Large Language Models (LLMs) have shown impressive capabilities in natural language processing, yet their use in sensitive domains like healthcare, particularly with Electronic Health Records (EHR), faces significant challenges due to privacy concerns and limited computational resources. This paper presents a compact LLM framework designed for local deployment in settings with strict privacy requirements and limited access to high-performance GPUs. We introduce a novel preprocessing technique that uses information extraction methods, e.g., regular expressions, to filter and emphasize critical information in clinical notes, enhancing the performance of smaller LLMs on EHR data. Our framework is evaluated using zero-shot and few-shot learning paradigms on both private and publicly available (MIMIC-IV) datasets, and we also compare its performance with fine-tuned LLMs on the MIMIC-IV dataset. The results demonstrate that our preprocessing approach significantly boosts the prediction accuracy of smaller LLMs, making them suitable for high-privacy, resource-constrained applications. This study offers valuable insights into optimizing LLM performance for sensitive, data-intensive tasks while addressing computational and privacy limitations.

We investigate biases in pretraining datasets for large language models (LLMs) through dataset classification experiments. Building on prior work demonstrating the existence of biases in popular computer vision datasets, we analyze popular open-source pretraining datasets for LLMs derived from CommonCrawl including C4, RefinedWeb, DolmaCC, RedPajama-V2, FineWeb, and DCLM-Baseline. Despite those datasets being obtained with similar filtering and deduplication steps, neural networks can classify surprisingly well which dataset a single text sequence belongs to, significantly better than a human can. This indicates that popular pretraining datasets have their own unique biases or fingerprints. Those biases remain even when the text is rewritten with LLMs. Moreover, these biases propagate through training: Random sequences generated by models trained on those datasets can be classified well by a classifier trained on the original datasets.

Recently, utilizing large language models (LLMs) for metaphor detection has achieved promising results. However, these methods heavily rely on the capabilities of closed-source LLMs, which come with relatively high inference costs and latency. To address this, we propose a method for metaphor detection by fine-tuning open-source LLMs, effectively reducing inference costs and latency with a single inference step. Furthermore, metaphor detection suffers from a severe data scarcity problem, which hinders effective fine-tuning of LLMs. To tackle this, we introduce Curriculum-style Data Augmentation (CDA). Specifically, before fine-tuning, we evaluate the training data to identify correctly predicted instances for fine-tuning, while incorrectly predicted instances are used as seed data for data augmentation. This approach enables the model to quickly learn simpler knowledge and progressively acquire more complex knowledge, thereby improving performance incrementally. Experimental results demonstrate that our method achieves state-of-the-art performance across all baselines. Additionally, we provide detailed ablation studies to validate the effectiveness of CDA.

This paper presents Patent-CR, the first dataset created for the patent claim revision task in English. It includes both initial patent applications rejected by patent examiners and the final granted versions. Unlike normal text revision tasks that predominantly focus on enhancing sentence quality, such as grammar correction and coherence improvement, patent claim revision aims at ensuring the claims meet stringent legal criteria. These criteria are beyond novelty and inventiveness, including clarity of scope, technical accuracy, language precision, and legal robustness. We assess various large language models (LLMs) through professional human evaluation, including general LLMs with different sizes and architectures, text revision models, and domain-specific models. Our results indicate that LLMs often bring ineffective edits that deviate from the target revisions. In addition, domain-specific models and the method of fine-tuning show promising results. Notably, GPT-4 outperforms other tested LLMs, but further revisions are still necessary to reach the examination standard. Furthermore, we demonstrate the inconsistency between automated and human evaluation results, suggesting that GPT-4-based automated evaluation has the highest correlation with human judgment. This dataset, along with our preliminary empirical research, offers invaluable insights for further exploration in patent claim revision.

Sycophancy refers to the tendency of a large language model to align its outputs with the user's perceived preferences, beliefs, or opinions, in order to look favorable, regardless of whether those statements are factually correct. This behavior can lead to undesirable consequences, such as reinforcing discriminatory biases or amplifying misinformation. Given that sycophancy is often linked to human feedback training mechanisms, this study explores whether sycophantic tendencies negatively impact user trust in large language models or, conversely, whether users consider such behavior as favorable. To investigate this, we instructed one group of participants to answer ground-truth questions with the assistance of a GPT specifically designed to provide sycophantic responses, while another group used the standard version of ChatGPT. Initially, participants were required to use the language model, after which they were given the option to continue using it if they found it trustworthy and useful. Trust was measured through both demonstrated actions and self-reported perceptions. The findings consistently show that participants exposed to sycophantic behavior reported and exhibited lower levels of trust compared to those who interacted with the standard version of the model, despite the opportunity to verify the accuracy of the model's output.

