Yuchen Zhang

Hugo Touvron, Louis Martin, Kevin Stone et al. · meta-ai

In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.

David Esiobu, Xiaoqing Tan, Saghar Hosseini et al. · meta-ai

As generative large language models (LLMs) grow more performant and prevalent, we must develop comprehensive enough tools to measure and improve their fairness. Different prompt-based datasets can be used to measure social bias across multiple text domains and demographic axes, meaning that testing LLMs on more datasets can potentially help us characterize their biases more fully, and better ensure equal and equitable treatment of marginalized demographic groups. In this work, our focus is two-fold: (1) Benchmarking: a comparison of 6 different prompt-based bias and toxicity metrics across 12 demographic axes and 5 families of generative LLMs. Out of those 6 metrics, AdvPromptSet and HolisticBiasR are novel datasets proposed in the paper. The comparison of those benchmarks gives us insights about the bias and toxicity of the compared models. Therefore, we explore the frequency of demographic terms in common LLM pre-training corpora and how this may relate to model biases. (2) Mitigation: we conduct a comprehensive study of how well 3 bias/toxicity mitigation techniques perform across our suite of measurements. ROBBIE aims to provide insights for practitioners while deploying a model, emphasizing the need to not only measure potential harms, but also understand how they arise by characterizing the data, mitigate harms once found, and balance any trade-offs. We open-source our analysis code in hopes of encouraging broader measurements of bias in future LLMs.

Wenyue Hua, Yuchen Zhang, Zhe Chen et al. · meta-ai, princeton

The application of Natural Language Processing (NLP) to specialized domains, such as the law, has recently received a surge of interest. As many legal services rely on processing and analyzing large collections of documents, automating such tasks with NLP tools emerges as a key challenge. Many popular language models, such as BERT or RoBERTa, are general-purpose models, which have limitations on processing specialized legal terminology and syntax. In addition, legal documents may contain specialized vocabulary from other domains, such as medical terminology in personal injury text. Here, we propose LegalRelectra, a legal-domain language model that is trained on mixed-domain legal and medical corpora. We show that our model improves over general-domain and single-domain medical and legal language models when processing mixed-domain (personal injury) text. Our training architecture implements the Electra framework, but utilizes Reformer instead of BERT for its generator and discriminator. We show that this improves the model's performance on processing long passages and results in better long-range text comprehension.

Shuo Sun, Yuchen Zhang, Jiahuan Yan et al.

The success of ChatGPT has ignited an AI race, with researchers striving to develop new large language models (LLMs) that can match or surpass the language understanding and generation abilities of commercial ones. In recent times, a number of models have emerged, claiming performance near that of GPT-3.5 or GPT-4 through various instruction-tuning methods. As practitioners of Text-to-SQL parsing, we are grateful for their valuable contributions to open-source research. However, it is important to approach these claims with a sense of scrutiny and ascertain the actual effectiveness of these models. Therefore, we pit six popular large language models against each other, systematically evaluating their Text-to-SQL parsing capability on nine benchmark datasets with five different prompting strategies, covering both zero-shot and few-shot scenarios. Regrettably, the open-sourced models fell significantly short of the performance achieved by closed-source models like GPT-3.5, highlighting the need for further work to bridge the performance gap between these models.

Xing Su, Jian Yang, Jia Wu et al.

Users' involvement in creating and propagating news is a vital aspect of fake news detection in online social networks. Intuitively, credible users are more likely to share trustworthy news, while untrusted users have a higher probability of spreading untrustworthy news. In this paper, we construct a dual-layer graph (i.e., the news layer and the user layer) to extract multiple relations of news and users in social networks to derive rich information for detecting fake news. Based on the dual-layer graph, we propose a fake news detection model named Us-DeFake. It learns the propagation features of news in the news layer and the interaction features of users in the user layer. Through the inter-layer in the graph, Us-DeFake fuses the user signals that contain credibility information into the news features, to provide distinctive user-aware embeddings of news for fake news detection. The training process conducts on multiple dual-layer subgraphs obtained by a graph sampler to scale Us-DeFake in large scale social networks. Extensive experiments on real-world datasets illustrate the superiority of Us-DeFake which outperforms all baselines, and the users' credibility signals learned by interaction relation can notably improve the performance of our model.

Guanhua Huang, Yuchen Zhang, Zhe Li et al.

