Yu Yu

Yongqi Li, Yu Yu, Tieyun Qian

Despite the recent success achieved by several two-stage prototypical networks in few-shot named entity recognition (NER) task, the overdetected false spans at the span detection stage and the inaccurate and unstable prototypes at the type classification stage remain to be challenging problems. In this paper, we propose a novel Type-Aware Decomposed framework, namely TadNER, to solve these problems. We first present a type-aware span filtering strategy to filter out false spans by removing those semantically far away from type names. We then present a type-aware contrastive learning strategy to construct more accurate and stable prototypes by jointly exploiting support samples and type names as references. Extensive experiments on various benchmarks prove that our proposed TadNER framework yields a new state-of-the-art performance. Our code and data will be available at https://github.com/NLPWM-WHU/TadNER.

Yu Yu, Chao-Han Huck Yang, Jari Kolehmainen et al.

We propose a neural language modeling system based on low-rank adaptation (LoRA) for speech recognition output rescoring. Although pretrained language models (LMs) like BERT have shown superior performance in second-pass rescoring, the high computational cost of scaling up the pretraining stage and adapting the pretrained models to specific domains limit their practical use in rescoring. Here we present a method based on low-rank decomposition to train a rescoring BERT model and adapt it to new domains using only a fraction (0.08%) of the pretrained parameters. These inserted matrices are optimized through a discriminative training objective along with a correlation-based regularization loss. The proposed low-rank adaptation Rescore-BERT (LoRB) architecture is evaluated on LibriSpeech and internal datasets with decreased training times by factors between 5.4 and 3.6.

Hongkang Yang, Zehao Lin, Wenjin Wang et al.

The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining "abstract knowledge". As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named $\text{Memory}^3$, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.

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.

Chanyeol Choi, Yoon Kim, Yu Yu et al.

Natural language processing (NLP) has been widely used in quantitative finance, but traditional methods often struggle to capture rich narratives in corporate disclosures, leaving potentially informative signals under-explored. Large language models (LLMs) offer a promising alternative due to their ability to extract nuanced semantics. In this paper, we ask whether semantic signals extracted by LLMs from corporate disclosures predict alpha, defined as abnormal returns beyond broad market movements and common risk factors. We introduce a simple framework, LLM as extractor, embedding as ruler, which extracts context-aware, metric-focused textual spans and quantifies semantic changes across consecutive disclosure periods using embedding-based similarity. This allows us to measure the degree of metric shifting -- how much firms move away from previously emphasized metrics, referred as moving targets. In experiments with portfolio and cross-sectional regression tests against a recent NER-based baseline, our method achieves more than twice the risk-adjusted alpha and shows significantly stronger predictive power. Qualitative analysis suggests that these gains stem from preserving contextual qualifiers and filtering out non-metric terms that keyword-based approaches often miss.

Boyi Zeng, Lizheng Wang, Yuncong Hu et al.

Protecting the copyright of large language models (LLMs) has become crucial due to their resource-intensive training and accompanying carefully designed licenses. However, identifying the original base model of an LLM is challenging due to potential parameter alterations. In this study, we introduce HuRef, a human-readable fingerprint for LLMs that uniquely identifies the base model without interfering with training or exposing model parameters to the public. We first observe that the vector direction of LLM parameters remains stable after the model has converged during pretraining, with negligible perturbations through subsequent training steps, including continued pretraining, supervised fine-tuning, and RLHF, which makes it a sufficient condition to identify the base model. The necessity is validated by continuing to train an LLM with an extra term to drive away the model parameters' direction and the model becomes damaged. However, this direction is vulnerable to simple attacks like dimension permutation or matrix rotation, which significantly change it without affecting performance. To address this, leveraging the Transformer structure, we systematically analyze potential attacks and define three invariant terms that identify an LLM's base model. Due to the potential risk of information leakage, we cannot publish invariant terms directly. Instead, we map them to a Gaussian vector using an encoder, then convert it into a natural image using StyleGAN2, and finally publish the image. In our black-box setting, all fingerprinting steps are internally conducted by the LLMs owners. To ensure the published fingerprints are honestly generated, we introduced Zero-Knowledge Proof (ZKP). Experimental results across various LLMs demonstrate the effectiveness of our method. The code is available at https://github.com/LUMIA-Group/HuRef.

