Jiaxuan Li

Zishan Guo, Renren Jin, Chuang Liu et al.

Large language models (LLMs) have demonstrated remarkable capabilities across a broad spectrum of tasks. They have attracted significant attention and been deployed in numerous downstream applications. Nevertheless, akin to a double-edged sword, LLMs also present potential risks. They could suffer from private data leaks or yield inappropriate, harmful, or misleading content. Additionally, the rapid progress of LLMs raises concerns about the potential emergence of superintelligent systems without adequate safeguards. To effectively capitalize on LLM capacities as well as ensure their safe and beneficial development, it is critical to conduct a rigorous and comprehensive evaluation of LLMs. This survey endeavors to offer a panoramic perspective on the evaluation of LLMs. We categorize the evaluation of LLMs into three major groups: knowledge and capability evaluation, alignment evaluation and safety evaluation. In addition to the comprehensive review on the evaluation methodologies and benchmarks on these three aspects, we collate a compendium of evaluations pertaining to LLMs' performance in specialized domains, and discuss the construction of comprehensive evaluation platforms that cover LLM evaluations on capabilities, alignment, safety, and applicability. We hope that this comprehensive overview will stimulate further research interests in the evaluation of LLMs, with the ultimate goal of making evaluation serve as a cornerstone in guiding the responsible development of LLMs. We envision that this will channel their evolution into a direction that maximizes societal benefit while minimizing potential risks. A curated list of related papers has been publicly available at https://github.com/tjunlp-lab/Awesome-LLMs-Evaluation-Papers.

Jiaxuan Li, Duc Minh Vo, Hideki Nakayama

Bias mitigation in image classification has been widely researched, and existing methods have yielded notable results. However, most of these methods implicitly assume that a given image contains only one type of known or unknown bias, failing to consider the complexities of real-world biases. We introduce a more challenging scenario, agnostic biases mitigation, aiming at bias removal regardless of whether the type of bias or the number of types is unknown in the datasets. To address this difficult task, we present the Partition-and-Debias (PnD) method that uses a mixture of biases-specific experts to implicitly divide the bias space into multiple subspaces and a gating module to find a consensus among experts to achieve debiased classification. Experiments on both public and constructed benchmarks demonstrated the efficacy of the PnD. Code is available at: https://github.com/Jiaxuan-Li/PnD.

Jiaxuan Li, Yue Ning

With the spreading of hate speech on social media in recent years, automatic detection of hate speech is becoming a crucial task and has attracted attention from various communities. This task aims to recognize online posts (e.g., tweets) that contain hateful information. The peculiarities of languages in social media, such as short and poorly written content, lead to the difficulty of learning semantics and capturing discriminative features of hate speech. Previous studies have utilized additional useful resources, such as sentiment hashtags, to improve the performance of hate speech detection. Hashtags are added as input features serving either as sentiment-lexicons or extra context information. However, our close investigation shows that directly leveraging these features without considering their context may introduce noise to classifiers. In this paper, we propose a novel approach to leverage sentiment hashtags to enhance hate speech detection in a natural language inference framework. We design a novel framework SRIC that simultaneously performs two tasks: (1) semantic relation inference between online posts and sentiment hashtags, and (2) sentiment classification on these posts. The semantic relation inference aims to encourage the model to encode sentiment-indicative information into representations of online posts. We conduct extensive experiments on two real-world datasets and demonstrate the effectiveness of our proposed framework compared with state-of-the-art representation learning models.

Jiaxuan Li, Lang Yu, Allyson Ettinger

Current pre-trained language models have enabled remarkable improvements in downstream tasks, but it remains difficult to distinguish effects of statistical correlation from more systematic logical reasoning grounded on understanding of the real world. In this paper we tease these factors apart by leveraging counterfactual conditionals, which force language models to predict unusual consequences based on hypothetical propositions. We introduce a set of tests drawn from psycholinguistic experiments, as well as larger-scale controlled datasets, to probe counterfactual predictions from a variety of popular pre-trained language models. We find that models are consistently able to override real-world knowledge in counterfactual scenarios, and that this effect is more robust in case of stronger baseline world knowledge -- however, we also find that for most models this effect appears largely to be driven by simple lexical cues. When we mitigate effects of both world knowledge and lexical cues to test knowledge of linguistic nuances of counterfactuals, we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity is also non-trivially impacted by lexical associative factors.