The widespread use of large language models (LLMs) has sparked concerns about the potential misuse of AI-generated text, as these models can produce content that closely resembles human-generated text. Current detectors for AI-generated text (AIGT) lack robustness against adversarial perturbations, with even minor changes in characters or words causing a reversal in distinguishing between human-created and AI-generated text. This paper investigates the robustness of existing AIGT detection methods and introduces a novel detector, the Siamese Calibrated Reconstruction Network (SCRN). The SCRN employs a reconstruction network to add and remove noise from text, extracting a semantic representation that is robust to local perturbations. We also propose a siamese calibration technique to train the model to make equally confidence predictions under different noise, which improves the model's robustness against adversarial perturbations. Experiments on four publicly available datasets show that the SCRN outperforms all baseline methods, achieving 6.5\%-18.25\% absolute accuracy improvement over the best baseline method under adversarial attacks. Moreover, it exhibits superior generalizability in cross-domain, cross-genre, and mixed-source scenarios. The code is available at \url{https://github.com/CarlanLark/Robust-AIGC-Detector}.

Jie Xiao, Changyuan Fan, Qingnan Ren et al.

Modern RL-based post-training for large language models (LLMs) co-locate trajectory sampling and policy optimisation on the same GPU cluster, forcing the system to switch between inference and training workloads. This serial context switching violates the single-program-multiple-data (SPMD) assumption underlying today's distributed training systems. We present Echo, the RL system that cleanly decouples these two phases across heterogeneous "inference" and "training" swarms while preserving statistical efficiency. Echo introduces two lightweight synchronization protocols: a sequential pull mode that refreshes policy weights according to API call for minimal bias, and an asynchronous push-pull mode that streams version-tagged rollouts through a replay buffer to maximise hardware utilisation. Training four representative RL workloads with Qwen3-4B, Qwen2.5-7B, Qwen3-30B-A3B-Thinking-2507 and Qwen3-32B on a geographically distributed cluster, Echo matches a fully co-located Verl baseline in convergence speed and final reward while off-loading trajectory generation to commodity edge hardware. These promising results demonstrate that large-scale RL for LLMs could achieve datacentre-grade performance using decentralised, heterogeneous resources.

Min Cai, Yuchen Zhang, Shichang Zhang et al.

We propose SelfControl, an inference-time model control method utilizing gradients to control the behavior of large language models (LLMs) without explicit human annotations. Given a desired behavior expressed in a natural language suffix string concatenated to the input prompt, SelfControl computes gradients of the LLM's self-evaluation of the suffix with respect to its latent representations. The gradients are used to directly control the auto-regressive generation process towards desired behaviors, which eliminates human supervision, achieves precise and transparent control, and offers on-the-fly adaptability. To further enhance efficiency, we introduce SelfControl_{Prefix}, a compact module that encapsulates the learned representations from gradients into a SelfControl_{Prefix}, facilitating efficient inference-time control with no latency compared to the original model and allowing control for multiple behaviors simultaneously. Our experiments demonstrate SelfControl's efficacy across multiple domains, where it improves over SOTA for 8.3% in detoxification, 3.1% in truthfulness enhancement, 4%~10% in controlling on emotion tones, and 48.2% in privacy protection, i.e., completely remove privacy leakage issue. Additionally, we demonstrate that SelfControl can be used for data synthesis and to improve reasoning abilities.

Hongqiu Wu, Shaohua Zhang, Yuchen Zhang et al.

In this paper, we study Chinese Spelling Correction (CSC) as a joint decision made by two separate models: a language model and an error model. Through empirical analysis, we find that fine-tuning BERT tends to over-fit the error model while under-fit the language model, resulting in poor generalization to out-of-distribution error patterns. Given that BERT is the backbone of most CSC models, this phenomenon has a significant negative impact. To address this issue, we are releasing a multi-domain benchmark LEMON, with higher quality and diversity than existing benchmarks, to allow a comprehensive assessment of the open domain generalization of CSC models. Then, we demonstrate that a very simple strategy, randomly masking 20\% non-error tokens from the input sequence during fine-tuning is sufficient for learning a much better language model without sacrificing the error model. This technique can be applied to any model architecture and achieves new state-of-the-art results on SIGHAN, ECSpell, and LEMON.

Zhen Huang, Xiaodan Zhuang, Daben Liu et al.

