Jinyang Gao

Junkang Wu, Yuexiang Xie, Zhengyi Yang et al.

Direct Preference Optimization (DPO) has emerged as a compelling approach for training Large Language Models (LLMs) to adhere to human preferences. However, the performance of DPO is sensitive to the fine-tuning of its trade-off parameter $β$, as well as to the quality of the preference data. We analyze the impact of $β$ and data quality on DPO, uncovering that optimal $β$ values vary with the informativeness of pairwise data. Addressing the limitations of static $β$ values, we introduce a novel framework that dynamically calibrates $β$ at the batch level, informed by data quality considerations. Additionally, our method incorporates $β$-guided data filtering to safeguard against the influence of outliers. Through empirical evaluation, we demonstrate that our dynamic $β$ adjustment technique significantly improves DPO's performance across a range of models and datasets, offering a more robust and adaptable training paradigm for aligning LLMs with human feedback. The code is available at \url{https://github.com/junkangwu/beta-DPO}.

Junkang Wu, Yuexiang Xie, Zhengyi Yang et al.

This study addresses the challenge of noise in training datasets for Direct Preference Optimization (DPO), a method for aligning Large Language Models (LLMs) with human preferences. We categorize noise into pointwise noise, which includes low-quality data points, and pairwise noise, which encompasses erroneous data pair associations that affect preference rankings. Utilizing Distributionally Robust Optimization (DRO), we enhance DPO's resilience to these types of noise. Our theoretical insights reveal that DPO inherently embeds DRO principles, conferring robustness to pointwise noise, with the regularization coefficient $β$ playing a critical role in its noise resistance. Extending this framework, we introduce Distributionally Robustifying DPO (Dr. DPO), which integrates pairwise robustness by optimizing against worst-case pairwise scenarios. The novel hyperparameter $β'$ in Dr. DPO allows for fine-tuned control over data pair reliability, providing a strategic balance between exploration and exploitation in noisy training environments. Empirical evaluations demonstrate that Dr. DPO substantially improves the quality of generated text and response accuracy in preference datasets, showcasing enhanced performance in both noisy and noise-free settings. The code is available at https://github.com/junkangwu/Dr_DPO.

Lizhe Fang, Yifei Wang, Zhaoyang Liu et al.

Handling long-context inputs is crucial for large language models (LLMs) in tasks such as extended conversations, document summarization, and many-shot in-context learning. While recent approaches have extended the context windows of LLMs and employed perplexity (PPL) as a standard evaluation metric, PPL has proven unreliable for assessing long-context capabilities. The underlying cause of this limitation has remained unclear. In this work, we provide a comprehensive explanation for this issue. We find that PPL overlooks key tokens, which are essential for long-context understanding, by averaging across all tokens and thereby obscuring the true performance of models in long-context scenarios. To address this, we propose \textbf{LongPPL}, a novel metric that focuses on key tokens by employing a long-short context contrastive method to identify them. Our experiments demonstrate that LongPPL strongly correlates with performance on various long-context benchmarks (e.g., Pearson correlation of -0.96), significantly outperforming traditional PPL in predictive accuracy. Additionally, we introduce \textbf{LongCE} (Long-context Cross-Entropy) loss, a re-weighting strategy for fine-tuning that prioritizes key tokens, leading to consistent improvements across diverse benchmarks. In summary, these contributions offer deeper insights into the limitations of PPL and present effective solutions for accurately evaluating and enhancing the long-context capabilities of LLMs. Code is available at https://github.com/PKU-ML/LongPPL.

Junkang Wu, Xue Wang, Zhengyi Yang et al.

Aligning large language models (LLMs) with human values and intentions is crucial for their utility, honesty, and safety. Reinforcement learning from human feedback (RLHF) is a popular approach to achieve this alignment, but it faces challenges in computational efficiency and training stability. Recent methods like Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO) have proposed offline alternatives to RLHF, simplifying the process by reparameterizing the reward function. However, DPO depends on a potentially suboptimal reference model, and SimPO's assumption of a fixed target reward margin may lead to suboptimal decisions in diverse data settings. In this work, we propose $α$-DPO, an adaptive preference optimization algorithm designed to address these limitations by introducing a dynamic reward margin. Specifically, $α$-DPO employs an adaptive preference distribution, balancing the policy model and the reference model to achieve personalized reward margins. We provide theoretical guarantees for $α$-DPO, demonstrating its effectiveness as a surrogate optimization objective and its ability to balance alignment and diversity through KL divergence control. Empirical evaluations on AlpacaEval 2 and Arena-Hard show that $α$-DPO consistently outperforms DPO and SimPO across various model settings, establishing it as a robust approach for fine-tuning LLMs. Our method achieves significant improvements in win rates, highlighting its potential as a powerful tool for LLM alignment. The code is available at https://github.com/junkangwu/alpha-DPO

Zhisheng Lin, Yifu Liu, Zhiling Luo et al.

