Knowledge distillation (KD) is a technique that compresses large teacher models by training smaller student models to mimic them. The success of KD in auto-regressive language models mainly relies on Reverse KL for mode-seeking and student-generated output (SGO) to combat exposure bias. Our theoretical analyses and experimental validation reveal that while Reverse KL effectively mimics certain features of the teacher distribution, it fails to capture most of its behaviors. Conversely, SGO incurs higher computational costs and presents challenges in optimization, particularly when the student model is significantly smaller than the teacher model. These constraints are primarily due to the immutable distribution of the teacher model, which fails to adjust adaptively to models of varying sizes. We introduce Online Knowledge Distillation (OKD), where the teacher network integrates small online modules to concurrently train with the student model. This strategy abolishes the necessity for on-policy sampling and merely requires minimal updates to the parameters of the teacher's online module during training, thereby allowing dynamic adaptation to the student's distribution to make distillation better. Extensive results across multiple generation datasets show that OKD achieves or exceeds the performance of leading methods in various model architectures and sizes, reducing training time by up to fourfold.

Recently, unsupervised sentence embedding models have received significant attention in downstream natural language processing tasks. Using large language models (LLMs) for data augmentation has led to considerable improvements in previous studies. Nevertheless, these strategies emphasize data augmentation with extensive generic corpora, neglecting the consideration of few-shot domain data. The synthesized data lacks fine-grained information and may introduce negative sample noise. This study introduces a novel pipeline-based data augmentation method that leverages LLM to synthesize the domain-specific dataset. It produces both positive and negative samples through entity- and quantity-aware augmentation, utilizing an entity knowledge graph to synthesize samples with fine-grained semantic distinctions, increasing training sample diversity and relevance. We then present a Gaussian-decayed gradient-assisted Contrastive Sentence Embedding (GCSE) model to reduce synthetic data noise and improve model discrimination to reduce negative sample noise. Experimental results demonstrate that our approach achieves state-of-the-art semantic textual similarity performance with fewer synthetic data samples and lesser LLM parameters, demonstrating its efficiency and robustness in varied backbones.

Enhancing small language models for real-life application deployment is a significant challenge facing the research community. Due to the difficulties and costs of using large language models, researchers are seeking ways to effectively deploy task-specific small models. In this work, we introduce a simple yet effective knowledge distillation method to improve the performance of small language models. Our approach utilizes a teacher model with approximately 3 billion parameters to identify the most influential tokens in its decision-making process. These tokens are extracted from the input based on their attribution scores relative to the output, using methods like saliency maps. These important tokens are then provided as rationales to a student model, aiming to distill the knowledge of the teacher model. This method has proven to be effective, as demonstrated by testing it on four diverse datasets, where it shows improvement over both standard fine-tuning methods and state-of-the-art knowledge distillation models. Furthermore, we explore explanations of the success of the model by analyzing the important tokens extracted from the teacher model. Our findings reveal that in 68\% of cases, specifically in datasets where labels are part of the answer, such as multiple-choice questions, the extracted tokens are part of the ground truth.

This paper introduces a novel framework combining LLM agents as proxies for human strategic behavior with reinforcement learning (RL) to engage these agents in evolving strategic interactions within team environments. Our approach extends traditional agent-based simulations by using strategic LLM agents (SLA) and introducing dynamic and adaptive governance through a pro-social promoting RL agent (PPA) that modulates information access across agents in a network, optimizing social welfare and promoting pro-social behavior. Through validation in iterative games, including the prisoner's dilemma, we demonstrate that SLA agents exhibit nuanced strategic adaptations. The PPA agent effectively learns to adjust information transparency, resulting in enhanced cooperation rates. This framework offers significant insights into AI-mediated social dynamics, contributing to the deployment of AI in real-world team settings.

Retrieval Augmented Generation (RAG) has shown notable advancements in software engineering tasks. Despite its potential, RAG's application in unit test generation remains under-explored. To bridge this gap, we take the initiative to investigate the efficacy of RAG-based LLMs in test generation. As RAGs can leverage various knowledge sources to enhance their performance, we also explore the impact of different sources of RAGs' knowledge bases on unit test generation to provide insights into their practical benefits and limitations. Specifically, we examine RAG built upon three types of domain knowledge: 1) API documentation, 2) GitHub issues, and 3) StackOverflow Q&As. Each source offers essential knowledge for creating tests from different perspectives, i.e., API documentations provide official API usage guidelines, GitHub issues offer resolutions of issues related to the APIs from the library developers, and StackOverflow Q&As present community-driven solutions and best practices. For our experiment, we focus on five widely used and typical Python-based machine learning (ML) projects, i.e., TensorFlow, PyTorch, Scikit-learn, Google JAX, and XGBoost to build, train, and deploy complex neural networks efficiently. We conducted experiments using the top 10% most widely used APIs across these projects, involving a total of 188 APIs. We investigate the effectiveness of four state-of-the-art LLMs (open and closed-sourced), i.e., GPT-3.5-Turbo, GPT-4o, Mistral MoE 8x22B, and Llamma 3.1 405B. Additionally, we compare three prompting strategies in generating unit test cases for the experimental APIs, i.e., zero-shot, a Basic RAG, and an API-level RAG on the three external sources. Finally, we compare the cost of different sources of knowledge used for the RAG.

