Mingchuan Yang

Zifan Zheng, Yezhaohui Wang, Yuxin Huang et al.

Since the advent of ChatGPT, Large Language Models (LLMs) have excelled in various tasks but remain as black-box systems. Understanding the reasoning bottlenecks of LLMs has become a critical challenge, as these limitations are deeply tied to their internal architecture. Among these, attention heads have emerged as a focal point for investigating the underlying mechanics of LLMs. In this survey, we aim to demystify the internal reasoning processes of LLMs by systematically exploring the roles and mechanisms of attention heads. We first introduce a novel four-stage framework inspired by the human thought process: Knowledge Recalling, In-Context Identification, Latent Reasoning, and Expression Preparation. Using this framework, we comprehensively review existing research to identify and categorize the functions of specific attention heads. Additionally, we analyze the experimental methodologies used to discover these special heads, dividing them into two categories: Modeling-Free and Modeling-Required methods. We further summarize relevant evaluation methods and benchmarks. Finally, we discuss the limitations of current research and propose several potential future directions.

Yiquan Wu, Bo Tang, Chenyang Xi et al.

Commentary provides readers with a deep understanding of events by presenting diverse arguments and evidence. However, creating commentary is a time-consuming task, even for skilled commentators. Large language models (LLMs) have simplified the process of natural language generation, but their direct application in commentary creation still faces challenges due to unique task requirements. These requirements can be categorized into two levels: 1) fundamental requirements, which include creating well-structured and logically consistent narratives, and 2) advanced requirements, which involve generating quality arguments and providing convincing evidence. In this paper, we introduce Xinyu, an efficient LLM-based system designed to assist commentators in generating Chinese commentaries. To meet the fundamental requirements, we deconstruct the generation process into sequential steps, proposing targeted strategies and supervised fine-tuning (SFT) for each step. To address the advanced requirements, we present an argument ranking model for arguments and establish a comprehensive evidence database that includes up-to-date events and classic books, thereby strengthening the substantiation of the evidence with retrieval augmented generation (RAG) technology. To evaluate the generated commentaries more fairly, corresponding to the two-level requirements, we introduce a comprehensive evaluation metric that considers five distinct perspectives in commentary generation. Our experiments confirm the effectiveness of our proposed system. We also observe a significant increase in the efficiency of commentators in real-world scenarios, with the average time spent on creating a commentary dropping from 4 hours to 20 minutes. Importantly, such an increase in efficiency does not compromise the quality of the commentaries.

Hang Zou, Chenxi Du, Ajian Liu et al.

Iris recognition is widely used in high-security scenarios due to its stability and distinctiveness. However, iris images captured by different devices exhibit certain and device-related consistent differences, which has a greater impact on the classification algorithm for anti-spoofing. The iris of various races would also affect the classification, causing the risk of identity theft. So it is necessary to improve the cross-domain capabilities of the iris anti-spoofing (IAS) methods to enable it more robust in facing different races and devices. However, there is no existing protocol that is comprehensively available. To address this gap, we propose an Iris Anti-Spoofing Cross-Domain-Testing (IAS-CDT) Protocol, which involves 10 datasets, belonging to 7 databases, published by 4 institutions, and collected with 6 different devices. It contains three sub-protocols hierarchically, aimed at evaluating average performance, cross-racial generalization, and cross-device generalization of IAS models. Moreover, to address the cross-device generalization challenge brought by the IAS-CDT Protocol, we employ multiple model parameter sets to learn from the multiple sub-datasets. Specifically, we utilize the Mixture of Experts (MoE) to fit complex data distributions using multiple sub-neural networks. To further enhance the generalization capabilities, we propose a novel method Masked-MoE (MMoE), which randomly masks a portion of tokens for some experts and requires their outputs to be similar to the unmasked experts, which can effectively mitigate the overfitting issue of MoE. For the evaluation, we selected ResNet50, VIT-B/16, CLIP, and FLIP as representative models and benchmarked them under the proposed IAS-CDT Protocol.

Yuan Cao, Zhiqiao Gao, Jie Hu et al.