Junyi Zhu, Shuochen Liu, Yu Yu et al.

Large language models (LLMs) excel in generating coherent text, but they often struggle with context awareness, leading to inaccuracies in tasks requiring faithful adherence to provided information. We introduce FastMem, a novel method designed to enhance instruction fine-tuned LLMs' context awareness through fast memorization of the prompt. FastMem maximizes the likelihood of the prompt before inference by updating only the last Feed-Forward Network (FFN) module. This targeted approach ensures efficient optimization without overfitting, significantly improving the model's ability to comprehend and accurately follow the context. Our experiments demonstrate substantial gains in reading comprehension, text summarization and adherence to output structures. For instance, FastMem improves the accuracy of Llama 3-8B-Inst on the NQ-SWAP dataset from 59.1% to 71.6%, and reduces the output structure failure rate of Qwen 1.5-4B-Chat from 34.9% to 25.5%. Extensive experimental results highlight FastMem's potential to offer a robust solution to enhance the reliability and accuracy of LLMs in various applications. Our code is available at: https://github.com/IAAR-Shanghai/FastMem

Zhengyi Li, Yakai Wang, Kang Yang et al.

For Transformer models, cryptographically secure inference ensures that the client learns only the final output, while the server learns nothing about the client's input. However, securely computing nonlinear layers remains a major efficiency bottleneck due to the substantial communication rounds and data transmission required. To address this issue, prior works reveal intermediate activations to the client, allowing nonlinear operations to be computed in plaintext. Although this approach significantly improves efficiency, exposing activations enables adversaries to extract model weights. To mitigate this risk, existing works employ a shuffling defense that reveals only randomly permuted activations to the client. In this work, we show that the shuffling defense is not as robust as previously claimed. We propose an attack that aligns differently shuffled activations to a common permutation and subsequently exploits them to extract model weights. Experiments on Pythia-70m and GPT-2 demonstrate that the proposed attack can align shuffled activations with mean squared errors ranging from $10^{-9}$ to $10^{-6}$. With a query cost of approximately \$1, the adversary can recover model weights with L1-norm differences ranging from $10^{-4}$ to $10^{-2}$ compared to the oracle weights.

Zhengyi Li, Kang Yang, Jin Tan et al.

Transformer models have gained significant attention due to their power in machine learning tasks. Their extensive deployment has raised concerns about the potential leakage of sensitive information during inference. However, when being applied to Transformers, existing approaches based on secure two-party computation (2PC) bring about efficiency limitations in two folds: (1) resource-intensive matrix multiplications in linear layers, and (2) complex non-linear activation functions like $\mathsf{GELU}$ and $\mathsf{Softmax}$. This work presents a new two-party inference framework $\mathsf{Nimbus}$ for Transformer models. For the linear layer, we propose a new 2PC paradigm along with an encoding approach to securely compute matrix multiplications based on an outer-product insight, which achieves $2.9\times \sim 12.5\times$ performance improvements compared to the state-of-the-art (SOTA) protocol. For the non-linear layer, through a new observation of utilizing the input distribution, we propose an approach of low-degree polynomial approximation for $\mathsf{GELU}$ and $\mathsf{Softmax}$, which improves the performance of the SOTA polynomial approximation by $2.9\times \sim 4.0\times$, where the average accuracy loss of our approach is 0.08\% compared to the non-2PC inference without privacy. Compared with the SOTA two-party inference, $\mathsf{Nimbus}$ improves the end-to-end performance of \bert{} inference by $2.7\times \sim 4.7\times$ across different network settings.

Yu Yu, Qian Xie, Nairen Cao et al.

