WPN: An Unlearning Method Based on N-pair Contrastive Learning in Language Models

The public release of ChatGPT, a generative artificial intelligence language model, caused wide-spread public interest in its abilities but also concern about the implications of the application on academia, depending on whether it was deemed benevolent (e.g., supporting analysis and simplification of tasks) or malevolent (e.g., assignment writing and academic misconduct). While ChatGPT has been shown to provide answers of sufficient quality to pass some university exams, its capacity to write essays that require an exploration of value concepts is unknown. This paper presents the results of a study where ChatGPT-4 (released May 2023) was tasked with writing a 1500-word essay to discuss the nature of values used in the assessment of cultural heritage significance. Based on an analysis of 36 iterations, ChatGPT wrote essays of limited length with about 50% of the stipulated word count being primarily descriptive and without any depth or complexity. The concepts, which are often flawed and suffer from inverted logic, are presented in an arbitrary sequence with limited coherence and without any defined line of argument. Given that it is a generative language model, ChatGPT often splits concepts and uses one or more words to develop tangential arguments. While ChatGPT provides references as tasked, many are fictitious, albeit with plausible authors and titles. At present, ChatGPT has the ability to critique its own work but seems unable to incorporate that critique in a meaningful way to improve a previous draft. Setting aside conceptual flaws such as inverted logic, several of the essays could possibly pass as a junior high school assignment but fall short of what would be expected in senior school, let alone at a college or university level.

The extensive spread of harmful content, including hate speech, harassment, violent and adult material across online platforms and media channels poses significant challenges and has raised considerable concern within various societal groups. Government bodies, educators, and parents frequently find themselves in disagreement with media providers over the best methods to regulate and restrict such content. Automated content moderation technologies have become critical tools in addressing these challenges, particularly through natural language processing (NLP) techniques that can automatically detect and filter sensitive textual content, such as offensive language, violence, and adult materials, enabling platforms to enforce moderation policies on a large scale. Despite their widespread use, current moderation technologies face challenges related to detection accuracy, often resulting in substantial false positives and false negatives. To enhance content moderation systems, more advanced algorithms capable of accurately interpreting textual context are necessary. In this study, we assess current large language model (LLM)-based moderation solutions, specifically OpenAI moderation model and Llama-Guard-3, examining their effectiveness in detecting sensitive content. Additionally, we investigate the capabilities of contemporary LLMs, including OpenAI generative pre-trained transformer (GPT), Google Gemini, Meta Llama, Anthropic Claude, and small language models (SLMs) such as Google Gemma, in recognizing inappropriate content from diverse media sources. We also studied the performance of these models under adversarial attacks such as input perturbation and prompt injection. Our evaluation and comparative analysis utilize various textual datasets, such as X (Twitter) posts, Amazon product reviews, and news articles. The findings indicate that LLM-based approaches significantly demonstrate high accuracy and low rates of false positives and negatives. They also indicate the robustness of the models under various adversarial attacks. These results underscore the considerable potential for integrating advanced LLMs into websites and social media platforms, thereby enhancing content regulation and moderation effectiveness.

Recent advancements in large language models (LLMs) have raised concerns about inference costs, increasing the need for research into model compression. While knowledge distillation (KD) is a prominent method for this, research on KD for generative language models like LLMs is relatively sparse, and the approach of distilling student-friendly knowledge, which has shown promising performance in KD for classification models, remains unexplored in generative language models. To explore this approach, we propose PromptKD, a simple yet effective method that utilizes prompt tuningfor the first time in KD -to enable generative language models to transfer student-friendly knowledge. Unlike previous works in classification that require fine-tuning the entire teacher model for extracting student-friendly knowledge, PromptKD achieves similar effects by adding a small number of prompt tokens and tuning only the prompt with student guidance. Extensive experiments on instruction-following datasets show that PromptKD achieves state-ofthe-art performance while adding only 0.0007% of the teacher's parameters as prompts. Further analysis suggests that distilling student-friendly knowledge alleviates exposure bias effectively throughout the entire training process, leading to performance enhancements. 1

The advent of powerful transformer-based discriminative language models and, more recently, generative GPT-family models, has led to notable advancements in natural language processing (NLP), particularly in commonsense reasoning tasks. One such task is commonsense reasoning, where performance is usually evaluated through multiple-choice question-answering benchmarks. Till date, many such benchmarks have been proposed and `leaderboards' tracking state-of-the-art performance on those benchmarks suggest that transformer-based models are approaching human-like performance. However, due to documented problems such as hallucination and bias, the research focus is shifting from merely quantifying accuracy on the task to an in-depth, context-sensitive probing of LLMs' generalization and robustness. To gain deeper insight into diagnosing these models' performance in commonsense reasoning scenarios, this thesis addresses three main studies: the generalization ability of transformer-based language models on commonsense reasoning, the trend in confidence distribution of these language models confronted with ambiguous inference tasks, and a proposed risk-centric evaluation framework for both discriminative and generative language models.