Layer normalization is a pivotal step in the transformer architecture. This paper delves into the less explored geometric implications of this process, examining how LayerNorm influences the norm and orientation of hidden vectors in the representation space. We show that the definition of LayerNorm is innately linked to the uniform vector, defined as $\boldsymbol{1} = [1, 1, 1, 1, \cdots, 1]^T \in \mathbb{R}^d$. We then show that the standardization step in LayerNorm can be understood in three simple steps: (i) remove the component of a vector along the uniform vector, (ii) normalize the remaining vector, and (iii) scale the resultant vector by $\sqrt{d}$, where $d$ is the dimensionality of the representation space. We also introduce the property of "irreversibility" for LayerNorm, where we show that the information lost during the normalization process cannot be recovered. In other words, unlike batch normalization, LayerNorm cannot learn an identity transform. While we present possible arguments for removing the component along the uniform vector, the choice of removing this component seems arbitrary and not well motivated by the original authors. To evaluate the usefulness of this step, we compare the hidden representations of LayerNorm-based LLMs with models trained using RMSNorm and show that all LLMs naturally align representations orthogonal to the uniform vector, presenting the first mechanistic evidence that removing the component along the uniform vector in LayerNorm is a redundant step. Our findings support the use of RMSNorm over LayerNorm as it is not only more computationally efficient with comparable downstream performance, but also learns a similar distribution of hidden representations that operate orthogonal to the uniform vector.

Large Language Models (LLMs) have achieved remarkable success in various fields, prompting several studies to explore their potential in recommendation systems. However, these attempts have so far resulted in only modest improvements over traditional recommendation models. Moreover, three critical questions remain under-explored: firstly, the real value of LLMs' pre-trained weights, often considered to encapsulate world knowledge; secondly, the necessity of fine-tuning for recommendation tasks; lastly, whether LLMs can exhibit the same scalability benefits in recommendation systems as they do in other domains. In this paper, we propose a novel Hierarchical Large Language Model (HLLM) architecture designed to enhance sequential recommendation systems. Our approach employs a two-tier model: the first Item LLM extracts rich content features from the detailed text description of the item, while the second User LLM utilizes these features to predict users' future interests based on their interaction history. Extensive experiments demonstrate that our method effectively leverages the pre-trained capabilities of open-source LLMs, and further fine-tuning leads to significant performance boosts. Additionally, HLLM achieves excellent scalability, with the largest configuration utilizing 7B parameters for both item feature extraction and user interest modeling. Moreover, HLLM offers excellent training and serving efficiency, making it practical in real-world applications. Evaluations on two large-scale datasets, PixelRec and Amazon Reviews, show that HLLM achieves state-of-the-art results, outperforming traditional ID-based models by a wide margin. In online A/B testing, HLLM showcases notable gains, validating its practical impact in real-world recommendation scenarios. Codes are available at https://github.com/bytedance/HLLM.

Large language models (LLMs) have been widely applied to assist test generation with the source code under test provided as the context. This paper aims to answer the question: If the source code under test is incorrect, will LLMs be misguided when generating tests? The effectiveness of test cases is measured by their accuracy, coverage, and bug detection effectiveness. Our evaluation results with five open- and six closed-source LLMs on four datasets demonstrate that incorrect code can significantly mislead LLMs in generating correct, high-coverage, and bug-revealing tests. For instance, in the HumanEval dataset, LLMs achieve 80.45% test accuracy when provided with task descriptions and correct code, but only 57.12% when given task descriptions and incorrect code. For the APPS dataset, prompts with correct code yield tests that detect 39.85% of the bugs, while prompts with incorrect code detect only 19.61%. These findings have important implications for the deployment of LLM-based testing: using it on mature code may help protect against future regression, but on early-stage immature code, it may simply bake in errors. Our findings also underscore the need for further research to improve LLMs resilience against incorrect code in generating reliable and bug-revealing tests.

E-commerce stores enable multilingual product discovery which require accurate product title translation. Multilingual large language models (LLMs) have shown promising capacity to perform machine translation tasks, and it can also enhance and translate product titles cross-lingually in one step. However, product title translation often requires more than just language conversion because titles are short, lack context, and contain specialized terminology. This study proposes a retrieval-augmented generation (RAG) approach that leverages existing bilingual product information in e-commerce by retrieving similar bilingual examples and incorporating them as few-shot prompts to enhance LLM-based product title translation. Experiment results show that our proposed RAG approach improve product title translation quality with chrF score gains of up to 15.3% for language pairs where the LLM has limited proficiency.

The pre-training phase of language models often begins with randomly initialized parameters. With the current trends in scaling models, training their large number of parameters can be extremely slow and costly. In contrast, small language models are less expensive to train, but they often cannot achieve the accuracy of large models. In this paper, we explore an intriguing idea to connect these two different regimes: Can we develop a method to initialize large language models using smaller pre-trained models? Will such initialization bring any benefits in terms of training time and final accuracy? In this paper, we introduce HyperCloning, a method that can expand the parameters of a pre-trained language model to those of a larger model with increased hidden dimensions. Our method ensures that the larger model retains the functionality of the smaller model. As a result, the larger model already inherits the predictive power and accuracy of the smaller model before the training starts. We demonstrate that training such an initialized model results in significant savings in terms of GPU hours required for pre-training large language models.