The development of Electronic Health Records summarization systems has revolutionized patient data management. Previous research advanced this field by adapting Large Language Models for clinical tasks, using diverse datasets to generate general EHR summaries. However, clinicians often require specific, focused summaries for quicker insights. This project builds on prior work by creating a system that generates clinician-preferred, focused summaries, improving EHR summarization for more efficient patient care. The proposed system leverages the Google Flan-T5 model to generate tailored EHR summaries based on clinician-specified topics. The approach involved fine-tuning the Flan-T5 model on an EHR question-answering dataset formatted in the Stanford Question Answering Dataset (SQuAD) style, which is a large-scale reading comprehension dataset with questions and answers. Fine-tuning utilized the Seq2SeqTrainer from the Hugging Face Transformers library with optimized hyperparameters. Key evaluation metrics demonstrated promising results: the system achieved an Exact Match (EM) score of 81.81%. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) metrics showed strong performance, with ROUGE-1 at 96.03%, ROUGE-2 at 86.67%, and ROUGE-L at 96.10%. Additionally, the Bilingual Evaluation Understudy (BLEU) score was 63%, reflecting the model's coherence in generating summaries. By enhancing EHR summarization through LLMs, this project supports digital transformation efforts in healthcare, streamlining workflows, and enabling more personalized patient care.

The goal of this paper is to improve (upcycle) an existing large language model without the prohibitive requirements of continued pre-training of the full-model. The idea is to split the pre-training data into semantically relevant groups and train an expert on each subset. An expert takes the form of a lightweight adapter added on the top of a frozen base model. During inference, an incoming query is first routed to the most relevant expert which is then loaded onto the base model for the forward pass. Unlike typical Mixture of Experts (MoE) models, the experts in our method do not work with other experts for a single query. Hence, we dub them "introvert" experts. Freezing the base model and keeping the experts as lightweight adapters allows extreme parallelism during training and inference. Training of all experts can be done in parallel without any communication channels between them. Similarly, the inference can also be heavily parallelized by distributing experts on different GPUs and routing each request to the GPU containing its relevant expert. We implement a proof-of-concept version of this method and show the validity of our approach.

Following natural instructions is crucial for the effective application of Retrieval-Augmented Generation (RAG) systems. Despite recent advancements in Large Language Models (LLMs), research on assessing and improving instruction-following (IF) alignment within the RAG domain remains limited. To address this issue, we propose VIF-RAG, the first automated, scalable, and verifiable synthetic pipeline for instruction-following alignment in RAG systems. We start by manually crafting a minimal set of atomic instructions (<100) and developing combination rules to synthesize and verify complex instructions for a seed set. We then use supervised models for instruction rewriting while simultaneously generating code to automate the verification of instruction quality via a Python executor. Finally, we integrate these instructions with extensive RAG and general data samples, scaling up to a high-quality VIF-RAG-QA dataset (>100k) through automated processes. To further bridge the gap in instruction-following auto-evaluation for RAG systems, we introduce FollowRAG Benchmark, which includes approximately 3K test samples, covering 22 categories of general instruction constraints and four knowledge-intensive QA datasets. Due to its robust pipeline design, FollowRAG can seamlessly integrate with different RAG benchmarks. Using FollowRAG and eight widely-used IF and foundational abilities benchmarks for LLMs, we demonstrate that VIF-RAG markedly enhances LLM performance across a broad range of general instruction constraints while effectively leveraging its capabilities in RAG scenarios. Further analysis offers practical insights for achieving IF alignment in RAG systems. Our code and datasets are released at https://FollowRAG.github.io.

Modern cryptographic methods for implementing privacy-preserving LLMs such as Homomorphic Encryption (HE) require the LLMs to have a polynomial form. Forming such a representation is challenging because Transformers include non-polynomial components, such as Softmax and layer normalization. Previous approaches have either directly approximated pre-trained models with large-degree polynomials, which are less efficient over HE, or replaced non-polynomial components with easier-to-approximate primitives before training, e.g., Softmax with pointwise attention. The latter approach might introduce scalability challenges. We present a new HE-friendly variant of self-attention that offers a stable form for training and is easy to approximate with polynomials for secure inference. Our work introduces the first polynomial LLMs with 32 layers and over a billion parameters, exceeding the size of previous models by more than tenfold. The resulting models demonstrate reasoning and in-context learning (ICL) capabilities comparable to standard transformers of the same size, representing a breakthrough in the field. Finally, we provide a detailed latency breakdown for each computation over encrypted data, paving the way for further optimization, and explore the differences in inductive bias between transformers relying on our HE-friendly variant and standard transformers. Our code is attached as a supplement.

In today's day and age where information is rapidly spread through online platforms, the rise of fake news poses an alarming threat to the integrity of public discourse, societal trust, and reputed news sources. Classical machine learning and Transformer-based models have been extensively studied for the task of fake news detection, however they are hampered by their reliance on training data and are unable to generalize on unseen headlines. To address these challenges, we propose our novel solution, leveraging web-scraping techniques and Natural Language Inference (NLI) models to retrieve external knowledge necessary for verifying the accuracy of a headline. Our system is evaluated on a diverse self-curated evaluation dataset spanning over multiple news channels and broad domains. Our best performing pipeline achieves an accuracy of 84.3% surpassing the best classical Machine Learning model by 33.3% and Bidirectional Encoder Representations from Transformers (BERT) by 31.0% . This highlights the efficacy of combining dynamic web-scraping with Natural Language Inference to find support for a claimed headline in the corresponding externally retrieved knowledge for the task of fake news detection.

