Pairwise evaluation of Large Language Models (LLMs) is a common paradigm, but it is prone to preference bias, where judges systematically favor certain outputs, such as their own. This bias leads to inconsistent and skewed rankings across different judges. To address this, we first empirically demonstrate significant and heterogeneous biases in cross-model evaluations. We then propose UDA (Unsupervised Debiasing Alignment), a framework that reduces inter-judge disagreement by dynamically adjusting the Elo rating system. For each pairwise comparison, a compact neural network learns to adaptively set the K-factor and refine win probabilities. Crucially, UDA operates in a fully unsupervised manner, guided solely by the objective of minimizing the dispersion among the Elo trajectories of all judges. This forces an alignment towards a collective consensus, which serves as an unsupervised proxy for a more stable and reproducible evaluation. In addition, we provide theoretical motivation demonstrating how alignment towards a consensus can reduce aggregate system bias. Experiments show that UDA significantly reduces the inter-judge rating standard deviation by up to 63.4% and improves the average correlation with human judgments by 24.7%. Notably, UDA elevates the performance of poorly performing judges to achieve parity with high-quality ones, fostering a more robust and reliable evaluation ecosystem.

Retrieval-Augmented Generation (RAG) systems are increasingly central to robust AI, enhancing large language model (LLM) faithfulness by incorporating external knowledge. However, our study unveils a critical, overlooked vulnerability: their profound susceptibility to subtle symbolic perturbations, particularly through near-imperceptible emotional icons (e.g., "(@_@)") that can catastrophically mislead retrieval, termed EmoRAG. We demonstrate that injecting a single emoticon into a query makes it nearly 100% likely to retrieve semantically unrelated texts, which contain a matching emoticon. Our extensive experiment across general question-answering and code domains, using a range of state-of-the-art retrievers and generators, reveals three key findings: (I) Single-Emoticon Disaster: Minimal emoticon injections cause maximal disruptions, with a single emoticon almost 100% dominating RAG output. (II) Positional Sensitivity: Placing an emoticon at the beginning of a query can cause severe perturbation, with F1-Scores exceeding 0.92 across all datasets. (III) Parameter-Scale Vulnerability: Counterintuitively, models with larger parameters exhibit greater vulnerability to the interference. We provide an in-depth analysis to uncover the underlying mechanisms of these phenomena. Furthermore, we raise a critical concern regarding the robustness assumption of current RAG systems, envisioning a threat scenario where an adversary exploits this vulnerability to manipulate the RAG system. We evaluate standard defenses and find them insufficient against EmoRAG. To address this, we propose targeted defenses, analyzing their strengths and limitations in mitigating emoticon-based perturbations. Finally, we outline future directions for building robust RAG systems.

Large Language Models excel at code generation yet struggle with complex programming tasks that demand sophisticated reasoning. To bridge this gap, traditional process supervision relies on learned reward models requiring costly training data and suffering from reward misalignment, while outcome supervision fails for complex tasks needing coordinated intermediate steps. We introduce Outcome Refining Process Supervision, which unifies process and outcome supervision by leveraging executable verification: a tree-structured search framework generates strategic alternatives, profiles execution metrics, and scores candidates via self-critique mechanisms that integrate runtime feedback with reasoning. Experiments across 5 models and 3 benchmarks show consistent gains, with 26.9% higher correctness and 42.2% improved code efficiency. The results demonstrate that ORPS enables LLMs to overcome local optima in code generation, suggesting a promising direction for combining verifiable outcomes with structured reasoning to tackle complex challenges.

Recent advances in latent diffusion models have demonstrated their effectiveness for high-resolution image synthesis. However, the properties of the latent space from tokenizer for better learning and generation of diffusion models remain under-explored. Theoretically and empirically, we find that improved generation quality is closely tied to the latent distributions with better structure, such as the ones with fewer Gaussian Mixture modes and more discriminative features. Motivated by these insights, we propose MAETok, an autoencoder (AE) leveraging mask modeling to learn semantically rich latent space while maintaining reconstruction fidelity. Extensive experiments validate our analysis, demonstrating that the variational form of autoencoders is not necessary, and a discriminative latent space from AE alone enables state-of-the-art performance on ImageNet generation using only 128 tokens. MAETok achieves significant practical improvements, enabling a gFID of 1.69 with 76× faster training and 31× higher inference throughput for 512×512 generation. Our findings show that the structure of the latent space, rather than variational constraints, is crucial for effective diffusion models. Code and trained models will be released.

