Recently, open source large language models We found that our model achieved better evaluation (LLMs) have grown drastically in both popularity scores on multilingual chat benchmarks compared and performance. Models such as Llama to the similarly sized state-of-the-art open 3 (AI@Meta, 2024b) have exceeded the performance source models, indicating the high quality and of previous state-of-the-art proprietary models diversity of our training dataset. We also find like GPT3.5 (Ouyang et al., 2022) on popular that our multilingual-trained LLM performs better robust benchmarks including the Chatbot Arena on Japanese chat benchmarks compared to our leaderboard (Chiang et al., 2024). These open Japanese-only-trained LLM, indicating that transfer source LLMs are also increasingly being used in learning from training on other languages is commerical AI chat products such as the Meta AI beneficial for training even monolingual models assistant (AI@Meta, 2024a).

This study evaluates the security of web application code generated by Large Language Models, analyzing 2,500 GPT-4 generated PHP websites. These were deployed in Docker containers and tested for vulnerabilities using a hybrid approach of Burp Suite active scanning, static analysis, and manual review. Our investigation focuses on identifying Insecure File Upload, SQL Injection, Stored XSS, and Reflected XSS in GPT-4 generated PHP code. This analysis highlights potential security risks and the implications of deploying such code in real-world scenarios. Overall, our analysis found 2,440 vulnerable parameters. According to Burp's Scan, 11.56% of the sites can be straight out compromised. Adding static scan results, 26% had at least one vulnerability that can be exploited through web interaction. Certain coding scenarios, like file upload functionality, are insecure 78% of the time, underscoring significant risks to software safety and security. To support further research, we have made the source codes and a detailed vulnerability record for each sample publicly available. This study emphasizes the crucial need for thorough testing and evaluation if generative AI technologies are used in software development.

In the domain of large language models (LLMs), arXiv:2305.16938 showed that few-shot full-model fine-tuning -- namely Vanilla Fine Tuning (FT) and Pattern-Based Fine Tuning (PBFT) --, and In-Context Learning (ICL) generalize similarly on Out-Of-Domain (OOD) datasets, but vary in terms of task adaptation. However, they both pose challenges, especially in term of memory requirements. In this paper, we further try to push the understanding of different fine-tuning strategies for LLM and aim to bring a myriad of these on the same pedestal for an elaborate comparison with full-model fine-tuning on two diverse datasets. To that end, we conducted a series of experiments, beginning with state-of-the-art methods like vanilla fine-tuning and Pattern-Based Fine-Tuning (PBFT) on pre-trained models across two datasets, COLA and MNLI. We then investigate adaptive fine-tuning and the efficiency of LoRA adapters in a few-shot setting. Finally, we also compare an alternative approach that has gained recent popularity -- context distillation -- with the vanilla FT and PBFT with and without few-shot setup. Our findings suggest that these alternative strategies that we explored can exhibit out-of-domain generalization comparable to that of vanilla FT and PBFT. PBFT under-performs Vanilla FT on out-of-domain (OOD) data, emphasizing the need for effective prompts. Further, our adaptive-fine tuning and LoRA experiments perform comparable or slightly worse than the standard fine-tunings as anticipated, since standard fine-tunings involve tuning the entire model. Finally, our context distillation experiments out-perform the standard fine-tuning methods. These findings underscore that eventually the choice of an appropriate fine-tuning method depends on the available resources (memory, compute, data) and task adaptability.

Recent work has aimed to improve LLM generations by filtering out hallucinations, thereby improving the precision of the information in responses. Correctness of a long-form response, however, also depends on the recall of multiple pieces of information relevant to the question. In this paper, we introduce Atomic Self-Consistency (ASC), a technique for improving the recall of relevant information in an LLM response. ASC follows recent work, Universal Self-Consistency (USC) in using multiple stochastic samples from an LLM to improve the long-form response. Unlike USC which only focuses on selecting the best single generation, ASC picks authentic subparts from the samples and merges them into a superior composite answer. Through extensive experiments and ablations, we show that merging relevant subparts of multiple samples performs significantly better than picking a single sample. ASC demonstrates significant gains over USC on multiple factoids and open-ended QA datasets - ASQA, QAMPARI, QUEST, ELI5 with ChatGPT and Llama2. Our analysis also reveals untapped potential for enhancing long-form generations using approach of merging multiple samples.

In the medical domain, numerous scenarios necessitate the long-form generation ability of large language models (LLMs). Specifically, when addressing patients' questions, it is essential that the model's response conveys factual claims, highlighting the need for an automated method to evaluate those claims. Thus, we introduce MedLFQA, a benchmark dataset reconstructed using long-form question-answering datasets related to the biomedical domain. We use MedLFQA to facilitate the automatic evaluations of factuality. We also propose OLAPH, a simple and novel framework that enables the improvement of factuality through automatic evaluations. The OLAPH framework iteratively trains LLMs to mitigate hallucinations using sampling predictions and preference optimization. In other words, we iteratively set the highest-scoring response as a preferred response derived from sampling predictions and train LLMs to align with the preferred response that improves factuality. We highlight that, even on evaluation metrics not used during training, LLMs trained with our OLAPH framework demonstrate significant performance improvement in factuality. Our findings reveal that a 7B LLM trained with our OLAPH framework can provide long answers comparable to the medical experts' answers in terms of factuality. We believe that our work could shed light on gauging the long-text generation ability of LLMs in the medical domain. Our code and datasets are available at https://github.com/dmis-lab/OLAPH}{https://github.com/dmis-lab/OLAPH.