Jiaxuan Li, Richard Futrell

The functional interpretation of language-related ERP components has been a central debate in psycholinguistics for decades. We advance an information-theoretic model of human language processing in the brain in which incoming linguistic input is processed at first shallowly and later with more depth, with these two kinds of information processing corresponding to distinct electroencephalographic signatures. Formally, we show that the information content (surprisal) of a word in context can be decomposed into two quantities: (A) shallow surprisal, which signals shallow processing difficulty for a word, and corresponds with the N400 signal; and (B) deep surprisal, which reflects the discrepancy between shallow and deep representations, and corresponds to the P600 signal and other late positivities. Both of these quantities can be estimated straightforwardly using modern NLP models. We validate our theory by successfully simulating ERP patterns elicited by a variety of linguistic manipulations in previously-reported experimental data from six experiments, with successful novel qualitative and quantitative predictions. Our theory is compatible with traditional cognitive theories assuming a `good-enough' shallow representation stage, but with a precise information-theoretic formulation. The model provides an information-theoretic model of ERP components grounded on cognitive processes, and brings us closer to a fully-specified neuro-computational model of language processing.

Jiaxuan Li, Qing Xu, Xiangjian He et al.

Masked Autoencoders (MAE) achieve self-supervised learning of image representations by randomly removing a portion of visual tokens and reconstructing the original image as a pretext task, thereby significantly enhancing pretraining efficiency and yielding excellent adaptability across downstream tasks. However, MAE and other MAE-style paradigms that adopt random masking generally require more pre-training epochs to maintain adaptability. Meanwhile, ViT in MAE suffers from inefficient parameter use due to fixed spatial resolution across layers. To overcome these limitations, we propose the Complementary Masked Autoencoders (CoMA), which employ a complementary masking strategy to ensure uniform sampling across all pixels, thereby improving effective learning of all features and enhancing the model's adaptability. Furthermore, we introduce DyViT, a hierarchical vision transformer that employs a Dynamic Multi-Window Self-Attention (DM-MSA), significantly reducing the parameters and FLOPs while improving fine-grained feature learning. Pre-trained on ImageNet-1K with CoMA, DyViT matches the downstream performance of MAE using only 12% of the pre-training epochs, demonstrating more effective learning. It also attains a 10% reduction in pre-training time per epoch, further underscoring its superior pre-training efficiency.

Jiaxuan Li, Richard Futrell

A large body of work in psycholinguistics has focused on the idea that online language comprehension can be shallow or `good enough': given constraints on time or available computation, comprehenders may form interpretations of their input that are plausible but inaccurate. However, this idea has not yet been linked with formal theories of computation under resource constraints. Here we use information theory to formulate a model of language comprehension as an optimal trade-off between accuracy and processing depth, formalized as bits of information extracted from the input, which increases with processing time. The model provides a measure of processing effort as the change in processing depth, which we link to EEG signals and reading times. We validate our theory against a large-scale dataset of garden path sentence reading times, and EEG experiments featuring N400, P600 and biphasic ERP effects. By quantifying the timecourse of language processing as it proceeds from shallow to deep, our model provides a unified framework to explain behavioral and neural signatures of language comprehension.

Songjie Xie, Youlong Wu, Jiaxuan Li et al.