Designing state encoders for reinforcement learning (RL) with multiple information sources -- such as sensor measurements, time-series signals, image observations, and textual instructions -- remains underexplored and often requires manual design. We formalize this challenge as a problem of composite neural architecture search (NAS), where multiple source-specific modules and a fusion module are jointly optimized. Existing NAS methods overlook useful side information from the intermediate outputs of these modules -- such as their representation quality -- limiting sample efficiency in multi-source RL settings. To address this, we propose an LLM-driven NAS pipeline that leverages language-model priors and intermediate-output signals to guide sample-efficient search for high-performing composite state encoders. On a mixed-autonomy traffic control task, our approach discovers higher-performing architectures with fewer candidate evaluations than traditional NAS baselines and the LLM-based GENIUS framework.

Zhengyi Li, Yue Guan, Kang Yang et al.

The wide deployment of the generative pre-trained transformer (GPT) has raised privacy concerns for both clients and servers. While cryptographic primitives can be employed for secure GPT inference to protect the privacy of both parties, they introduce considerable performance overhead.To accelerate secure inference, this study proposes a public decoding and secure verification approach that utilizes public GPT models, motivated by the observation that securely decoding one and multiple tokens takes a similar latency. The client uses the public model to generate a set of tokens, which are then securely verified by the private model for acceptance. The efficiency of our approach depends on the acceptance ratio of tokens proposed by the public model, which we improve from two aspects: (1) a private sampling protocol optimized for cryptographic primitives and (2) model alignment using knowledge distillation. Our approach improves the efficiency of secure decoding while maintaining the same level of privacy and generation quality as standard secure decoding. Experiments demonstrate a $2.1\times \sim 6.0\times$ speedup compared to standard decoding across three pairs of public-private models and different network conditions.

Fake Lin, Xi Zhu, Ziwei Zhao et al.

Recent years have witnessed the prosperity of knowledge graph based recommendation system (KGRS), which enriches the representation of users, items, and entities by structural knowledge with striking improvement. Nevertheless, its unaffordable computational cost still limits researchers from exploring more sophisticated models. We observe that the bottleneck for training efficiency arises from the knowledge graph, which is plagued by the well-known issue of knowledge explosion. Recently, some works have attempted to slim the inflated KG via summarization techniques. However, these summarized nodes may ignore the collaborative signals and deviate from the facts that nodes in knowledge graph represent symbolic abstractions of entities from the real-world. To this end, in this paper, we propose a novel approach called KGTrimmer for knowledge graph pruning tailored for recommendation, to remove the unessential nodes while minimizing performance degradation. Specifically, we design an importance evaluator from a dual-view perspective. For the collective view, we embrace the idea of collective intelligence by extracting community consensus based on abundant collaborative signals, i.e. nodes are considered important if they attract attention of numerous users. For the holistic view, we learn a global mask to identify the valueless nodes from their inherent properties or overall popularity. Next, we build an end-to-end importance-aware graph neural network, which injects filtered knowledge to enhance the distillation of valuable user-item collaborative signals. Ultimately, we generate a pruned knowledge graph with lightweight, stable, and robust properties to facilitate the following-up recommendation task. Extensive experiments are conducted on three publicly available datasets to prove the effectiveness and generalization ability of KGTrimmer.

George Ma, Anurag Koul, Qi Chen et al. · amazon-science

Large Language Models (LLMs) excel at code-related tasks but often struggle in realistic software repositories, where project-specific APIs and cross-file dependencies are crucial. Retrieval-augmented methods mitigate this by injecting repository context at inference time. The low inference-time latency budget affects either retrieval quality or the added latency adversely impacts user experience. We address this limitation with SpecAgent, an agent that improves both latency and code-generation quality by proactively exploring repository files during indexing and constructing speculative context that anticipates future edits in each file. This indexing-time asynchrony allows thorough context computation, masking latency, and the speculative nature of the context improves code-generation quality. Additionally, we identify the problem of future context leakage in existing benchmarks, which can inflate reported performance. To address this, we construct a synthetic, leakage-free benchmark that enables a more realistic evaluation of our agent against baselines. Experiments show that SpecAgent consistently achieves absolute gains of 9-11% (48-58% relative) compared to the best-performing baselines, while significantly reducing inference latency.

Yuan Tian, Min Zhou, Yitong Chen et al.

