As Large Language Models (LLMs) become increasingly integrated into various facets of society, a significant portion of online text consequently become synthetic. This raises concerns about bias amplification, a phenomenon where models trained on synthetic data amplify the pre-existing biases over successive training iterations. Previous literature seldom discusses bias amplification as an independent issue from model collapse. In this work, we address the gap in understanding the bias amplification of LLMs with four main contributions. Firstly, we propose a theoretical framework, defining the necessary and sufficient conditions for its occurrence, and emphasizing that it occurs independently of model collapse. Using statistical simulations with weighted maximum likelihood estimation, we demonstrate the framework and show how bias amplification arises without the sampling and functional form issues that typically drive model collapse. Secondly, we conduct experiments with GPT-2 to empirically demonstrate bias amplification, specifically examining open-ended generational political bias with a benchmark we developed. We observe that GPT-2 exhibits a right-leaning bias in sentence continuation tasks and that the bias progressively increases with iterative fine-tuning on synthetic data generated by previous iterations. Thirdly, we explore three potential mitigation strategies: Overfitting, Preservation, and Accumulation. We find that both Preservation and Accumulation effectively mitigate bias amplification and model collapse. Finally, using novel mechanistic interpretation techniques, we demonstrate that in the GPT-2 experiments, bias amplification and model collapse are driven by distinct sets of neurons, which aligns with our theoretical framework.

The ability for individuals to constructively engage with one another across lines of difference is a critical feature of a healthy pluralistic society. This is also true in online discussion spaces like social media platforms. To date, much social media research has focused on preventing ills -- like political polarization and the spread of misinformation. While this is important, enhancing the quality of online public discourse requires not just reducing ills but also promoting foundational human virtues. In this study, we focus on one particular virtue: ``intellectual humility'' (IH), or acknowledging the potential limitations in one's own beliefs. Specifically, we explore the development of computational methods for measuring IH at scale. We manually curate and validate an IH codebook on 350 posts about religion drawn from subreddits and use them to develop LLM-based models for automating this measurement. Our best model achieves a Macro-F1 score of 0.64 across labels (and 0.70 when predicting IH/IA/Neutral at the coarse level), higher than an expected naive baseline of 0.51 (0.32 for IH/IA/Neutral) but lower than a human annotator-informed upper bound of 0.85 (0.83 for IH/IA/Neutral). Our results both highlight the challenging nature of detecting IH online -- opening the door to new directions in NLP research -- and also lay a foundation for computational social science researchers interested in analyzing and fostering more IH in online public discourse.

Users of the transit system flood social networks daily with messages that contain valuable insights crucial for improving service quality. These posts help transit agencies quickly identify emerging issues. Parsing topics and sentiments is key to gaining comprehensive insights to foster service excellence. However, the volume of messages makes manual analysis impractical, and standard NLP techniques like Term Frequency-Inverse Document Frequency (TF-IDF) fall short in nuanced interpretation. Traditional sentiment analysis separates topics and sentiments before integrating them, often missing the interaction between them. This incremental approach complicates classification and reduces analytical productivity. To address these challenges, we propose a novel approach to extracting and analyzing transit-related information, including sentiment and sarcasm detection, identification of unusual system problems, and location data from social media. Our method employs Large Language Models (LLM), specifically Llama 3, for a streamlined analysis free from pre-established topic labels. To enhance the model's domain-specific knowledge, we utilize Retrieval-Augmented Generation (RAG), integrating external knowledge sources into the information extraction pipeline. We validated our method through extensive experiments comparing its performance with traditional NLP approaches on user tweet data from the real world transit system. Our results demonstrate the potential of LLMs to transform social media data analysis in the public transit domain, providing actionable insights and enhancing transit agencies' responsiveness by extracting a broader range of information.

Amyotrophic lateral sclerosis (ALS), a progressive neuromuscular degenerative disease, severely restricts patient communication capacity within a few years of onset, resulting in a significant deterioration of quality of life. The P300 speller brain computer interface (BCI) offers an alternative communication medium by leveraging a subject's EEG response to characters traditionally highlighted on a character grid on a graphical user interface (GUI). A recurring theme in P300-based research is enhancing performance to enable faster subject interaction. This study builds on that theme by addressing key limitations, particularly in the training of multi-subject classifiers, and by integrating advanced language models to optimize stimuli presentation and word prediction, thereby improving communication efficiency. Furthermore, various advanced large language models such as Generative Pre-Trained Transformer (GPT2), BERT, and BART, alongside Dijkstra's algorithm, are utilized to optimize stimuli and provide word completion choices based on the spelling history. In addition, a multi-layered smoothing approach is applied to allow for out-of-vocabulary (OOV) words. By conducting extensive simulations based on randomly sampled EEG data from subjects, we show substantial speed improvements in typing passages that include rare and out-of-vocabulary (OOV) words, with the extent of improvement varying depending on the language model utilized. The gains through such character-level interface optimizations are approximately 10%, and GPT2 for multi-word prediction provides gains of around 40%. In particular, some large language models achieve performance levels within 10% of the theoretical performance limits established in this study. In addition, both within and across subjects, training techniques are explored, and speed improvements are shown to hold in both cases.