Large Vision-Language Models (LVLMs) have achieved significant success in multimodal tasks, with multimodal chain-of-thought (MCoT) further enhancing performance and interpretability. Recent MCoT methods fall into two categories: (i) Textual-MCoT (T-MCoT), which takes multimodal input and produces textual output; and (ii) Interleaved-MCoT (I-MCoT), which generates interleaved image-text outputs. Despite advances in both approaches, the mechanisms driving these improvements are not fully understood. To fill this gap, we first reveal that MCoT boosts LVLMs by incorporating $\textit{visual thoughts}$, which convey image information to the reasoning process regardless of the MCoT format, depending only on clarity and conciseness of expression. Furthermore, to explore visual thoughts systematically, we define four distinct forms of visual thought expressions and analyze them comprehensively. Our findings demonstrate that these forms differ in clarity and conciseness, yielding varying levels of MCoT improvement. Additionally, we explore the internal nature of visual thoughts, finding that visual thoughts serve as intermediaries between the input image and reasoning to deeper transformer layers, enabling more advanced visual information transmission. We hope that the visual thoughts can inspire further breakthroughs for future MCoT research.

Watermarking LLM-generated text is critical for content attribution and misinformation prevention, yet existing methods compromise text quality and require white-box model access with logit manipulation or training, which exclude API-based models and multilingual scenarios. We propose SAEMark, an **inference-time framework** for *multi-bit* watermarking that embeds personalized information through *feature-based rejection sampling*, fundamentally different from logit-based or rewriting-based approaches: we **do not modify model outputs directly** and require only **black-box access**, while naturally supporting multi-bit message embedding and generalizing across diverse languages and domains. We instantiate the framework using *Sparse Autoencoders* as deterministic feature extractors and provide theoretical worst-case analysis relating watermark accuracy to computational budget. Experiments across 4 datasets demonstrate strong watermarking performance on English, Chinese, and code while preserving text quality. SAEMark establishes a new paradigm for **scalable, quality-preserving watermarks** that work seamlessly with closed-source LLMs across languages and domains.

Researching genes and their interactions is crucial for deciphering the fundamental laws of cellular activity, advancing disease treatment, drug discovery, and more. Large language Models (LLMs), with their profound text comprehension and generation capabilities, have made significant strides across various natural science fields. However, their application in cell biology remains limited and a systematic evaluation of their performance is lacking. To address this gap, in this article, we select seven mainstream LLMs and evaluate their performance across nine gene-related problem scenarios. Our findings indicate that LLMs possess a certain level of understanding of genes and cells, but still lag behind domain-specific models in comprehending transcriptional expression profiles. Moreover, we have improved the current method of textual representation of cells, enhancing the LLMs’ ability to tackle cell annotation tasks. We encourage cell biology researchers to leverage LLMs for problem-solving while being mindful of the associated challenges. We release our code and data at https://github.com/epang-ucas/Evaluate_LLMs_to_Genes .

Large Language Models (LLMs) have showcased exceptional capabilities in various domains, attracting significant interest from both academia and industry. Despite their impressive performance, the substantial size and computational demands of LLMs pose considerable challenges for practical deployment, particularly in environments with limited resources. The endeavor to compress language models while maintaining their accuracy has become a focal point of research. Among the various methods, knowledge distillation has emerged as an effective technique to enhance inference speed without greatly compromising performance. This article presents a thorough survey from three aspects: method, evaluation, and application, exploring knowledge distillation techniques tailored specifically for LLMs. Specifically, we divide the methods into white-box KD and black-box KD to better illustrate their differences. Furthermore, we also explored the evaluation tasks and distillation effects between different distillation methods and proposed directions for future research. Through in-depth understanding of the latest advancements and practical applications, this survey provides valuable resources for researchers, paving the way for sustained progress in this field.