Large Language Models (LLMs) perform outstandingly in various downstream tasks, and the use of the Retrieval-Augmented Generation (RAG) architecture has been shown to improve performance for legal question answering (Nuruzzaman and Hussain, 2020; Louis et al., 2024). However, there are limited applications in insurance questions-answering, a specific type of legal document. This paper introduces two corpora: the Quebec Automobile Insurance Expertise Reference Corpus and a set of 82 Expert Answers to Layperson Automobile Insurance Questions. Our study leverages both corpora to automatically and manually assess a GPT4-o, a state-of-the-art LLM, to answer Quebec automobile insurance questions. Our results demonstrate that, on average, using our expertise reference corpus generates better responses on both automatic and manual evaluation metrics. However, they also highlight that LLM QA is unreliable enough for mass utilization in critical areas. Indeed, our results show that between 5% to 13% of answered questions include a false statement that could lead to customer misunderstanding.

Collaborative learning of large language models (LLMs) has emerged as a new paradigm for utilizing private data from different parties to guarantee efficiency and privacy. Meanwhile, Knowledge Editing (KE) for LLMs has also garnered increased attention due to its ability to manipulate the behaviors of LLMs explicitly, yet leaves the collaborative KE case (in which knowledge edits of multiple parties are aggregated in a privacy-preserving and continual manner) unexamined. To this end, this manuscript dives into the first investigation of collaborative KE, in which we start by carefully identifying the unique three challenges therein, including knowledge overlap, knowledge conflict, and knowledge forgetting. We then propose a non-destructive collaborative KE framework, COLLABEDIT, which employs a novel model merging mechanism to mimic the global KE behavior while preventing the severe performance drop. Extensive experiments on two canonical datasets demonstrate the superiority of COLLABEDIT compared to other destructive baselines, and results shed light on addressing three collaborative KE challenges and future applications.

While server-side Large Language Models (LLMs) demonstrate proficiency in function calling and complex reasoning, deploying Small Language Models (SLMs) directly on devices brings opportunities to improve latency and privacy but also introduces unique challenges for accuracy and memory. We introduce CAMPHOR, an innovative on-device SLM multi-agent framework designed to handle multiple user inputs and reason over personal context locally, ensuring privacy is maintained. CAMPHOR employs a hierarchical architecture where a high-order reasoning agent decomposes complex tasks and coordinates expert agents responsible for personal context retrieval, tool interaction, and dynamic plan generation. By implementing parameter sharing across agents and leveraging prompt compression, we significantly reduce model size, latency, and memory usage. To validate our approach, we present a novel dataset capturing multi-agent task trajectories centered on personalized mobile assistant use-cases. Our experiments reveal that fine-tuned SLM agents not only surpass closed-source LLMs in task completion F1 by~35\% but also eliminate the need for server-device communication, all while enhancing privacy.

Rapid advancements in artificial intelligence (AI) have made it crucial to integrate moral reasoning into AI systems. However, existing models and datasets often overlook regional and cultural differences. To address this shortcoming, we have expanded the JCommonsenseMorality (JCM) dataset, the only publicly available dataset focused on Japanese morality. The Extended JCM (eJCM) has grown from the original 13,975 sentences to 31,184 sentences using our proposed sentence expansion method called Masked Token and Label Enhancement (MTLE). MTLE selectively masks important parts of sentences related to moral judgment and replaces them with alternative expressions generated by a large language model (LLM), while re-assigning appropriate labels. The model trained using our eJCM achieved an F1 score of 0.857, higher than the scores for the original JCM (0.837), ChatGPT one-shot classification (0.841), and data augmented using AugGPT, a state-of-the-art augmentation method (0.850). Specifically, in complex moral reasoning tasks unique to Japanese culture, the model trained with eJCM showed a significant improvement in performance (increasing from 0.681 to 0.756) and achieved a performance close to that of GPT-4 Turbo (0.787). These results demonstrate the validity of the eJCM dataset and the importance of developing models and datasets that consider the cultural context.

Low-Rank Adaptation (LoRA) is the bread and butter of Large Language Model (LLM) finetuning. LoRA learns an additive low-rank perturbation, $AB$, of a pretrained matrix parameter $W$ to align the model to a new task or dataset with $W+AB$. We identify three core limitations to LoRA for finetuning--a setting that employs limited amount of data and training steps. First, LoRA employs Dropout to prevent overfitting. We prove that Dropout is only suitable for long training episodes but fails to converge to a reliable regularizer for short training episodes. Second, LoRA's initialization of $B$ at $0$ creates a slow training dynamic between $A$ and $B$. That dynamic is also exacerbated by Dropout that further slows the escape from $0$ for $B$ which is particularly harmful for short training episodes. Third, the scaling factor multiplying each LoRA additive perturbation creates ``short-sighted'' interactions between the LoRA modules of different layers. Motivated by principled analysis of those limitations, we find an elegant solution: a Dropout-free, scaling-free, LoRA with Adaptive Learning rate--coined ALLoRA. By scaling the per sample and per parameter gradients with a coefficient inversely proportional to parameters' $\ell_2$ norm, ALLoRA alleviates those three limitations. As a by-product, ALLoRA removes two hyper-parameters from LoRA: the scaling factor and the dropout rate. Empirical results show that ALLoRA admits better accuracy than LoRA on various settings, including against recent LoRA variants such as Weight-Decomposed Low-Rank Adaptation (DoRA). Ablation studies show our solution is the optimal in a family of weight-dependent / output-dependent approaches on various LLMs including the latest Llama3.