In recent years, Graph Neural Networks (GNNs) have become successful in molecular property prediction tasks such as toxicity analysis. However, due to the black-box nature of GNNs, their outputs can be concerning in high-stakes decision-making scenarios, e.g., drug discovery. Facing such an issue, Graph Counterfactual Explanation (GCE) has emerged as a promising approach to improve GNN transparency. However, current GCE methods usually fail to take domain-specific knowledge into consideration, which can result in outputs that are not easily comprehensible by humans. To address this challenge, we propose a novel GCE method, LLM-GCE, to unleash the power of large language models (LLMs) in explaining GNNs for molecular property prediction. Specifically, we utilize an autoencoder to generate the counterfactual graph topology from a set of counterfactual text pairs (CTPs) based on an input graph. Meanwhile, we also incorporate a CTP dynamic feedback module to mitigate LLM hallucination, which provides intermediate feedback derived from the generated counterfactuals as an attempt to give more faithful guidance. Extensive experiments demonstrate the superior performance of LLM-GCE. Our code is released on https://github.com/YinhanHe123/new\_LLM4GNNExplanation.

In retrieval-augmented systems, context ranking techniques are commonly employed to reorder the retrieved contexts based on their relevance to a user query. A standard approach is to measure this relevance through the similarity between contexts and queries in the embedding space. However, such similarity often fails to capture the relevance. Alternatively, large language models (LLMs) have been used for ranking contexts. However, they can encounter scalability issues when the number of candidate contexts grows and the context window sizes of the LLMs remain constrained. Additionally, these approaches require fine-tuning LLMs with domain-specific data. In this work, we introduce a scalable ranking framework that combines embedding similarity and LLM capabilities without requiring LLM fine-tuning. Our framework uses a pre-trained LLM to hypothesize the user query based on the retrieved contexts and ranks the context based on the similarity between the hypothesized queries and the user query. Our framework is efficient at inference time and is compatible with many other retrieval and ranking techniques. Experimental results show that our method improves the ranking performance across multiple benchmarks. The complete code and data are available at https://github.com/zwc662/hyqe

Topic modeling is a powerful technique for uncovering hidden themes within a collection of documents. However, the effectiveness of traditional topic models often relies on sufficient word co-occurrence, which is lacking in short texts. Therefore, existing approaches, whether probabilistic or neural, frequently struggle to extract meaningful patterns from such data, resulting in incoherent topics. To address this challenge, we propose a novel approach that leverages large language models (LLMs) to extend short texts into more detailed sequences before applying topic modeling. To further improve the efficiency and solve the problem of semantic inconsistency from LLM-generated texts, we propose to use prefix tuning to train a smaller language model coupled with a variational autoencoder for short-text topic modeling. Our method significantly improves short-text topic modeling performance, as demonstrated by extensive experiments on real-world datasets with extreme data sparsity, outperforming current state-of-the-art topic models.

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 $\alpha$-DPO, an adaptive preference optimization algorithm designed to address these limitations by introducing a dynamic reward margin. Specifically, $\alpha$-DPO employs an adaptive preference distribution, balancing the policy model and the reference model to achieve personalized reward margins. We provide theoretical guarantees for $\alpha$-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 $\alpha$-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

Explanation generation plays a more pivotal role than fact verification in producing interpretable results and facilitating comprehensive fact-checking, which has recently garnered considerable attention. However, previous studies on explanation generation has shown several limitations, such as being confined to English scenarios, involving overly complex inference processes, and not fully unleashing the potential of the mutual feedback between veracity labels and explanation texts. To address these issues, we construct two complex fact-checking datasets in the Chinese scenarios: CHEF-EG and TrendFact. These datasets involve complex facts in areas such as health, politics, and society, presenting significant challenges for fact verification methods. In response to these challenges, we propose a unified framework called FactISR (Augmenting Fact-Checking via Iterative Self-Revision) to perform mutual feedback between veracity and explanations by leveraging the capabilities of large language models(LLMs). FactISR uses a single model to address tasks such as fact verification and explanation generation. Its self-revision mechanism can further revision the consistency between veracity labels, explanation texts, and evidence, as well as eliminate irrelevant noise. We conducted extensive experiments with baselines and FactISR on the proposed datasets. The experimental results demonstrate the effectiveness of our method.

The thematic fit estimation task measures the compatibility between a predicate (typically a verb), an argument (typically a noun phrase), and a specific semantic role assigned to the argument. Previous state-of-the-art work has focused on modeling thematic fit through distributional or neural models of event representation, trained in a supervised fashion with indirect labels. In this work, we assess whether pre-trained autoregressive LLMs possess consistent, expressible knowledge about thematic fit. We evaluate both closed and open state-of-the-art LLMs on several psycholinguistic datasets, along three axes: (1) Reasoning Form: multi-step logical reasoning (chain-of-thought prompting) vs. simple prompting. (2) Input Form: providing context (generated sentences) vs. raw tuples . (3) Output Form: categorical vs. numeric. Our results show that chain-of-thought reasoning is more effective on datasets with self-explanatory semantic role labels, especially Location. Generated sentences helped only in few settings, and lowered results in many others. Predefined categorical (compared to numeric) output raised GPT's results across the board with few exceptions, but lowered Llama's. We saw that semantically incoherent generated sentences, which the models lack the ability to consistently filter out, hurt reasoning and overall performance too. Our GPT-powered methods set new state-of-the-art on all tested datasets.