As deep learning becomes the mainstream in the field of natural language processing, the need for suitable active learning method are becoming unprecedented urgent. Active Learning (AL) methods based on nearest neighbor classifier are proposed and demonstrated superior results. However, existing nearest neighbor classifier are not suitable for classifying mutual exclusive classes because inter-class discrepancy cannot be assured by nearest neighbor classifiers. As a result, informative samples in the margin area can not be discovered and AL performance are damaged. To this end, we propose a novel Nearest neighbor Classifier with Margin penalty for Active Learning(NCMAL). Firstly, mandatory margin penalty are added between classes, therefore both inter-class discrepancy and intra-class compactness are both assured. Secondly, a novel sample selection strategy are proposed to discover informative samples within the margin area. To demonstrate the effectiveness of the methods, we conduct extensive experiments on for datasets with other state-of-the-art methods. The experimental results demonstrate that our method achieves better results with fewer annotated samples than all baseline methods.

Tianshu Yu, Chao Xiang, Mingchuan Yang et al.

Large language models (LLMs) have demonstrated remarkable evaluation and critique capabilities, providing insightful feedback and identifying flaws in various tasks. However, limited research has explored which types of critiques are most effective for improving model responses or how to generate such critiques. To address this gap, we introduce \textbf{R}efinement-oriented \textbf{C}ritique \textbf{O}ptimization (RCO), a novel framework designed to train critic models using refinement signals. RCO uses a feedback loop where critiques, generated by the critic model, guide the actor model in refining its responses. The critique utility (CU) quantifies the effectiveness of these refinements, serving as the reward signal for training the critic model. By focusing on critiques that lead to better refinements, RCO eliminates the need for direct critique preference assessment, ensuring that critiques driving meaningful improvements are rewarded. We evaluate RCO across five tasks, i.e., dialog generation, summarization, question answering, mathematical reasoning, and code generation, and show that it significantly outperforms traditional methods and open-source models in terms of critique quality and refinement outcomes. Our contributions include the introduction of RCO, a novel supervision scheme based on refined response preferences, and comprehensive experimental results that highlight the method's effectiveness in enhancing LLM critique-refinement loops.

Zhiyu Li, Shichao Song, Chenyang Xi et al.

Large Language Models (LLMs) have become an essential infrastructure for Artificial General Intelligence (AGI), yet their lack of well-defined memory management systems hinders the development of long-context reasoning, continual personalization, and knowledge consistency.Existing models mainly rely on static parameters and short-lived contextual states, limiting their ability to track user preferences or update knowledge over extended periods.While Retrieval-Augmented Generation (RAG) introduces external knowledge in plain text, it remains a stateless workaround without lifecycle control or integration with persistent representations.Recent work has modeled the training and inference cost of LLMs from a memory hierarchy perspective, showing that introducing an explicit memory layer between parameter memory and external retrieval can substantially reduce these costs by externalizing specific knowledge. Beyond computational efficiency, LLMs face broader challenges arising from how information is distributed over time and context, requiring systems capable of managing heterogeneous knowledge spanning different temporal scales and sources. To address this challenge, we propose MemOS, a memory operating system that treats memory as a manageable system resource. It unifies the representation, scheduling, and evolution of plaintext, activation-based, and parameter-level memories, enabling cost-efficient storage and retrieval. As the basic unit, a MemCube encapsulates both memory content and metadata such as provenance and versioning. MemCubes can be composed, migrated, and fused over time, enabling flexible transitions between memory types and bridging retrieval with parameter-based learning. MemOS establishes a memory-centric system framework that brings controllability, plasticity, and evolvability to LLMs, laying the foundation for continual learning and personalized modeling.

Yu Zhu, Chuxiong Sun, Wenfei Yang et al.

Reinforcement Learning from Human Feedback (RLHF) is the prevailing approach to ensure Large Language Models (LLMs) align with human values. However, existing RLHF methods require a high computational cost, one main reason being that RLHF assigns both the generation and alignment tasks to the LLM simultaneously. In this paper, we introduce Proxy-RLHF, which decouples the generation and alignment processes of LLMs, achieving alignment with human values at a much lower computational cost. We start with a novel Markov Decision Process (MDP) designed for the alignment process and employ Reinforcement Learning (RL) to train a streamlined proxy model that oversees the token generation of the LLM, without altering the LLM itself. Experiments show that our method achieves a comparable level of alignment with only 1\% of the training parameters of other methods.

Yuanfu Wang, Pengyu Wang, Chenyang Xi et al.

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

Mu Chen, Wenyu Chen, Mingchuan Yang et al.