The knowledge within large language models (LLMs) may become outdated quickly. While in-context editing (ICE) is currently the most effective method for knowledge editing (KE), it is constrained by the black-box modeling of LLMs and thus lacks interpretability. Our work aims to elucidate the superior performance of ICE on the KE by analyzing the impacts of in-context new knowledge on token-wise distributions. We observe that despite a significant boost in logits of the new knowledge, the performance of is still hindered by stubborn knowledge. Stubborn knowledge refers to as facts that have gained excessive confidence during pretraining, making it hard to edit effectively. To address this issue and further enhance the performance of ICE, we propose a novel approach termed $\textbf{De}$coding by $\textbf{C}$ontrasting $\textbf{K}$nowledge (DeCK). DeCK derives the distribution of the next token by contrasting the logits obtained from the newly edited knowledge guided by ICE with those from the unedited parametric knowledge. Our experiments consistently demonstrate that DeCK enhances the confidence of LLMs in edited facts. For instance, it improves the performance of LLaMA3-8B-instruct on MQuAKE by up to 219%, demonstrating its capability to strengthen ICE in the editing of stubborn knowledge. Our work paves the way to develop the both effective and accountable KE methods for LLMs. (The source code is available at: https://deck-llm.meirtz.com)

Large Language Models (LLMs) have revolutionised the field of Natural Language Processing (NLP) and have achieved state-of-the-art performance in practically every task in this field. However, the prevalent approach used in text generation, Causal Language Modelling (CLM), which generates text sequentially from left to right, inherently limits the freedom of the model, which does not decide when and where each token is generated. In contrast, Masked Language Modelling (MLM), primarily used for language understanding tasks, can generate tokens anywhere in the text and any order. This paper conducts an extensive comparison of MLM and CLM approaches for text generation tasks. To do so, we pre-train several language models of comparable sizes on three different datasets, namely 1) medical discharge summaries, 2) movie plot synopses, and 3) authorship verification datasets. To assess the quality of the generations, we first employ quantitative metrics and then perform a qualitative human evaluation to analyse coherence and grammatical correctness. In addition, we evaluate the usefulness of the generated texts by using them in three different downstream tasks: 1) Entity Recognition, 2) Text Classification, and 3) Authorship Verification. The results show that MLM consistently outperforms CLM in text generation across all datasets, with higher quantitative scores and better coherence in the generated text. The study also finds \textit{no strong correlation} between the quality of the generated text and the performance of the models in the downstream tasks. With this study, we show that MLM for text generation has great potential for future research and provides direction for future studies in this area.

Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.

Understanding the process of emotion generation is crucial for analyzing the causes behind emotions. Causal Emotion Entailment (CEE), an emotion-understanding task, aims to identify the causal utterances in a conversation that stimulate the emotions expressed in a target utterance. However, current works in CEE mainly focus on modeling semantic and emotional interactions in conversations, neglecting the exploration of the emotion-generation process. This hinders the models from deeply understanding emotions, restricting their ability to produce explainable predictions. In this work, inspired by the emotion generation process of "stimulus-appraisal-emotion" in the cognitive appraisal theory, we introduce a step-by-step reasoning method, Emotion-Cause Reasoning Chain (ECR-Chain), to infer the stimulus from the target emotional expressions in conversations. Specifically, we first introduce the ECR-Chain to ChatGPT via few-shot prompting, which significantly improves its performance on the CEE task. We further propose an automated construction process to utilize ChatGPT in building an ECR-Chain set, which can enhance the reasoning abilities of smaller models through supervised training and assist the Vicuna-7B model in achieving state-of-the-art CEE performance. Moreover, our methods can enable these generative language models to effectively perform emotion-cause reasoning in an explainable manner. Our code, data and more details are at https://github.com/hzp3517/ECR-Chain.

Software developers use natural language to interact not only with other humans, but increasingly also with chatbots. These interactions have different properties and flow differently based on what goal the developer wants to achieve and who they interact with. In this paper, we aim to understand the dynamics of conversations that occur during modern software development after the integration of AI and chatbots, enabling a deeper recognition of the advantages and disadvantages of including chatbot interactions in addition to human conversations in collaborative work. We compile existing conversation attributes with humans and NLU-based chatbots and adapt them to the context of software development. Then, we extend the comparison to include LLM-powered chatbots based on an observational study. We present similarities and differences between human-to-human and human-to-bot conversations, also distinguishing between NLU- and LLM-based chatbots. Furthermore, we discuss how understanding the differences among the conversation styles guides the developer on how to shape their expectations from a conversation and consequently support the communication within a software team. We conclude that the recent conversation styles that we observe with LLM-chatbots can not replace conversations with humans due to certain attributes regarding social aspects despite their ability to support productivity and decrease the developers' mental load.