Zhuohao Yu, Chang Gao, Wenjin Yao, Yidong Wang, Wei Ye, Jindong Wang, Xing Xie, Yue Zhang, Shikun Zhang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Xiaoling Zhou, Wei Ye, Yidong Wang, Chaoya Jiang, Zhemg Lee, Rui Xie, Shikun Zhang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Zero-Shot Hashing (ZSH) has aroused significant attention due to its efficiency and generalizability in multi-modal retrieval scenarios, which aims to encode semantic information into hash codes without needing unseen labeled training samples. In addition to commonly used visual images as visual semantics and class labels as global semantics, the corresponding attribute descriptions contain critical local semantics with detailed information. However, most existing methods focus on leveraging the extracted attribute numerical values, without exploring the textual semantics in attribute descriptions. To bridge this gap, in this paper, we propose Prompt-based zero-shot hashing via vIsual and teXtual sEmantic aLignment, namely PIXEL. Concretely, we design the attribute prompt template depending on attribute descriptions to make the model capture the corresponding local semantics. Then, achieving the textual embedding and visual embedding, we proposed an alignment module to model the intra- and inter-class contrastive distances. In addition, the attribute-wise constraint and class-wise constraint are utilized to collaboratively learn the hash code, image representation, and visual attributes more effectively. Finally, extensive experimental results demonstrate the superiority of PIXEL.

Instruction tuning large language models (LLMs) remains a challenging task, owing to the complexity of hyperparameter selection and the difficulty involved in evaluating the tuned models. To determine the optimal hyperparameters, an automatic, robust, and reliable evaluation benchmark is essential. However, establishing such a benchmark is not a trivial task due to the challenges associated with evaluation accuracy and privacy protection. In response to these challenges, we introduce a judge large language model, named PandaLM, which is trained to distinguish the superior model given several LLMs. PandaLM's focus extends beyond just the objective correctness of responses, which is the main focus of traditional evaluation datasets. It addresses vital subjective factors such as relative conciseness, clarity, adherence to instructions, comprehensiveness, and formality. To ensure the reliability of PandaLM, we collect a diverse human-annotated test dataset, where all contexts are generated by humans and labels are aligned with human preferences. Our findings reveal that PandaLM-7B offers a performance comparable to both GPT-3.5 and GPT-4. Impressively, PandaLM-70B surpasses their performance. PandaLM enables the evaluation of LLM to be fairer but with less cost, evidenced by significant improvements achieved by models tuned through PandaLM compared to their counterparts trained with default Alpaca's hyperparameters. In addition, PandaLM does not depend on API-based evaluations, thus avoiding potential data leakage.

Large Language Models (LLMs) exhibit emerging in-context learning abilities through prompt engineering. The recent progress in large-scale generative models has further expanded their use in real-world language applications. However, the critical challenge of improving the generalizability and factuality of LLMs in natural language understanding and question answering remains under-explored. While previous in-context learning research has focused on enhancing models to adhere to users' specific instructions and quality expectations, and to avoid undesired outputs, little to no work has explored the use of task-specific fine-tuned Language Models (SLMs) to improve LLMs' in-context learning during the inference stage. Our primary contribution is the establishment of a simple yet effective framework that enhances the reliability of LLMs as it: 1) generalizes out-of-distribution data, 2) elucidates how LLMs benefit from discriminative models, and 3) minimizes hallucinations in generative tasks. Using our proposed plug-in method, enhanced versions of Llama 2 and ChatGPT surpass their original versions regarding generalizability and factuality. We offer a comprehensive suite of resources, including 16 curated datasets, prompts, model checkpoints, and LLM outputs across 9 distinct tasks. Our empirical analysis sheds light on the advantages of incorporating discriminative models into LLMs and highlights the potential of our methodology in fostering more reliable LLMs.

Weakly supervised learning generally faces challenges in applicability to various scenarios with diverse weak supervision and in scalability due to the complexity of existing algorithms, thereby hindering the practical deployment. This paper introduces a general framework for learning from weak supervision (GLWS) with a novel algorithm. Central to GLWS is an Expectation-Maximization (EM) formulation, adeptly accommodating various weak supervision sources, including instance partial labels, aggregate statistics, pairwise observations, and unlabeled data. We further present an advanced algorithm that significantly simplifies the EM computational demands using a Non-deterministic Finite Automaton (NFA) along with a forward-backward algorithm, which effectively reduces time complexity from quadratic or factorial often required in existing solutions to linear scale. The problem of learning from arbitrary weak supervision is therefore converted to the NFA modeling of them. GLWS not only enhances the scalability of machine learning models but also demonstrates superior performance and versatility across 11 weak supervision scenarios. We hope our work paves the way for further advancements and practical deployment in this field.