As machine learning (ML) becomes more prevalent in human-centric applications, there is a growing emphasis on algorithmic fairness and privacy protection. While previous research has explored these areas as separate objectives, there is a growing recognition of the complex relationship between privacy and fairness. However, previous works have primarily focused on examining the interplay between privacy and fairness through empirical investigations, with limited attention given to theoretical exploration. This study aims to bridge this gap by introducing a theoretical framework that enables a comprehensive examination of their interrelation. We shall develop and analyze an information bottleneck (IB) based information obfuscation method with local differential privacy (LDP) for fair representation learning. In contrast to many empirical studies on fairness in ML, we show that the incorporation of LDP randomizers during the encoding process can enhance the fairness of the learned representation. Our analysis will demonstrate that the disclosure of sensitive information is constrained by the privacy budget of the LDP randomizer, thereby enabling the optimization process within the IB framework to effectively suppress sensitive information while preserving the desired utility through obfuscation. Based on the proposed method, we further develop a variational representation encoding approach that simultaneously achieves fairness and LDP. Our variational encoding approach offers practical advantages. It is trained using a non-adversarial method and does not require the introduction of any variational prior. Extensive experiments will be presented to validate our theoretical results and demonstrate the ability of our proposed approach to achieve both LDP and fairness while preserving adequate utility.

Nan Wang, Jiaxuan Li

A central question in psycholinguistics is the timing of syntax in sentence processing. Much of the existing evidence comes from violation paradigms, which conflate two separable processes - syntactic category detection and phrase structure construction - and implicitly assume that phrase structure follows category detection. In this study, we use co-registered EEG and eye-tracking data from the ZuCo corpus to disentangle these processes and test their temporal order under naturalistic reading conditions. Analyses of gaze transitions showed that readers preferentially moved between syntactic heads, suggesting that phrase structures, rather than serial word order, organize scanpaths. Bayesian network modeling further revealed that structural depth was the strongest driver of deviations from linear reading, outweighing lexical familiarity and surprisal. Finally, fixation-related potentials demonstrated that syntactic surprisal influences neural activity before word onset (-184 to -10 ms) and during early integration (48 to 300 ms). These findings extend current models of syntactic timing by showing that phrase structure construction can precede category detection and dominate lexical influences, supporting a predictive "tree-scaffolding" account of comprehension.

Jiaxuan Li, Minxi Yang, Dahua Gao et al.

Current communication technologies face limitations in terms of theoretical capacity, spectrum availability, and power resources. Pragmatic communication, leveraging terminal intelligence for selective data transmission, offers resource conservation. Existing research lacks universal intention resolution tools, limiting applicability to specific tasks. This paper proposes an image pragmatic communication framework based on a Pragmatic Agent for Communication Efficiency (PACE) using Large Language Models (LLM). In this framework, PACE sequentially performs semantic perception, intention resolution, and intention-oriented coding. To ensure the effective utilization of LLM in communication, a knowledge base is designed to supplement the necessary knowledge, dedicated prompts are introduced to facilitate understanding of pragmatic communication scenarios and task requirements, and a chain of thought is designed to assist in making reasonable trade-offs between transmission efficiency and cost. For experimental validation, this paper constructs an image pragmatic communication dataset along with corresponding evaluation standards. Simulation results indicate that the proposed method outperforms traditional and non-LLM-based pragmatic communication in terms of transmission efficiency.

Jiaxuan Li, Lang Yu, Allyson Ettinger

Current pre-trained language models have enabled remarkable improvements in downstream tasks, but it remains difficult to distinguish effects of statistical correlation from more systematic logical reasoning grounded on the understanding of real world. We tease these factors apart by leveraging counterfactual conditionals, which force language models to predict unusual consequences based on hypothetical propositions. We introduce a set of tests from psycholinguistic experiments, as well as larger-scale controlled datasets, to probe counterfactual predictions from five pre-trained language models. We find that models are consistently able to override real-world knowledge in counterfactual scenarios, and that this effect is more robust in case of stronger baseline world knowledge -- however, we also find that for most models this effect appears largely to be driven by simple lexical cues. When we mitigate effects of both world knowledge and lexical cues to test knowledge of linguistic nuances of counterfactuals, we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity is also non-trivially impacted by lexical associative factors.