The improvement in translating natural language to structured query language (SQL) can be attributed to the advancements in large language models (LLMs). Open-source LLMs, tailored for specific database dialects such as MySQL, have shown great performance. However, cloud service providers are looking for a unified database manager service (e.g., Cosmos DB from Azure, Amazon Aurora from AWS, Lindorm from AlibabaCloud) that can support multiple dialects. This requirement has led to the concept of multi-dialect query generation, which presents challenges to LLMs. These challenges include syntactic differences among dialects and imbalanced data distribution across multiple dialects. To tackle these challenges, we propose MoMQ, a novel Mixture-of-Experts-based multi-dialect query generation framework across both relational and non-relational databases. MoMQ employs a dialect expert group for each dialect and a multi-level routing strategy to handle dialect-specific knowledge, reducing interference during query generation. Additionally, a shared expert group is introduced to address data imbalance, facilitating the transfer of common knowledge from high-resource dialects to low-resource ones. Furthermore, we have developed a high-quality multi-dialect query generation benchmark that covers relational and non-relational databases such as MySQL, PostgreSQL, Cypher for Neo4j, and nGQL for NebulaGraph. Extensive experiments have shown that MoMQ performs effectively and robustly even in resource-imbalanced scenarios.

Yingqi Gao, Yifu Liu, Xiaoxia Li et al.

To tackle the challenges of large language model performance in natural language to SQL tasks, we introduce XiYan-SQL, an innovative framework that employs a multi-generator ensemble strategy to improve candidate generation. We introduce M-Schema, a semi-structured schema representation method designed to enhance the understanding of database structures. To enhance the quality and diversity of generated candidate SQL queries, XiYan-SQL integrates the significant potential of in-context learning (ICL) with the precise control of supervised fine-tuning. On one hand, we propose a series of training strategies to fine-tune models to generate high-quality candidates with diverse preferences. On the other hand, we implement the ICL approach with an example selection method based on named entity recognition to prevent overemphasis on entities. The refiner optimizes each candidate by correcting logical or syntactical errors. To address the challenge of identifying the best candidate, we fine-tune a selection model to distinguish nuances of candidate SQL queries. The experimental results on multiple dialect datasets demonstrate the robustness of XiYan-SQL in addressing challenges across different scenarios. Overall, our proposed XiYan-SQL achieves the state-of-the-art execution accuracy of 75.63% on Bird benchmark, 89.65% on the Spider test set, 69.86% on SQL-Eval, 41.20% on NL2GQL. The proposed framework not only enhances the quality and diversity of SQL queries but also outperforms previous methods.

Yifu Liu, Yin Zhu, Yingqi Gao et al.

To leverage the advantages of LLM in addressing challenges in the Text-to-SQL task, we present XiYan-SQL, an innovative framework effectively generating and utilizing multiple SQL candidates. It consists of three components: 1) a Schema Filter module filtering and obtaining multiple relevant schemas; 2) a multi-generator ensemble approach generating multiple highquality and diverse SQL queries; 3) a selection model with a candidate reorganization strategy implemented to obtain the optimal SQL query. Specifically, for the multi-generator ensemble, we employ a multi-task fine-tuning strategy to enhance the capabilities of SQL generation models for the intrinsic alignment between SQL and text, and construct multiple generation models with distinct generation styles by fine-tuning across different SQL formats. The experimental results and comprehensive analysis demonstrate the effectiveness and robustness of our framework. Overall, XiYan-SQL achieves a new SOTA performance of 75.63% on the notable BIRD benchmark, surpassing all previous methods. It also attains SOTA performance on the Spider test set with an accuracy of 89.65%.

Xu Xie, Zhaoyang Liu, Shiwen Wu et al.