In the evolving landscape of large language models (LLMs) tailored for software engineering, the need for benchmarks that accurately reflect real-world development scenarios is paramount. Current benchmarks are either too simplistic or fail to capture the multi-tasking nature of software development. To address this, we introduce CoderUJB, a new benchmark designed to evaluate LLMs across diverse Java programming tasks that are executable and reflective of actual development scenarios, acknowledging Java's prevalence in real-world software production. CoderUJB comprises 2,239 programming questions derived from 17 real open-source Java projects and spans five practical programming tasks. Our empirical study on this benchmark investigates the coding abilities of various open-source and closed-source LLMs, examining the effects of continued pre-training in specific programming languages code and instruction fine-tuning on their performance. The findings indicate that while LLMs exhibit strong potential, challenges remain, particularly in non-functional code generation (e.g., test generation and defect detection). Importantly, our results advise caution in the specific programming languages continued pre-training and instruction fine-tuning, as these techniques could hinder model performance on certain tasks, suggesting the need for more nuanced strategies. CoderUJB thus marks a significant step towards more realistic evaluations of programming capabilities in LLMs, and our study provides valuable insights for the future development of these models in software engineering.

Learning with reduced labeling standards, such as noisy label, partial label, and supplementary unlabeled data, which we generically refer to as imprecise label, is a commonplace challenge in machine learning tasks. Previous methods tend to propose specific designs for every emerging imprecise label configuration, which is usually unsustainable when multiple configurations of imprecision coexist. In this paper, we introduce imprecise label learning (ILL), a framework for the unification of learning with various imprecise label configurations. ILL leverages expectation-maximization (EM) for modeling the imprecise label information, treating the precise labels as latent variables. Instead of approximating the correct labels for training, it considers the entire distribution of all possible labeling entailed by the imprecise information. We demonstrate that ILL can seamlessly adapt to partial label learning, semi-supervised learning, noisy label learning, and, more importantly, a mixture of these settings, with closed-form learning objectives derived from the unified EM modeling. Notably, ILL surpasses the existing specified techniques for handling imprecise labels, marking the first practical and unified framework with robust and effective performance across various challenging settings. We hope our work will inspire further research on this topic, unleashing the full potential of ILL in wider scenarios where precise labels are expensive and complicated to obtain.

This paper introduces AutoSurvey, a speedy and well-organized methodology for automating the creation of comprehensive literature surveys in rapidly evolving fields like artificial intelligence. Traditional survey paper creation faces challenges due to the vast volume and complexity of information, prompting the need for efficient survey methods. While large language models (LLMs) offer promise in automating this process, challenges such as context window limitations, parametric knowledge constraints, and the lack of evaluation benchmarks remain. AutoSurvey addresses these challenges through a systematic approach that involves initial retrieval and outline generation, subsection drafting by specialized LLMs, integration and refinement, and rigorous evaluation and iteration. Our contributions include a comprehensive solution to the survey problem, a reliable evaluation method, and experimental validation demonstrating AutoSurvey's effectiveness.

The distribution shifts between training and test data typically undermine the performance of models. In recent years, lots of work pays attention to domain generalization (DG) where distribution shifts exist and target data are unseen. Despite the progress in algorithm design, two foundational factors have long been ignored: 1) the optimization for regularization-based objectives, and 2) the model selection for DG since no knowledge about the target domain can be utilized. In this paper, we propose Mixup guided optimization and selection techniques for DG. For optimization, we utilize an adapted Mixup to generate an out-of-distribution dataset that can guide the preference direction and optimize with Pareto optimization. For model selection, we generate a validation dataset with a closer distance to the target distribution, and thereby it can better represent the target data. We also present some theoretical insights behind our proposals. Comprehensive experiments demonstrate that our model optimization and selection techniques can largely improve the performance of existing domain generalization algorithms and even achieve new state-of-the-art results.