Patient face images provide a convenient mean for evaluating eye diseases, while also raising privacy concerns. Here, we introduce ROFI, a deep learning-based privacy protection framework for ophthalmology. Using weakly supervised learning and neural identity translation, ROFI anonymizes facial features while retaining disease features (over 98\% accuracy, $κ> 0.90$). It achieves 100\% diagnostic sensitivity and high agreement ($κ> 0.90$) across eleven eye diseases in three cohorts, anonymizing over 95\% of images. ROFI works with AI systems, maintaining original diagnoses ($κ> 0.80$), and supports secure image reversal (over 98\% similarity), enabling audits and long-term care. These results show ROFI's effectiveness of protecting patient privacy in the digital medicine era.

Bo Tang, Junyi Zhu, Chenyang Xi et al.

Traditional search engines struggle to synthesize fragmented information for complex queries, while generative AI search engines face challenges in relevance, comprehensiveness, and presentation. To address these limitations, we introduce Xinyu AI Search, a novel system that incorporates a query-decomposition graph to dynamically break down complex queries into sub-queries, enabling stepwise retrieval and generation. Our retrieval pipeline enhances diversity through multi-source aggregation and query expansion, while filtering and re-ranking strategies optimize passage relevance. Additionally, Xinyu AI Search introduces a novel approach for fine-grained, precise built-in citation and innovates in result presentation by integrating timeline visualization and textual-visual choreography. Evaluated on recent real-world queries, Xinyu AI Search outperforms eight existing technologies in human assessments, excelling in relevance, comprehensiveness, and insightfulness. Ablation studies validate the necessity of its key sub-modules. Our work presents the first comprehensive framework for generative AI search engines, bridging retrieval, generation, and user-centric presentation.

Yidan Wu, Yu Yu, Jianan Zhang et al.

We consider the traffic control problem of dynamic routing over parallel servers, which arises in a variety of engineering systems such as transportation and data transmission. We propose a semi-gradient, on-policy algorithm that learns an approximate optimal routing policy. The algorithm uses generic basis functions with flexible weights to approximate the value function across the unbounded state space. Consequently, the training process lacks Lipschitz continuity of the gradient, boundedness of the temporal-difference error, and a prior guarantee on ergodicity, which are the standard prerequisites in existing literature on reinforcement learning theory. To address this, we combine a Lyapunov approach and an ordinary differential equation-based method to jointly characterize the behavior of traffic state and approximation weights. Our theoretical analysis proves that the training scheme guarantees traffic state stability and ensures almost surely convergence of the weights to the approximate optimum. We also demonstrate via simulations that our algorithm attains significantly faster convergence than neural network-based methods with an insignificant approximation error.

Yu Yu, Chao-Han Huck Yang, Tuan Dinh et al.

The use of low-rank adaptation (LoRA) with frozen pretrained language models (PLMs) has become increasing popular as a mainstream, resource-efficient modeling approach for memory-constrained hardware. In this study, we first explore how to enhance model performance by introducing various LoRA training strategies, achieving relative word error rate reductions of 3.50\% on the public Librispeech dataset and of 3.67\% on an internal dataset in the messaging domain. To further characterize the stability of LoRA-based second-pass speech recognition models, we examine robustness against input perturbations. These perturbations are rooted in homophone replacements and a novel metric called N-best Perturbation-based Rescoring Robustness (NPRR), both designed to measure the relative degradation in the performance of rescoring models. Our experimental results indicate that while advanced variants of LoRA, such as dynamic rank-allocated LoRA, lead to performance degradation in $1$-best perturbation, they alleviate the degradation in $N$-best perturbation. This finding is in comparison to fully-tuned models and vanilla LoRA tuning baselines, suggesting that a comprehensive selection is needed when using LoRA-based adaptation for compute-cost savings and robust language modeling.

Nikhil Muralidhar, Sathappah Muthiah, Patrick Butler et al.

