Natural language processing (NLP) has seen remarkable advancements with the development of large language models (LLMs). Despite these advancements, LLMs often produce socially biased outputs. Recent studies have mainly addressed this problem by prompting LLMs to behave ethically, but this approach results in unacceptable performance degradation. In this paper, we propose a multi-objective approach within a multi-agent framework (MOMA) to mitigate social bias in LLMs without significantly compromising their performance. The key idea of MOMA involves deploying multiple agents to perform causal interventions on bias-related contents of the input questions, breaking the shortcut connection between these contents and the corresponding answers. Unlike traditional debiasing techniques leading to performance degradation, MOMA substantially reduces bias while maintaining accuracy in downstream tasks. Our experiments conducted on two datasets and two models demonstrate that MOMA reduces bias scores by up to 87.7%, with only a marginal performance degradation of up to 6.8% in the BBQ dataset. Additionally, it significantly enhances the multi-objective metric icat in the StereoSet dataset by up to 58.1%. Code will be made available at https://github.com/Cortantse/MOMA.

Large language models (LLMs) have shown remarkable effectiveness across various domains, with data augmentation methods utilizing GPT for synthetic data generation becoming prevalent. However, the quality and utility of augmented data remain questionable, and current methods lack clear metrics for evaluating data characteristics. To address these challenges, we propose ResoFilter, a novel method that integrates models, data, and tasks to refine datasets. ResoFilter leverages the fine-tuning process to obtain Data-Parameter features for data selection, offering improved interpretability by representing data characteristics through model weights. Our experiments demonstrate that ResoFilter achieves comparable results to full-scale fine-tuning using only half the data in mathematical tasks and exhibits strong generalization across different models and domains. This method provides valuable insights for constructing synthetic datasets and evaluating high-quality data, offering a promising solution for enhancing data augmentation techniques and improving training dataset quality for LLMs. For reproducibility, we will release our code and data upon acceptance.

Prior research indicates that to be able to mediate conflict, observers of disagreements between parties must be able to reliably distinguish the sources of their disagreement as stemming from differences in beliefs about what is true (causality) vs. differences in what they value (morality). In this paper, we test if OpenAI's Large Language Models GPT 3.5 and GPT 4 can perform this task and whether one or other type of disagreement proves particularly challenging for LLM's to diagnose. We replicate study 1 in Ko\c{c}ak et al. (2003), which employes a vignette design, with OpenAI's GPT 3.5 and GPT 4. We find that both LLMs have similar semantic understanding of the distinction between causal and moral codes as humans and can reliably distinguish between them. When asked to diagnose the source of disagreement in a conversation, both LLMs, compared to humans, exhibit a tendency to overestimate the extent of causal disagreement and underestimate the extent of moral disagreement in the moral misalignment condition. This tendency is especially pronounced for GPT 4 when using a proximate scale that relies on concrete language specific to an issue. GPT 3.5 does not perform as well as GPT4 or humans when using either the proximate or the distal scale. The study provides a first test of the potential for using LLMs to mediate conflict by diagnosing the root of disagreements in causal and evaluative codes.

With the adoption of retrieval-augmented generation (RAG), large language models (LLMs) are expected to ground their generation to the retrieved contexts. Yet, this is hindered by position bias of LLMs, failing to evenly attend to all contexts. Previous work has addressed this by synthesizing contexts with perturbed positions of gold segment, creating a position-diversified train set. We extend this intuition to propose consistency regularization with augmentation and distillation. First, we augment each training instance with its position perturbation to encourage consistent predictions, regardless of ordering. We also distill behaviors of this pair, although it can be counterproductive in certain RAG scenarios where the given order from the retriever is crucial for generation quality. We thus propose CORD, balancing COnsistency and Rank Distillation. CORD adaptively samples noise-controlled perturbations from an interpolation space, ensuring both consistency and respect for the rank prior. Empirical results show this balance enables CORD to outperform consistently in diverse RAG benchmarks.

Masked diffusion models (MDMs) have shown promise in language modeling, yet their scalability and effectiveness in core language tasks, such as text generation and language understanding, remain underexplored. This paper establishes the first scaling law for MDMs, demonstrating a scaling rate comparable to autoregressive models (ARMs) and a relatively small compute gap. Motivated by their scalability, we train a family of MDMs with up to 1.1 billion (B) parameters to systematically evaluate their performance against ARMs of comparable or larger sizes. Fully leveraging the probabilistic formulation of MDMs, we propose a simple yet effective unsupervised classifier-free guidance that effectively exploits large-scale unpaired data, boosting performance for conditional inference. In language understanding, the 1.1B MDM outperforms the 1.1B TinyLlama model trained on the same data across four of eight zero-shot benchmarks. Notably, it achieves competitive math reasoning ability with the 7B Llama-2 model on the GSM8K dataset. In text generation, MDMs provide a flexible trade-off compared to ARMs utilizing KV-cache: MDMs match the performance of ARMs while being 1.4 times faster or achieving higher quality than ARMs at a higher computational cost. Moreover, MDMs address challenging tasks for ARMs by effectively handling bidirectional reasoning and adapting to temporal shifts in data. Notably, a 1.1B MDM breaks the reverse curse encountered by much larger ARMs with significantly more data and computation, such as 13B Llama-2 and 175B GPT-3. Our code is available at https://github.com/ML-GSAI/SMDM. Figure 1: IsoFLOP curves plot optimal model sizes under fixed computation budgets. The optimal MDMs validation loss exhibits power-law scaling, decreasing at a rate comparable to that of ARMs. Work done during Shen Nie's internship at Sea AI Lab. Autoregressive models (ARMs) have long been regarded as the gold standard in probabilistic language modeling. However, ARMs exhibit inherent limitations, particularly in reasoning tasks that require bidirectional context understanding or handling temporal shifts in data. These shortcomings, widely recognized as the reverse curse (Berglund et al., 2023) and temporal quality degradation (Vela et al., 2022), significantly hinder their applicability in complex language modeling scenarios.