Large language models (LLMs) are gaining increasing popularity in both academia and industry, owing to their unprecedented performance in various applications. As LLMs continue to play a vital role in both research and daily use, their evaluation becomes increasingly critical, not only at the task level, but also at the society level for better understanding of their potential risks. Over the past years, significant efforts have been made to examine LLMs from various perspectives. This paper presents a comprehensive review of these evaluation methods for LLMs, focusing on three key dimensions: what to evaluate , where to evaluate , and how to evaluate . Firstly, we provide an overview from the perspective of evaluation tasks, encompassing general natural language processing tasks, reasoning, medical usage, ethics, education, natural and social sciences, agent applications, and other areas. Secondly, we answer the ‘where’ and ‘how’ questions by diving into the evaluation methods and benchmarks, which serve as crucial components in assessing the performance of LLMs. Then, we summarize the success and failure cases of LLMs in different tasks. Finally, we shed light on several future challenges that lie ahead in LLMs evaluation. Our aim is to offer invaluable insights to researchers in the realm of LLMs evaluation, thereby aiding the development of more proficient LLMs. Our key point is that evaluation should be treated as an essential discipline to better assist the development of LLMs. We consistently maintain the related open-source materials at: https://github.com/MLGroupJLU/LLM-eval-survey

Federated learning (FL) is important for privacy-preserving services by training models without collecting raw user data. Most FL algorithms assume all data is annotated, which is impractical due to the high cost of labeling data in real applications. To alleviate the reliance on labeled data, semi-supervised federated learning (SSFL) has been proposed to utilize unlabeled data on clients to improve model performance. However, most existing methods either have privacy issues which share models trained on other clients, or generate pseudo-labels for unlabeled local datasets with the global model, which is usually biased towards the global data distribution. The latter may lead to sub-optimal accuracy of pseudo-labels, due to the gap between the local data distribution and the global model, especially in non-IID settings. In this paper, we propose a semi-supervised heterogeneous federated learning method with local knowledge enhancement, called FedLoKe, which aims to train an accurate global model from both labeled and unlabeled local data with non-IID distributions. Specifically, in FedLoKe, the server maintains a global model to capture global data distribution, and each client learns a local model to capture local data distribution. Since the distribution captured by the local model is aligned with the local data distribution, we utilize it to generate high-accuracy pseudo-labels of the unlabeled dataset for global model training. To prevent the local model from severely overfitting the small number of local labeled data, we further use the exponential moving average and apply the global model to generate pseudo-labels for local modeling training. Experiments on four datasets show the effectiveness of FedLoKe. Our code is available at: https://github.com/zcfinal/FedLoKe.

This study focuses on the evaluation of the Open Question Answering (Open-QA) task, which can directly estimate the factuality of large language models (LLMs). Current automatic evaluation methods have shown limitations, indicating that human evaluation still remains the most reliable approach. We introduce a new task, QA Evaluation (QA-Eval) and the corresponding dataset EVOUNA, designed to assess the accuracy of AI-generated answers in relation to standard answers within Open-QA. Our evaluation of these methods utilizes human-annotated results to measure their performance. Specifically, the work investigates methods that show high correlation with human evaluations, deeming them more reliable. We also discuss the pitfalls of current methods and methods to improve LLM-based evaluators. We believe this new QA-Eval task and corresponding dataset EVOUNA will facilitate the development of more effective automatic evaluation tools and prove valuable for future research in this area. All resources are available at https://github.com/wangcunxiang/QA-Eval and it is under the Apache-2.0 License.

Target-oriented Opinion Words Extraction (TOWE) is a fine-grained sentiment analysis task that aims to extract the corresponding opinion words of a given opinion target from the sentence. Recently, deep learning approaches have made remarkable progress on this task. Nevertheless, the TOWE task still suffers from the scarcity of training data due to the expensive data annotation process. Limited labeled data increase the risk of distribution shift between test data and training data. In this paper, we propose exploiting massive unlabeled data to reduce the risk by increasing the exposure of the model to varying distribution shifts. Specifically, we propose a novel Multi-Grained Consistency Regularization (MGCR) method to make use of unlabeled data and design two filters specifically for TOWE to filter noisy data at different granularity. Extensive experimental results on four TOWE benchmark datasets indicate the superiority of MGCR compared with current state-of-the-art methods. The in-depth analysis also demonstrates the effectiveness of the different-granularity filters.