3D semantic occupancy has rapidly become a research focus in the fields of robotics and autonomous driving environment perception due to its ability to provide more realistic geometric perception and its closer integration with downstream tasks. By performing occupancy prediction of the 3D space in the environment, the ability and robustness of scene understanding can be effectively improved. However, existing occupancy prediction tasks are primarily modeled using voxel or point cloud-based approaches: voxel-based network structures often suffer from the loss of spatial information due to the voxelization process, while point cloud-based methods, although better at retaining spatial location information, face limitations in representing volumetric structural details. To address this issue, we propose a dual-modal prediction method based on 3D Gaussian sets and sparse points, which balances both spatial location and volumetric structural information, achieving higher accuracy in semantic occupancy prediction. Specifically, our method adopts a Transformer-based architecture, taking 3D Gaussian sets, sparse points, and queries as inputs. Through the multi-layer structure of the Transformer, the enhanced queries and 3D Gaussian sets jointly contribute to the semantic occupancy prediction, and an adaptive fusion mechanism integrates the semantic outputs of both modalities to generate the final prediction results. Additionally, to further improve accuracy, we dynamically refine the point cloud at each layer, allowing for more precise location information during occupancy prediction. We conducted experiments on the Occ3DnuScenes dataset, and the experimental results demonstrate superior performance of the proposed method on IoU based metrics.

Mingchuan Yang, Ziyuan Huang

This study explores the optimization of the DRAGON Longformer base model for clinical text classification, specifically targeting the binary classification of medical case descriptions. A dataset of 500 clinical cases containing structured medical observations was used, with 400 cases for training and 100 for validation. Enhancements to the pre-trained joeranbosma/dragon-longformer-base-mixed-domain model included hyperparameter tuning, domain-specific preprocessing, and architectural adjustments. Key modifications involved increasing sequence length from 512 to 1024 tokens, adjusting learning rates from 1e-05 to 5e-06, extending training epochs from 5 to 8, and incorporating specialized medical terminology. The optimized model achieved notable performance gains: accuracy improved from 72.0% to 85.2%, precision from 68.0% to 84.1%, recall from 75.0% to 86.3%, and F1-score from 71.0% to 85.2%. Statistical analysis confirmed the significance of these improvements (p < .001). The model demonstrated enhanced capability in interpreting medical terminology, anatomical measurements, and clinical observations. These findings contribute to domain-specific language model research and offer practical implications for clinical natural language processing applications. The optimized model's strong performance across diverse medical conditions underscores its potential for broad use in healthcare settings.

Mingxu Tao, Jie Hu, Mingchuan Yang et al.

The remarkable performance of Large language models (LLMs) relies heavily on the availability of abundant high-quality training data. However, the high cost of acquiring annotated data often prevents models from obtaining capabilities to tackle downstream tasks. In this paper, we introduce a novel method, EpiCoDe that boosts model performance in data-scarcity scenarios without extra training. We first employ model extrapolation to enhance a finetuned model with its inferior version, and then adopt contrastive decoding to further reduce predicted errors, by comparing the logit scores given by the extrapolated and the vanilla finetuned model. Experiments across three tasks over four different LLMs show that EpiCoDe consistently outperforms existing methods with significant and robust improvement. We also propose a new theoretical framework to reveal the mechanism behind contrastive decoding in data-scarcity scenarios, which further helps us better understand the effectiveness of EpiCoDe.

Chuxiong Sun, Jie Hu, Hongming Gu et al.

Graph Neural Networks (GNNs) have received much attention in the graph deep learning domain. However, recent research empirically and theoretically shows that deep GNNs suffer from over-fitting and over-smoothing problems. The usual solutions either cannot solve extensive runtime of deep GNNs or restrict graph convolution in the same feature space. We propose the Adaptive Graph Diffusion Networks (AGDNs) which perform multi-layer generalized graph diffusion in different feature spaces with moderate complexity and runtime. Standard graph diffusion methods combine large and dense powers of the transition matrix with predefined weighting coefficients. Instead, AGDNs combine smaller multi-hop node representations with learnable and generalized weighting coefficients. We propose two scalable mechanisms of weighting coefficients to capture multi-hop information: Hop-wise Attention (HA) and Hop-wise Convolution (HC). We evaluate AGDNs on diverse, challenging Open Graph Benchmark (OGB) datasets with semi-supervised node classification and link prediction tasks. Until the date of submission (Aug 26, 2022), AGDNs achieve top-1 performance on the ogbn-arxiv, ogbn-proteins and ogbl-ddi datasets and top-3 performance on the ogbl-citation2 dataset. On the similar Tesla V100 GPU cards, AGDNs outperform Reversible GNNs (RevGNNs) with 13% complexity and 1% training runtime of RevGNNs on the ogbn-proteins dataset. AGDNs also achieve comparable performance to SEAL with 36% training and 0.2% inference runtime of SEAL on the ogbl-citation2 dataset.