Targeted protein degradation of pathogenic proteins represents a powerful new treatment strategy for multiple cancers. Unfortunately, a sizable portion of these proteins are considered “undruggable” by standard small molecule-based approaches, including PROTACs and molecular glues, largely due to their disordered nature, instability, and lack of binding site accessibility. As a more modular strategy, we have developed a genetically-encoded protein architecture by fusing target-specific peptides to E3 ubiquitin ligase domains for selective and potent intracellular degradation of oncoproteins. To enable programmability of our system, we develop a suite of algorithms that enable the design of target-specific peptides via protein language model (pLM) embeddings, without the requirement of 3D structures. First, we train a model that leverages pLM embeddings to efficiently select high-affinity peptides from natural protein interaction interfaces. Next, we develop a high-accuracy discriminator, based on the contrastive language-image pretraining (CLIP) architecture underlying OpenAI's DALL-E model, to prioritize and screen peptides with selectivity to a specified target oncoprotein. As input to the discriminator, we create a Gaussian diffusion generator to sample a pLM latent space, fine-tuned on experimentally-valid peptide sequences. Finally, to enable de novo design of binding peptides, we train an instance of GPT-2 with protein interacting sequences to enable peptide generation conditioned on target oncoprotein sequences. Our models demonstrate low perplexities across both existing and generated peptide sequences, highlighting their robust generative capability. By experimentally fusing model-derived peptides to E3 ubiquitin ligase domains, we reliably identify candidates exhibiting robust and selective endogenous degradation of diverse, "undruggable" oncoproteins in cancer cell models, including tumorigenic regulators such as β-catenin and TRIM8, as well as oncogenic fusion proteins, such as EWS-FLI1, PAX3-FOXO1, and DNAJB1-PRKACA. We further show that our peptide-guided degraders have negligible off-target effects via whole-cell proteomics and demonstrate their modulation of transcriptional and apoptotic pathways, motivating further translation of our therapeutic platform. Together, our work establishes a CRISPR-analogous system for programmable protein degradation applications across the oncoproteome. Citation Format: Suhaas Bhat, Garyk Brixi, Kalyan Palepu, Lauren Hong, Vivian Yudistyra, Tianlai Chen, Sophia Vincoff, Lin Zhao, Pranam Chatterjee. Design of programmable peptide-guided oncoprotein degraders via generative language models [abstract]. In: Proceedings of the AACR-NCI-EORTC Virtual International Conference on Molecular Targets and Cancer Therapeutics; 2023 Oct 11-15; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2023;22(12 Suppl):Abstract nr C118.

Extractive approaches have been the mainstream paradigm for identifying overlapping entity–relation extraction. However, limited by their inherently methodological flaws, which hardly deal with three issues: hierarchical dependent entity–relations, implicit entity–relations, and entity normalization. Recent advances have proposed an effective solution based on generative language models, which cast entity–relation extraction as a sequence-to-sequence text generation task. Inspired by the observation that humans learn by getting to the bottom of things, we propose a novel framework, namely GenRE, Generative multi-turn question answering with contrastive learning for entity–relation extraction. Specifically, a template-based question prompt generation first is designed to answer in different turns. We then formulate entity–relation extraction as a generative question answering task based on the general language model instead of span-based machine reading comprehension. Meanwhile, the contrastive learning strategy in fine-tuning is introduced to add negative samples to mitigate the exposure bias inherent in generative models. Our extensive experiments demonstrate that GenRE performs competitively on two public datasets and a custom dataset, highlighting its superiority in entity normalization and implicit entity–relation extraction. (The code is available at https://github.com/lovelyllwang/GenRE).