We describe lessons learned from developing and deploying machine learning models at scale across the enterprise in a range of financial analytics applications. These lessons are presented in the form of antipatterns. Just as design patterns codify best software engineering practices, antipatterns provide a vocabulary to describe defective practices and methodologies. Here we catalog and document numerous antipatterns in financial ML operations (MLOps). Some antipatterns are due to technical errors, while others are due to not having sufficient knowledge of the surrounding context in which ML results are used. By providing a common vocabulary to discuss these situations, our intent is that antipatterns will support better documentation of issues, rapid communication between stakeholders, and faster resolution of problems. In addition to cataloging antipatterns, we describe solutions, best practices, and future directions toward MLOps maturity.

Yu Yu, Jean-Marc Odobez

Although automatic gaze estimation is very important to a large variety of application areas, it is difficult to train accurate and robust gaze models, in great part due to the difficulty in collecting large and diverse data (annotating 3D gaze is expensive and existing datasets use different setups). To address this issue, our main contribution in this paper is to propose an effective approach to learn a low dimensional gaze representation without gaze annotations, which to the best of our best knowledge, is the first work to do so. The main idea is to rely on a gaze redirection network and use the gaze representation difference of the input and target images (of the redirection network) as the redirection variable. A redirection loss in image domain allows the joint training of both the redirection network and the gaze representation network. In addition, we propose a warping field regularization which not only provides an explicit physical meaning to the gaze representations but also avoids redirection distortions. Promising results on few-shot gaze estimation (competitive results can be achieved with as few as <= 100 calibration samples), cross-dataset gaze estimation, gaze network pretraining, and another task (head pose estimation) demonstrate the validity of our framework.

Yu Yu, Gang Liu, Jean-Marc Odobez

As an indicator of human attention gaze is a subtle behavioral cue which can be exploited in many applications. However, inferring 3D gaze direction is challenging even for deep neural networks given the lack of large amount of data (groundtruthing gaze is expensive and existing datasets use different setups) and the inherent presence of gaze biases due to person-specific difference. In this work, we address the problem of person-specific gaze model adaptation from only a few reference training samples. The main and novel idea is to improve gaze adaptation by generating additional training samples through the synthesis of gaze-redirected eye images from existing reference samples. In doing so, our contributions are threefold: (i) we design our gaze redirection framework from synthetic data, allowing us to benefit from aligned training sample pairs to predict accurate inverse mapping fields; (ii) we proposed a self-supervised approach for domain adaptation; (iii) we exploit the gaze redirection to improve the performance of person-specific gaze estimation. Extensive experiments on two public datasets demonstrate the validity of our gaze retargeting and gaze estimation framework.

Gang Liu, Yu Yu, Kenneth A. Funes Mora et al.

Non-invasive gaze estimation methods usually regress gaze directions directly from a single face or eye image. However, due to important variabilities in eye shapes and inner eye structures amongst individuals, universal models obtain limited accuracies and their output usually exhibit high variance as well as biases which are subject dependent. Therefore, increasing accuracy is usually done through calibration, allowing gaze predictions for a subject to be mapped to his/her actual gaze. In this paper, we introduce a novel image differential method for gaze estimation. We propose to directly train a differential convolutional neural network to predict the gaze differences between two eye input images of the same subject. Then, given a set of subject specific calibration images, we can use the inferred differences to predict the gaze direction of a novel eye sample. The assumption is that by allowing the comparison between two eye images, annoyance factors (alignment, eyelid closing, illumination perturbations) which usually plague single image prediction methods can be much reduced, allowing better prediction altogether. Experiments on 3 public datasets validate our approach which constantly outperforms state-of-the-art methods even when using only one calibration sample or when the latter methods are followed by subject specific gaze adaptation.

Yi Li, Yitao Duan, Yu Yu et al.

We introduce PrivPy, a practical privacy-preserving collaborative computation framework, especially optimized for machine learning tasks. PrivPy provides an easy-to-use and highly compatible Python programming front-end which supports high-level array operations and different secure computation engines to allow for security assumptions and performance trade-offs. With PrivPy, programmers can write modern machine learning algorithms conveniently and efficiently in Python. We also design and implement a new efficient computation engine, with which people can use competing cloud providers to efficiently perform general arithmetics over real numbers. We demonstrate the usability and scalability of PrivPy using common machine learning models (e.g. logistic regression and convolutional neural networks) and real-world datasets (including a 5000-by-1-million matrix).