In this study, we develop and assess new corpus selection and training methodologies to improve the effectiveness of Turkish language models. Specifically, we adapted Large Language Model generated datasets and translated English datasets into Turkish, integrating these resources into the training process. This approach led to substantial enhancements in model accuracy for both few-shot and zero-shot learning scenarios. Furthermore, the merging of these adapted models was found to markedly improve their performance. Human evaluative metrics, including task-specific performance assessments, further demonstrated that these adapted models possess a greater aptitude for comprehending the Turkish language and addressing logic-based queries. This research underscores the importance of refining corpus selection strategies to optimize the performance of multilingual models, particularly for under-resourced languages like Turkish.

Recent English Common Crawl datasets like FineWeb-Edu and DCLM achieved significant benchmark gains via aggressive model-based filtering, but at the cost of removing 90% of data. This limits their suitability for long token horizon training, such as 15T tokens for Llama 3.1. In this paper, we show how to achieve better trade-offs between accuracy and data quantity by a combination of classifier ensembling, synthetic data rephrasing, and reduced reliance on heuristic filters. When training 8B parameter models for 1T tokens, using a high-quality subset of our data improves MMLU by 5.6 over DCLM, demonstrating the efficacy of our methods for boosting accuracies over a relatively short token horizon. Furthermore, our full 6.3T token dataset matches DCLM on MMLU, but contains four times more unique real tokens than DCLM. This unlocks state-of-the-art training over a long token horizon: an 8B parameter model trained for 15T tokens, of which 7.2T came from our dataset, is better than the Llama 3.1 8B model: +5 on MMLU, +3.1 on ARC-Challenge, and +0.5 on average across ten diverse tasks. The dataset is available at https://data.commoncrawl.org/contrib/Nemotron/Nemotron-CC/index.html

What factors contribute to the relative success and corresponding difficulties of in-context learning for Large Language Models (LLMs)? Drawing on insights from the literature on human concept learning, we test LLMs on carefully designed concept learning tasks, and show that task performance highly correlates with the Boolean complexity of the concept. This suggests that in-context learning exhibits a learning bias for simplicity in a way similar to humans.

With the rise of sophisticated phishing attacks, there is a growing need for effective and economical detection solutions. This paper explores the use of large multimodal agents, specifically Gemini 1.5 Flash and GPT-4o mini, to analyze both URLs and webpage screenshots via APIs, thus avoiding the complexities of training and maintaining AI systems. Our findings indicate that integrating these two data types substantially enhances detection performance over using either type alone. However, API usage incurs costs per query that depend on the number of input and output tokens. To address this, we propose a two-tiered agentic approach: initially, one agent assesses the URL, and if inconclusive, a second agent evaluates both the URL and the screenshot. This method not only maintains robust detection performance but also significantly reduces API costs by minimizing unnecessary multi-input queries. Cost analysis shows that with the agentic approach, GPT-4o mini can process about 4.2 times as many websites per $100 compared to the multimodal approach (107,440 vs. 25,626), and Gemini 1.5 Flash can process about 2.6 times more websites (2,232,142 vs. 862,068). These findings underscore the significant economic benefits of the agentic approach over the multimodal method, providing a viable solution for organizations aiming to leverage advanced AI for phishing detection while controlling expenses.

Reinforcement learning from human feedback (RLHF) has been crucial in aligning large language models (LLMs) with human values. Traditionally, RLHF involves generating responses to a query and using a reward model to assign a reward to the entire response. However, this approach faces challenges due to its reliance on a single, sparse reward, which makes it challenging for the model to identify which parts of the sequence contribute most significantly to the final reward. Recent methods have attempted to address this limitation by introducing token-level rewards. However, these methods often rely on either a trained credit assignment model or AI annotators, raising concerns about the quality and reliability of the rewards. In this paper, we propose token-level reward regularization (T-REG), a novel approach that leverages both sequence-level and token-level rewards for preference optimization. Harnessing the self-refinement capabilities of LLMs, our method uses contrastive prompting to enable LLMs to self-generate token-level rewards. These self-generated rewards then act as reward regularization, guiding the model to more effectively distribute sequence-level rewards across tokens. This facilitates better token-level credit assignment and enhances alignment performance. Experiments on the instruction following benchmarks, including Alpaca Eval 2 and Arena-Hard, show that our method consistently outperforms baseline methods by up to 3.8% and 4.4%, respectively. We will release the code and models at https://github.com/wzhouad/T-REG.