In many practical applications, large language models (LLMs) need to incorporate new knowledge not present in their pre-training data. The primary methods for this are fine-tuning and retrieval-augmented generation (RAG). Although RAG has emerged as the industry standard for knowledge injection, fine-tuning has not yet achieved comparable success. In this paper, we propose a new fine-tuning technique for learning new knowledge and show that it can reach the performance of RAG. The proposed method is based on the self-distillation approach, which we call prompt distillation. First, we generate question-answer pairs about the new knowledge. Then, we fine-tune a student model on the question-answer pairs to imitate the output distributions of a teacher model, which additionally receives the new knowledge in its prompt. The student model is identical to the teacher, except it is equipped with a LoRA adapter. This training procedure facilitates distilling the new knowledge from the teacher's prompt into the student's weights.

This study investigates the internal representations of verb-particle combinations within transformer-based large language models (LLMs), specifically examining how these models capture lexical and syntactic nuances at different neural network layers. Employing the BERT architecture, we analyse the representational efficacy of its layers for various verb-particle constructions such as 'agree on', 'come back', and 'give up'. Our methodology includes a detailed dataset preparation from the British National Corpus, followed by extensive model training and output analysis through techniques like multi-dimensional scaling (MDS) and generalized discrimination value (GDV) calculations. Results show that BERT's middle layers most effectively capture syntactic structures, with significant variability in representational accuracy across different verb categories. These findings challenge the conventional uniformity assumed in neural network processing of linguistic elements and suggest a complex interplay between network architecture and linguistic representation. Our research contributes to a better understanding of how deep learning models comprehend and process language, offering insights into the potential and limitations of current neural approaches to linguistic analysis. This study not only advances our knowledge in computational linguistics but also prompts further research into optimizing neural architectures for enhanced linguistic precision.

The proliferation of radical content on online platforms poses significant risks, including inciting violence and spreading extremist ideologies. Despite ongoing research, existing datasets and models often fail to address the complexities of multilingual and diverse data. To bridge this gap, we introduce a publicly available multilingual dataset annotated with radicalization levels, calls for action, and named entities in English, French, and Arabic. This dataset is pseudonymized to protect individual privacy while preserving contextual information. Beyond presenting our freely available dataset, we analyze the annotation process, highlighting biases and disagreements among annotators and their implications for model performance. Additionally, we use synthetic data to investigate the influence of socio-demographic traits on annotation patterns and model predictions. Our work offers a comprehensive examination of the challenges and opportunities in building robust datasets for radical content detection, emphasizing the importance of fairness and transparency in model development.

Large language models (LLMs) have gained significant attention over the past few years, particularly following the emergence of ChatGPT, both from researchers Movva et al. [2024] and in terms of their adoption in the private sector. This hype has revolutionized how we use AI in different fields and helped redefine how various domains utilize AI, such as in medicine Alghamdi and Mostafa [2024], Yuan et al. [2024], finance Xie et al. [2024], Malaysha et al. [2024], and even agriculture Gupta et al. [2024]. These advancements demonstrate the profound potential of AI to transform industries, driving innovation and efficiency in ways previously unimaginable. However, one domain that still has room for growth and the potential to bring about significant change is the legal domain Martin et al. [2024], Maree et al. [2024]. Although sectors such as healthcare and finance have rapidly adopted AI to address their unique challenges, the legal industry has been relatively slow to embrace these technologies Legg and Bell [2020]. The complex nature of legal language, coupled with jurisdictional variations and the high stakes involved, and the lack of AI regulations de Almeida et al. [2021], Nadjia [2024], has presented significant obstacles to developing effective AI-powered legal

The recent advancements in large language models (LLMs) have significantly improved language understanding and generation capabilities. However, it is difficult to deploy LLMs on resource-constrained edge devices due to their high computational and storage resource demands. To address this issue, we propose a novel LLM model pruning method, namely structurally-aware adaptive pruning (SAAP), to significantly reduce the computational and memory costs while maintaining model performance. We first define an adaptive importance fusion metric to evaluate the importance of all coupled structures in LLMs by considering their homoscedastic uncertainty. Then, we rank the importance of all modules to determine the specific layers that should be pruned to meet particular performance requirements. Furthermore, we develop a new group fine-tuning strategy to improve the inference efficiency of LLMs. Finally, we evaluate the proposed SAAP method on multiple LLMs across two common tasks, i.e., zero-shot classification and text generation. Experimental results show that our SAAP method outperforms several state-of-the-art baseline methods, achieving 2.17%, 2.37%, and 2.39% accuracy gains on LLaMA-7B, Vicuna-7B, and LLaMA-13B. Additionally, SAAP improves the token generation speed by 5%, showcasing its practical advantages in resource-constrained scenarios.