To improve user experience and profits of corporations, modern industrial recommender systems usually aim to select the items that are most likely to be interacted with (e.g., clicks and purchases). However, they overlook the fact that users may purchase the items even without recommendations. To select these effective items, it is essential to estimate the causal effect of recommendations. The real effective items are the ones which can contribute to purchase probability uplift. Nevertheless, it is difficult to obtain the real causal effect since we can only recommend or not recommend an item to a user at one time. Furthermore, previous works usually rely on the randomized controlled trial~(RCT) experiment to evaluate their performance. However, it is usually not practicable in the recommendation scenario due to its unavailable time consuming. To tackle these problems, in this paper, we propose a causal collaborative filtering~(CausCF) method inspired by the widely adopted collaborative filtering~(CF) technique. It is based on the idea that similar users not only have a similar taste on items, but also have similar treatment effect under recommendations. CausCF extends the classical matrix factorization to the tensor factorization with three dimensions -- user, item, and treatment. Furthermore, we also employs regression discontinuity design (RDD) to evaluate the precision of the estimated causal effects from different models. With the testable assumptions, RDD analysis can provide an unbiased causal conclusion without RCT experiments. Through dedicated experiments on both the public datasets and the industrial application, we demonstrate the effectiveness of our proposed CausCF on the causal effect estimation and ranking performance improvement.

Xu Xie, Fei Sun, Xiaoyong Yang et al.

Recommender systems play a vital role in modern online services, such as Amazon and Taobao. Traditional personalized methods, which focus on user-item (UI) relations, have been widely applied in industrial settings, owing to their efficiency and effectiveness. Despite their success, we argue that these approaches ignore local information hidden in similar users. To tackle this problem, user-based methods exploit similar user relations to make recommendations in a local perspective. Nevertheless, traditional user-based methods, like userKNN and matrix factorization, are intractable to be deployed in the real-time applications since such transductive models have to be recomputed or retrained with any new interaction. To overcome this challenge, we propose a framework called self-complementary collaborative filtering~(SCCF) which can make recommendations with both global and local information in real time. On the one hand, it utilizes UI relations and user neighborhood to capture both global and local information. On the other hand, it can identify similar users for each user in real time by inferring user representations on the fly with an inductive model. The proposed framework can be seamlessly incorporated into existing inductive UI approach and benefit from user neighborhood with little additional computation. It is also the first attempt to apply user-based methods in real-time settings. The effectiveness and efficiency of SCCF are demonstrated through extensive offline experiments on four public datasets, as well as a large scale online A/B test in Taobao.

Xu Xie, Fei Sun, Zhaoyang Liu et al.

Sequential recommendation methods play a crucial role in modern recommender systems because of their ability to capture a user's dynamic interest from her/his historical interactions. Despite their success, we argue that these approaches usually rely on the sequential prediction task to optimize the huge amounts of parameters. They usually suffer from the data sparsity problem, which makes it difficult for them to learn high-quality user representations. To tackle that, inspired by recent advances of contrastive learning techniques in the computer version, we propose a novel multi-task model called \textbf{C}ontrastive \textbf{L}earning for \textbf{S}equential \textbf{Rec}ommendation~(\textbf{CL4SRec}). CL4SRec not only takes advantage of the traditional next item prediction task but also utilizes the contrastive learning framework to derive self-supervision signals from the original user behavior sequences. Therefore, it can extract more meaningful user patterns and further encode the user representation effectively. In addition, we propose three data augmentation approaches to construct self-supervision signals. Extensive experiments on four public datasets demonstrate that CL4SRec achieves state-of-the-art performance over existing baselines by inferring better user representations.

Chen Xu, Quan Li, Junfeng Ge et al.

Features play an important role in the prediction tasks of e-commerce recommendations. To guarantee the consistency of off-line training and on-line serving, we usually utilize the same features that are both available. However, the consistency in turn neglects some discriminative features. For example, when estimating the conversion rate (CVR), i.e., the probability that a user would purchase the item if she clicked it, features like dwell time on the item detailed page are informative. However, CVR prediction should be conducted for on-line ranking before the click happens. Thus we cannot get such post-event features during serving. We define the features that are discriminative but only available during training as the privileged features. Inspired by the distillation techniques which bridge the gap between training and inference, in this work, we propose privileged features distillation (PFD). We train two models, i.e., a student model that is the same as the original one and a teacher model that additionally utilizes the privileged features. Knowledge distilled from the more accurate teacher is transferred to the student to improve its accuracy. During serving, only the student part is extracted and it relies on no privileged features. We conduct experiments on two fundamental prediction tasks at Taobao recommendations, i.e., click-through rate (CTR) at coarse-grained ranking and CVR at fine-grained ranking. By distilling the interacted features that are prohibited during serving for CTR and the post-event features for CVR, we achieve significant improvements over their strong baselines. During the on-line A/B tests, the click metric is improved by +5.0% in the CTR task. And the conversion metric is improved by +2.3% in the CVR task. Besides, by addressing several issues of training PFD, we obtain comparable training speed as the baselines without any distillation.