Pre-trained large language models have demonstrated a strong ability to learn from context, known as in-context learning (ICL). Despite a surge of recent applications that leverage such capabilities, it is by no means clear, at least theoretically, how the ICL capabilities arise, and in particular, what is the precise role played by key factors such as pre-training procedure as well as context construction. In this work, we propose a new framework to analyze the ICL performance, for a class of realistic settings, which includes network architectures, data encoding, data generation, and prompt construction process. As a first step, we construct a simple example with a one-layer transformer, and show an interesting result, namely when the pre-train data distribution is different from the query task distribution, a properly constructed context can shift the output distribution towards the query task distribution, in a quantifiable manner, leading to accurate prediction on the query topic. We then extend the findings in the previous step to a more general case, and derive the precise relationship between ICL performance, context length and the KL divergence between pre-train and query task distribution. Finally, we provide experiments to validate our theoretical results.

We give these results in Table 1. The first column in the table shows the sequence id of each task sequence in each mixed sequence dataset. By comparing the rows for "EMNIST -10/20" and "CIFAR100-10/20", we can see how each model in the dissimilar task dataset are performing. "F-EMNIST" and "F-CelebA ", we can see the performance on similar task datasets. CA T consistently gives the best results for overall and similar task datasets.

Electronic health records (EHRs) provide a comprehensive source of longitudinal patient data, encompassing structured modalities such as laboratory results, imaging data, and vital signs, and unstructured clinical notes. These datasets, after necessary preprocessing to clean and format the data for analysis, often remain in their raw EHR form, representing numerical or categorical values without further transformation into task-agnostic embeddings. While such raw EHR data enables predictive modeling, its reliance on manual feature engineering or downstream task-specific optimization limits its utility for general-purpose applications. Deep learning (DL) techniques, such as recurrent neural networks (RNNs) and Transformers, have facilitated predictive tasks like disease progression and diagnosis prediction. However, these methods often struggle to fully exploit the temporal and multimodal dependencies inherent in EHR data due to their reliance on pre-processed but untransformed raw EHR inputs. In this study, we introduce CAAT-EHR, a novel architecture designed to bridge this gap by generating robust, task-agnostic longitudinal embeddings from raw EHR data. CAAT-EHR leverages self- and cross-attention mechanisms in its encoder to integrate temporal and contextual relationships across multiple modalities, transforming the data into enriched embeddings that capture complex dependencies. An autoregressive decoder complements the encoder by predicting future time points data during pre-training, ensuring that the resulting embeddings maintain temporal consistency and alignment. CAAT-EHR eliminates the need for manual feature engineering and enables seamless transferability across diverse downstream tasks. Extensive evaluations on benchmark datasets, demonstrate the superiority of CAAT-EHR-generated embeddings over pre-processed raw EHR data and other baseline approaches.

In-context learning (ICL) of large language models (LLMs) has attracted increasing attention in the community where LLMs make predictions only based on instructions augmented with a few examples. Existing example selection methods for ICL utilize sparse or dense retrievers and derive effective performance. However, these methods do not utilize direct feedback of LLM to train the retriever and the examples selected can not necessarily improve the analogy ability of LLM. To tackle this, we propose our policy-based reinforcement learning framework for example selection (RLS), which consists of a language model (LM) selector and an LLM generator. The LM selector encodes the candidate examples into dense representations and selects the top-k examples into the demonstration for LLM. The outputs of LLM are adopted to compute the reward and policy gradient to optimize the LM selector. We conduct experiments on different datasets and significantly outperform existing example selection methods. Moreover, our approach shows advantages over supervised finetuning (SFT) models in few shot setting. Further experiments show the balance of abundance and the similarity with the test case of examples is important for ICL performance of LLM.

This work presents a dynamic vocabulary adaptation strategy, MEDVOC, for fine-tuning pre-trained language models (PLMs) like BertSumAbs, BART, and PEGASUS for improved medical text summarization. In contrast to existing domain adaptation approaches in summarization, MEDVOC treats vocabulary as an optimizable parameter and optimizes the PLM vocabulary based on fragment score conditioned only on the downstream task's reference summaries. Unlike previous works on vocabulary adaptation (limited only to classification tasks), optimizing vocabulary based on summarization tasks requires an extremely costly intermediate fine-tuning step on large summarization datasets. To that end, our novel fragment score-based hyperparameter search very significantly reduces this fine-tuning time -- from 450 days to less than 2 days on average. Furthermore, while previous works on vocabulary adaptation are often primarily tied to single PLMs, MEDVOC is designed to be deployable across multiple PLMs (with varying model vocabulary sizes, pre-training objectives, and model sizes) -- bridging the limited vocabulary overlap between the biomedical literature domain and PLMs. MEDVOC outperforms baselines by 15.74% in terms of Rouge-L in zero-shot setting and shows gains of 17.29% in high Out-Of-Vocabulary (OOV) concentrations. Our human evaluation shows MEDVOC generates more faithful medical summaries (88% compared to 59% in baselines). We make the codebase publicly available at https://github.com/gb-kgp/MEDVOC.

