareness for RGB-D Salient Object Detection

RGB-D salient object detection (SOD), aiming to highlight prominent regions of a given scene by jointly modeling RGB and depth information, is one of the challenging pixel-level prediction tasks. Recently, the dual-attention mechanism has been devoted to this area due to its ability to strengthen the detection process. However, most existing methods directly fuse attentional cross-modality features under a manual-mandatory fusion paradigm without considering the inherent discrepancy between the RGB and depth, which may lead to a reduction in performance. Moreover, the long-range dependencies derived from global and local information make it difficult to leverage a unified efficient fusion strategy. Hence, in this paper, we propose the GL-DMNet, a novel dual mutual learning network with global-local awareness. Specifically, we present a position mutual fusion module and a channel mutual fusion module to exploit the interdependencies among different modalities in spatial and channel dimensions. Besides, we adopt an efficient decoder based on cascade transformer-infused reconstruction to integrate multi-level fusion features jointly. Extensive experiments on six benchmark datasets demonstrate that our proposed GL-DMNet performs better than 24 RGB-D SOD methods, achieving an average improvement of ~3% across four evaluation metrics compared to the second-best model (S3Net). Codes and results are available at https://github.com/kingkung2016/GL-DMNet.

Deep Learning Models for Colloidal Nanocrystal Synthesis

Colloidal synthesis of nanocrystals usually includes complex chemical reactions and multi-step crystallization processes. Despite the great success in the past 30 years, it remains challenging to clarify the correlations between synthetic parameters of chemical reaction and physical properties of nanocrystals. Here, we developed a deep learning-based nanocrystal synthesis model that correlates synthetic parameters with the final size and shape of target nanocrystals, using a dataset of 3500 recipes covering 348 distinct nanocrystal compositions. The size and shape labels were obtained from transmission electron microscope images using a segmentation model trained with a semi-supervised algorithm on a dataset comprising 1.2 million nanocrystals. By applying the reaction intermediate-based data augmentation method and elaborated descriptors, the synthesis model was able to predict nanocrystal's size with a mean absolute error of 1.39 nm, while reaching an 89% average accuracy for shape classification. The synthesis model shows knowledge transfer capabilities across different nanocrystals with inputs of new recipes. With that, the influence of chemicals on the final size of nanocrystals was further evaluated, revealing the importance order of nanocrystal composition, precursor or ligand, and solvent. Overall, the deep learning-based nanocrystal synthesis model offers a powerful tool to expedite the development of high-quality nanocrystals.

AI and Deep Learning for THz Ultra-Massive MIMO: From Model-Driven Approaches to Foundation Models

In this paper, we explore the potential of artificial intelligence (AI) to address the challenges posed by terahertz ultra-massive multiple-input multiple-output (THz UM-MIMO) systems. We begin by outlining the characteristics of THz UM-MIMO systems, and identify three primary challenges for the transceiver design: 'hard to compute', 'hard to model', and 'hard to measure'. We argue that AI can provide a promising solution to these challenges. We then propose two systematic research roadmaps for developing AI algorithms tailored for THz UM-MIMO systems. The first roadmap, called model-driven deep learning (DL), emphasizes the importance to leverage available domain knowledge and advocates for adopting AI only to enhance the bottleneck modules within an established signal processing or optimization framework. We discuss four essential steps to make it work, including algorithmic frameworks, basis algorithms, loss function design, and neural architecture design. Afterwards, we present a forward-looking vision through the second roadmap, i.e., physical layer foundation models. This approach seeks to unify the design of different transceiver modules by focusing on their common foundation, i.e., the wireless channel. We propose to train a single, compact foundation model to estimate the score function of wireless channels, which can serve as a versatile prior for designing a wide variety of transceiver modules. We will also guide the readers through four essential steps, including general frameworks, conditioning, site-specific adaptation, and the joint design of foundation models and model-driven DL.