In this paper, we present our initial efforts for building a code-switching (CS) speech recognition system leveraging existing acoustic models (AMs) and language models (LMs), i.e., no training required, and specifically targeting intra-sentential switching. To achieve such an ambitious goal, new mechanisms for foreign pronunciation generation and language model (LM) enrichment have been devised. Specifically, we have designed an automatic approach to obtain high quality pronunciation of foreign language (FL) words in the native language (NL) phoneme set using existing acoustic phone decoders and an LSTM-based grapheme-to-phoneme (G2P) model. Improved accented pronunciations have thus been obtained by learning foreign pronunciations directly from data. Furthermore, a code-switching LM was deployed by converting the original NL LM into a CS LM using translated word pairs and borrowing statistics for the NL LM. Experimental evidence clearly demonstrates that our approach better deals with accented foreign pronunciations than techniques based on human labeling. Moreover, our best system achieves a 55.5% relative word error rate reduction from 34.4%, obtained with a conventional monolingual ASR system, to 15.3% on an intra-sentential CS task without harming the monolingual recognition accuracy.

Yuchen Zhang, Mingsheng Long, Jianmin Wang et al.

We propose the discrepancy-based generalization theories for unsupervised domain adaptation. Previous theories introduced distribution discrepancies defined as the supremum over complete hypothesis space. The hypothesis space may contain hypotheses that lead to unnecessary overestimation of the risk bound. This paper studies the localized discrepancies defined on the hypothesis space after localization. First, we show that these discrepancies have desirable properties. They could be significantly smaller than the pervious discrepancies. Their values will be different if we exchange the two domains, thus can reveal asymmetric transfer difficulties. Next, we derive improved generalization bounds with these discrepancies. We show that the discrepancies could influence the rate of the sample complexity. Finally, we further extend the localized discrepancies for achieving super transfer and derive generalization bounds that could be even more sample-efficient on source domain.

Yuchen Zhang, Tianle Liu, Mingsheng Long et al.

This paper addresses the problem of unsupervised domain adaption from theoretical and algorithmic perspectives. Existing domain adaptation theories naturally imply minimax optimization algorithms, which connect well with the domain adaptation methods based on adversarial learning. However, several disconnections still exist and form the gap between theory and algorithm. We extend previous theories (Mansour et al., 2009c; Ben-David et al., 2010) to multiclass classification in domain adaptation, where classifiers based on the scoring functions and margin loss are standard choices in algorithm design. We introduce Margin Disparity Discrepancy, a novel measurement with rigorous generalization bounds, tailored to the distribution comparison with the asymmetric margin loss, and to the minimax optimization for easier training. Our theory can be seamlessly transformed into an adversarial learning algorithm for domain adaptation, successfully bridging the gap between theory and algorithm. A series of empirical studies show that our algorithm achieves the state of the art accuracies on challenging domain adaptation tasks.

Yuchen Zhang, Nianwen Xue

We design and build the first neural temporal dependency parser. It utilizes a neural ranking model with minimal feature engineering, and parses time expressions and events in a text into a temporal dependency tree structure. We evaluate our parser on two domains: news reports and narrative stories. In a parsing-only evaluation setup where gold time expressions and events are provided, our parser reaches 0.81 and 0.70 f-score on unlabeled and labeled parsing respectively, a result that is very competitive against alternative approaches. In an end-to-end evaluation setup where time expressions and events are automatically recognized, our parser beats two strong baselines on both data domains. Our experimental results and discussions shed light on the nature of temporal dependency structures in different domains and provide insights that we believe will be valuable to future research in this area.

Yuchen Zhang, Nianwen Xue

Temporal relations between events and time expressions in a document are often modeled in an unstructured manner where relations between individual pairs of time expressions and events are considered in isolation. This often results in inconsistent and incomplete annotation and computational modeling. We propose a novel annotation approach where events and time expressions in a document form a dependency tree in which each dependency relation corresponds to an instance of temporal anaphora where the antecedent is the parent and the anaphor is the child. We annotate a corpus of 235 documents using this approach in the two genres of news and narratives, with 48 documents doubly annotated. We report a stable and high inter-annotator agreement on the doubly annotated subset, validating our approach, and perform a quantitative comparison between the two genres of the entire corpus. We make this corpus publicly available.

Yuchen Zhang, Percy Liang, Martin J. Wainwright

We describe the class of convexified convolutional neural networks (CCNNs), which capture the parameter sharing of convolutional neural networks in a convex manner. By representing the nonlinear convolutional filters as vectors in a reproducing kernel Hilbert space, the CNN parameters can be represented as a low-rank matrix, which can be relaxed to obtain a convex optimization problem. For learning two-layer convolutional neural networks, we prove that the generalization error obtained by a convexified CNN converges to that of the best possible CNN. For learning deeper networks, we train CCNNs in a layer-wise manner. Empirically, CCNNs achieve performance competitive with CNNs trained by backpropagation, SVMs, fully-connected neural networks, stacked denoising auto-encoders, and other baseline methods.