Large language models (LLMs) have received a lot of attention in natural language processing (NLP) research because of their exceptional performance in understanding and generating human languages. However, low-resource languages are left behind due to the unavailability of resources. In this work, we focus on enhancing the LLaMA-2-Amharic model by integrating task-specific and generative datasets to improve language model performance for Amharic. We compile an Amharic instruction fine-tuning dataset and fine-tuned LLaMA-2-Amharic model. The fine-tuned model shows promising results in different NLP tasks. We open-source our dataset creation pipeline, instruction datasets, trained models, and evaluation outputs to promote language-specific studies on these models.

Recent advancements in multimodal techniques open exciting possibilities for models excelling in diverse tasks involving text, audio, and image processing. Models like GPT-4V, blending computer vision and language modeling, excel in complex text and image tasks. Numerous prior research endeavors have diligently examined the performance of these Vision Large Language Models (VLLMs) across tasks like object detection, image captioning and others. However, these analyses often focus on evaluating the performance of each modality in isolation, lacking insights into their cross-modal interactions. Specifically, questions concerning whether these vision-language models execute vision and language tasks consistently or independently have remained unanswered. In this study, we draw inspiration from recent investigations into multilingualism and conduct a comprehensive analysis of model's cross-modal interactions. We introduce a systematic framework that quantifies the capability disparities between different modalities in the multi-modal setting and provide a set of datasets designed for these evaluations. Our findings reveal that models like GPT-4V tend to perform consistently modalities when the tasks are relatively simple. However, the trustworthiness of results derived from the vision modality diminishes as the tasks become more challenging. Expanding on our findings, we introduce "Vision Description Prompting," a method that effectively improves performance in challenging vision-related tasks.

Video is a promising source of knowledge for embodied agents to learn models of the world's dynamics. Large deep networks have become increasingly effective at modeling complex video data in a self-supervised manner, as evaluated by metrics based on human perceptual similarity or pixel-wise comparison. However, it remains unclear whether current metrics are accurate indicators of performance on downstream tasks. We find empirically that for planning robotic manipulation, existing metrics can be unreliable at predicting execution success. To address this, we propose a benchmark for action-conditioned video prediction in the form of a control benchmark that evaluates a given model for simulated robotic manipulation through sampling-based planning. Our benchmark, Video Prediction for Visual Planning ($VP^2$), includes simulated environments with 11 task categories and 310 task instance definitions, a full planning implementation, and training datasets containing scripted interaction trajectories for each task category. A central design goal of our benchmark is to expose a simple interface -- a single forward prediction call -- so it is straightforward to evaluate almost any action-conditioned video prediction model. We then leverage our benchmark to study the effects of scaling model size, quantity of training data, and model ensembling by analyzing five highly-performant video prediction models, finding that while scale can improve perceptual quality when modeling visually diverse settings, other attributes such as uncertainty awareness can also aid planning performance.

The performance of deep neural networks can be highly sensitive to the choice of a variety of meta-parameters, such as optimizer parameters and model hyperparameters. Tuning these well, however, often requires extensive and costly experimentation. Bayesian optimization (BO) is a principled approach to solve such expensive hyperparameter tuning problems efficiently. Key to the performance of BO is specifying and refining a distribution over functions, which is used to reason about the optima of the underlying function being optimized. In this work, we consider the scenario where we have data from similar functions that allows us to specify a tighter distribution a priori. Specifically, we focus on the common but potentially costly task of tuning optimizer parameters for training neural networks. Building on the meta BO method from Wang et al. (2018), we develop practical improvements that (a) boost its performance by leveraging tuning results on multiple tasks without requiring observations for the same meta-parameter points across all tasks, and (b) retain its regret bound for a special case of our method. As a result, we provide a coherent BO solution for iterative optimization of continuous optimizer parameters. To verify our approach in realistic model training setups, we collected a large multi-task hyperparameter tuning dataset by training tens of thousands of configurations of near-state-of-the-art models on popular image and text datasets, as well as a protein sequence dataset. Our results show that on average, our method is able to locate good hyperparameters at least 3 times more efficiently than the best competing methods.