Statistical language models improve the performance of speech recognition systems by providing estimates of a priori probabilities of word sequences. The commonly used trigram language models obtain the conditional probability estimate of a word given the previous two words, from a large corpus of text. The text corpus is often a collection of several small diverse segments such as newspaper articles, or conversations on different topics. Knowledge of the current topic could be utilized to adapt the general trigram language models to match that topic closely. For example, an interpolation of the general language model with one built on the topic data could be used. The authors first discuss the adaptation of general trigram language models to a known topic using the minimum discrimination information (MDI) method. They then present results on the switchboard corpus which consists of telephone conversations on several topics.

Introduction. Modern education faces a contradiction between the active integration of generative artificial intelligence and its underexplored potential for providing evaluative feedback in development students’ mathematical literacy. The purpose of the article is to identify the potential of using a generative language model as a teacher’s tool for generating expert-level evaluative feedback when assessing open-ended mathematical problems Materials and Methods. The research is based on systemic-activity, criteria-oriented, and comparative approaches. Methods employed included theoretical analysis of scholarly literature, criteria-based assessment combined with prompt engineering techniques, as well as quantitative and qualitative analysis to determine the agreement between the evaluative feedback generated by the language model and that provided by a human expert. The sample consisted of 51 students Results. The research experimentally confirmed the feasibility of using generative artificial intelligence for providing evaluative feedback in mathematics education. An effective strategy for automating the assessment of open-ended mathematical problems was developed and substantiated, based on criteria-based assessment and prompt engineering techniques using GigaChat Pro language model. Empirical data revealed a moderate agreement between the evaluative feedback generated by GigaChat Pro and that provided by an expert teacher: accuracy reached 73%, Cohen’s coefficient (k) was 0,57, and the semantic similarity of textual comments (BertScore F1) was 0,614.

We introduce a novel approach for building language models based on a systematic, recursive exploration of skip n-gram models which are interpolated using modified Kneser-Ney smoothing. Our approach generalizes language models as it contains the classical interpolation with lower order models as a special case. In this paper we motivate, formalize and present our approach. In an extensive empirical experiment over English text corpora we demonstrate that our generalized language models lead to a substantial reduction of perplexity between 3.1% and 12.7% in comparison to traditional language models using modified Kneser-Ney smoothing. Furthermore, we investigate the behaviour over three other languages and a domain specific corpus where we observed consistent improvements. Finally, we also show that the strength of our approach lies in its ability to cope in particular with sparse training data. Using a very small training data set of only 736 KB text we yield improvements of even 25.7% reduction of perplexity.

Despite the superior performance of large language models to generate natural language texts, it is hard to generate texts with correct logic according to a given task, due to the difficulties for neural models to capture strict logic from free-form texts. In this paper, we propose a novel graph-based language model, Logical-GLM, to extract strict logic from free-form texts and then infuse into language models. Specifically, we first capture information from natural language instructions and construct logical probability graphs that generally describe domains. Next, we generate logical skeletons to guide language model training, infusing domain knowledge into language models. At last, we alternately optimize the searching policy of graphs and language models until convergence. The experimental results show that Logical-GLM is both effective and efficient compared with traditional language models, despite using smaller-scale training data and fewer parameters. Our approach can generate instructional texts with more correct logic owing to the internalized domain knowledge. Moreover, the search of logical graphs reflects the inner mechanism of the language models, which improves the interpretability of black-box models.