Semi-supervised learning (SSL) improves model generalization by leveraging massive unlabeled data to augment limited labeled samples. However, currently, popular SSL evaluation protocols are often constrained to computer vision (CV) tasks. In addition, previous work typically trains deep neural networks from scratch, which is time-consuming and environmentally unfriendly. To address the above issues, we construct a Unified SSL Benchmark (USB) for classification by selecting 15 diverse, challenging, and comprehensive tasks from CV, natural language processing (NLP), and audio processing (Audio), on which we systematically evaluate the dominant SSL methods, and also open-source a modular and extensible codebase for fair evaluation of these SSL methods. We further provide the pre-trained versions of the state-of-the-art neural models for CV tasks to make the cost affordable for further tuning. USB enables the evaluation of a single SSL algorithm on more tasks from multiple domains but with less cost. Specifically, on a single NVIDIA V100, only 39 GPU days are required to evaluate FixMatch on 15 tasks in USB while 335 GPU days (279 GPU days on 4 CV datasets except for ImageNet) are needed on 5 CV tasks with TorchSSL.

Cross-lingual speech adaptation aims to solve the problem of leveraging multiple rich-resource languages to build models for a low-resource target language. Since the low-resource language has limited training data, speech recognition models can easily overfit. Adapter is a versatile module that can be plugged into Transformer for parameter-efficient learning. In this paper, we propose to use adapters for parameter-efficient cross-lingual speech adaptation. Based on our previous MetaAdapter that implicitly leverages adapters, we propose a novel algorithm called SimAdapter for explicitly learning knowledge from adapters. Our algorithms can be easily integrated into the Transformer structure. MetaAdapter leverages meta-learning to transfer the general knowledge from training data to the test language. SimAdapter aims to learn the similarities between the source and target languages during fine-tuning using the adapters. We conduct extensive experiments on five-low-resource languages in the Common Voice dataset. Results demonstrate that MetaAdapter and SimAdapter can reduce WER by 2.98% and 2.55% with only 2.5% and 15.5% of trainable parameters compared to the strong full-model fine-tuning baseline. Moreover, we show that these two novel algorithms can be integrated for better performance with up to 3.55% relative WER reduction.

Transfer learning from a multilingual model has shown favorable results on low-resource automatic speech recognition (ASR). However, full-model fine-tuning generates a separate model for every target language and is not suitable for deploying and maintaining in production. The key challenge lies in how to efficiently extend the pre-trained model with fewer parameters. In this paper, we propose to combine the adapter module with meta-learning algorithms to achieve high recognition performance under low-resource settings and improve the parameter-efficiency of the model. Extensive experiments show that our methods can achieve comparable or even superior recognition rates than the state-of-the-art baselines on low-resource languages, especially under very-low-resource conditions, with a significantly smaller model profile.

The recently proposed FixMatch achieved state-of-the-art results on most semi-supervised learning (SSL) benchmarks. However, like other modern SSL algorithms, FixMatch uses a pre-defined constant threshold for all classes to select unlabeled data that contribute to the training, thus failing to consider different learning status and learning difficulties of different classes. To address this issue, we propose Curriculum Pseudo Labeling (CPL), a curriculum learning approach to leverage unlabeled data according to the model's learning status. The core of CPL is to flexibly adjust thresholds for different classes at each time step to let pass informative unlabeled data and their pseudo labels. CPL does not introduce additional parameters or computations (forward or backward propagation). We apply CPL to FixMatch and call our improved algorithm FlexMatch. FlexMatch achieves state-of-the-art performance on a variety of SSL benchmarks, with especially strong performances when the labeled data are extremely limited or when the task is challenging. For example, FlexMatch achieves 13.96% and 18.96% error rate reduction over FixMatch on CIFAR-100 and STL-10 datasets respectively, when there are only 4 labels per class. CPL also significantly boosts the convergence speed, e.g., FlexMatch can use only 1/5 training time of FixMatch to achieve even better performance. Furthermore, we show that CPL can be easily adapted to other SSL algorithms and remarkably improve their performances. We open-source our code at https://github.com/TorchSSL/TorchSSL.