Understanding the Generalization of In-Context Learning in Transformers: An Empirical Study

Large language models (LLMs) like GPT-4 and LLaMA-3 utilize the powerful in-context learning (ICL) capability of Transformer architecture to learn on the fly from limited examples. While ICL underpins many LLM applications, its full potential remains hindered by a limited understanding of its generalization boundaries and vulnerabilities. We present a systematic investigation of transformers' generalization capability with ICL relative to training data coverage by defining a task-centric framework along three dimensions: inter-problem, intra-problem, and intra-task generalization. Through extensive simulation and real-world experiments, encompassing tasks such as function fitting, API calling, and translation, we find that transformers lack inter-problem generalization with ICL, but excel in intra-task and intra-problem generalization. When the training data includes a greater variety of mixed tasks, it significantly enhances the generalization ability of ICL on unseen tasks and even on known simple tasks. This guides us in designing training data to maximize the diversity of tasks covered and to combine different tasks whenever possible, rather than solely focusing on the target task for testing.

Error Slice Discovery via Manifold Compactness

Despite the great performance of deep learning models in many areas, they still make mistakes and underperform on certain subsets of data, i.e. error slices. Given a trained model, it is important to identify its semantically coherent error slices that are easy to interpret, which is referred to as the error slice discovery problem. However, there is no proper metric of slice coherence without relying on extra information like predefined slice labels. Current evaluation of slice coherence requires access to predefined slices formulated by metadata like attributes or subclasses. Its validity heavily relies on the quality and abundance of metadata, where some possible patterns could be ignored. Besides, current algorithms cannot directly incorporate the constraint of coherence into their optimization objective due to the absence of an explicit coherence metric, which could potentially hinder their effectiveness. In this paper, we propose manifold compactness, a coherence metric without reliance on extra information by incorporating the data geometry property into its design, and experiments on typical datasets empirically validate the rationality of the metric. Then we develop Manifold Compactness based error Slice Discovery (MCSD), a novel algorithm that directly treats risk and coherence as the optimization objective, and is flexible to be applied to models of various tasks. Extensive experiments on the benchmark and case studies on other typical datasets demonstrate the superiority of MCSD.

Hidden Question Representations Tell Non-Factuality Within and Across Large Language Models

Despite the remarkable advance of large language models (LLMs), the prevalence of non-factual responses remains a common issue. This work studies non-factuality prediction (NFP), which predicts whether an LLM will generate non-factual responses to a question before the generation process. Previous efforts on NFP usually rely on extensive computation. In this work, we conduct extensive analysis to explore the capabilities of using a lightweight probe to elicit ``whether an LLM knows'' from the hidden representations of questions. Additionally, we discover that the non-factuality probe employs similar patterns for NFP across multiple LLMs. Motivated by the intriguing finding, we conduct effective transfer learning for cross-LLM NFP and propose a question-aligned strategy to ensure the efficacy of mini-batch based training.

Geometry-Calibrated DRO: Combating Over-Pessimism with Free Energy Implications

Machine learning algorithms minimizing average risk are susceptible to distributional shifts. Distributionally Robust Optimization (DRO) addresses this issue by optimizing the worst-case risk within an uncertainty set. However, DRO suffers from over-pessimism, leading to low-confidence predictions, poor parameter estimations as well as poor generalization. In this work, we conduct a theoretical analysis of a probable root cause of over-pessimism: excessive focus on noisy samples. To alleviate the impact of noise, we incorporate data geometry into calibration terms in DRO, resulting in our novel Geometry-Calibrated DRO (GCDRO) for regression. We establish the connection between our risk objective and the Helmholtz free energy in statistical physics, and this free-energy-based risk can extend to standard DRO methods. Leveraging gradient flow in Wasserstein space, we develop an approximate minimax optimization algorithm with a bounded error ratio and elucidate how our approach mitigates noisy sample effects. Comprehensive experiments confirm GCDRO's superiority over conventional DRO methods.