BACKGROUND Although uncertainties exist regarding implementation, artificial intelligence–driven generative language models (GLMs) have enormous potential in medicine. Deployment of GLMs could improve patient comprehension of clinical texts and improve low health literacy. OBJECTIVE The goal of this study is to evaluate the potential of ChatGPT-3.5 and GPT-4 to tailor the complexity of medical information to patient-specific input education level, which is crucial if it is to serve as a tool in addressing low health literacy. METHODS Input templates related to 2 prevalent chronic diseases—type II diabetes and hypertension—were designed. Each clinical vignette was adjusted for hypothetical patient education levels to evaluate output personalization. To assess the success of a GLM (GPT-3.5 and GPT-4) in tailoring output writing, the readability of pre- and posttransformation outputs were quantified using the Flesch reading ease score (FKRE) and the Flesch-Kincaid grade level (FKGL). RESULTS Responses (n=80) were generated using GPT-3.5 and GPT-4 across 2 clinical vignettes. For GPT-3.5, FKRE means were 57.75 (SD 4.75), 51.28 (SD 5.14), 32.28 (SD 4.52), and 28.31 (SD 5.22) for 6th grade, 8th grade, high school, and bachelor’s, respectively; FKGL mean scores were 9.08 (SD 0.90), 10.27 (SD 1.06), 13.4 (SD 0.80), and 13.74 (SD 1.18). GPT-3.5 only aligned with the prespecified education levels at the bachelor’s degree. Conversely, GPT-4’s FKRE mean scores were 74.54 (SD 2.6), 71.25 (SD 4.96), 47.61 (SD 6.13), and 13.71 (SD 5.77), with FKGL mean scores of 6.3 (SD 0.73), 6.7 (SD 1.11), 11.09 (SD 1.26), and 17.03 (SD 1.11) for the same respective education levels. GPT-4 met the target readability for all groups except the 6th-grade FKRE average. Both GLMs produced outputs with statistically significant differences (<i>P</i><.001; 8th grade <i>P</i><.001; high school <i>P</i><.001; bachelors <i>P</i>=.003; FKGL: 6th grade <i>P</i>=.001; 8th grade <i>P</i><.001; high school <i>P</i><.001; bachelors <i>P</i><.001) between mean FKRE and FKGL across input education levels. CONCLUSIONS GLMs can change the structure and readability of medical text outputs according to input-specified education. However, GLMs categorize input education designation into 3 broad tiers of output readability: easy (6th and 8th grade), medium (high school), and difficult (bachelor’s degree). This is the first result to suggest that there are broader boundaries in the success of GLMs in output text simplification. Future research must establish how GLMs can reliably personalize medical texts to prespecified education levels to enable a broader impact on health care literacy. CLINICALTRIAL

There has been an explosion of available pre-trained and fine-tuned Generative Language Models (LM). They vary in the number of parameters, architecture, training strategy, and training set size. Aligned with it, alternative strategies exist to exploit these models, such as Fine-tuning and Prompt Engineering. However, many questions may arise throughout this process: Which model to apply for a given task? Which strategies to use? Will Prompt Engineering solve all tasks? What are the computational and financial costs involved? This tutorial will introduce and explore typical modern LM architectures with a hands-on approach to the available strategies.

Continuous glucose monitors (CGM) provide valuable insights about glycemic control that aid in diabetes management. However, interpreting metrics and charts and synthesizing them into linguistic summaries is often non-trivial for patients and providers. The advent of large language models (LLMs) has enabled real-time text generation and summarization of medical data. The objective of this study was to assess the strengths and limitations of using an LLM to analyze raw CGM data and produce summaries of 14 days of data for patients with type 1 diabetes. We first evaluated the ability of GPT-4 to compute quantitative metrics specific to diabetes found in an Ambulatory Glucose Profile (AGP). Then, using two independent clinician graders, we evaluated the accuracy, completeness, safety, and suitability of qualitative descriptions produced by GPT-4 across five different CGM analysis tasks. GPT-4 performed 9 out of the 10 quantitative metrics tasks with perfect accuracy across all 10 cases. The clinician-evaluated CGM analysis tasks had good performance across measures of accuracy [lowest task mean score 8/10, highest task mean score 10/10], completeness [lowest task mean score 7.5/10, highest task mean score 10/10], and safety [lowest task mean score 9.5/10, highest task mean score 10/10]. Our work serves as a preliminary study on how generative language models can be integrated into diabetes care through data summarization and, more broadly, the potential to leverage LLMs for streamlined medical time series analysis.

The retrieval of similar documents in the Web from a given document is different in many aspects from information retrieval based on queries generated by regular search engine users. In this work, a new method is proposed for Web similarity document retrieval based on generative language models and meta search engines. Probabilistic language models are used as a random query generator for the given document. Queries are submitted to a customizable set of Web search engines. Once all results obtained are gathered, its evaluation is determined by a proposed scoring function based on the Zipf law. Results obtained showed that the proposed methodology for query generation and scoring procedure solves the problem with acceptable levels of precision.

The growing capability and availability of generative language models has enabled a wide range of new downstream tasks. Academic research has identified, quantified and mitigated biases present in language models but is rarely tailored to downstream tasks where wider impact on individuals and society can be felt. In this work, we leverage one popular generative language model, GPT-3, with the goal of writing unbiased and realistic job advertisements. We first assess the bias and realism of zero-shot generated advertisements and compare them to real-world advertisements. We then evaluate prompt-engineering and fine-tuning as debiasing methods. We find that prompt-engineering with diversity-encouraging prompts gives no significant improvement to bias, nor realism. Conversely, fine-tuning, especially on unbiased real advertisements, can improve realism and reduce bias.