CVSNet: A Computer Implementation for Central Visual System of The Brain

In computer vision, different basic blocks are created around different matrix operations, and models based on different basic blocks have achieved good results. Good results achieved in vision tasks grants them rationality. However, these experimental-based models also make deep learning long criticized for principle and interpretability. Deep learning originated from the concept of neurons in neuroscience, but recent designs detached natural neural networks except for some simple concepts. In this paper, we build an artificial neural network, CVSNet, which can be seen as a computer implementation for central visual system of the brain. Each block in CVSNet represents the same vision information as that in brains. In CVSNet, blocks differs from each other and visual information flows through three independent pathways and five different blocks. Thus CVSNet is completely different from the design of all previous models, in which basic blocks are repeated to build model and information between channels is mixed at the outset. In ablation experiment, we show the information extracted by blocks in CVSNet and compare with previous networks, proving effectiveness and rationality of blocks in CVSNet from experiment side. And in the experiment of object recognition, CVSNet achieves comparable results to ConvNets, Vision Transformers and MLPs.

Diffusion Models in NLP: A Survey

This survey paper provides a comprehensive review of the use of diffusion models in natural language processing (NLP). Diffusion models are a class of mathematical models that aim to capture the diffusion of information or signals across a network or manifold. In NLP, diffusion models have been used in a variety of applications, such as natural language generation, sentiment analysis, topic modeling, and machine translation. This paper discusses the different formulations of diffusion models used in NLP, their strengths and limitations, and their applications. We also perform a thorough comparison between diffusion models and alternative generative models, specifically highlighting the autoregressive (AR) models, while also examining how diverse architectures incorporate the Transformer in conjunction with diffusion models. Compared to AR models, diffusion models have significant advantages for parallel generation, text interpolation, token-level controls such as syntactic structures and semantic contents, and robustness. Exploring further permutations of integrating Transformers into diffusion models would be a valuable pursuit. Also, the development of multimodal diffusion models and large-scale diffusion language models with notable capabilities for few-shot learning would be important directions for the future advance of diffusion models in NLP.

Exploring and Exploiting Data Heterogeneity in Recommendation

Massive amounts of data are the foundation of data-driven recommendation models. As an inherent nature of big data, data heterogeneity widely exists in real-world recommendation systems. It reflects the differences in the properties among sub-populations. Ignoring the heterogeneity in recommendation data could limit the performance of recommendation models, hurt the sub-populational robustness, and make the models misled by biases. However, data heterogeneity has not attracted substantial attention in the recommendation community. Therefore, it inspires us to adequately explore and exploit heterogeneity for solving the above problems and assisting data analysis. In this work, we focus on exploring two representative categories of heterogeneity in recommendation data that is the heterogeneity of prediction mechanism and covariate distribution and propose an algorithm that explores the heterogeneity through a bilevel clustering method. Furthermore, the uncovered heterogeneity is exploited for two purposes in recommendation scenarios which are prediction with multiple sub-models and supporting debias. Extensive experiments on real-world data validate the existence of heterogeneity in recommendation data and the effectiveness of exploring and exploiting data heterogeneity in recommendation.

Gradient Norm Aware Minimization Seeks First-Order Flatness and Improves Generalization

Recently, flat minima are proven to be effective for improving generalization and sharpness-aware minimization (SAM) achieves state-of-the-art performance. Yet the current definition of flatness discussed in SAM and its follow-ups are limited to the zeroth-order flatness (i.e., the worst-case loss within a perturbation radius). We show that the zeroth-order flatness can be insufficient to discriminate minima with low generalization error from those with high generalization error both when there is a single minimum or multiple minima within the given perturbation radius. Thus we present first-order flatness, a stronger measure of flatness focusing on the maximal gradient norm within a perturbation radius which bounds both the maximal eigenvalue of Hessian at local minima and the regularization function of SAM. We also present a novel training procedure named Gradient norm Aware Minimization (GAM) to seek minima with uniformly small curvature across all directions. Experimental results show that GAM improves the generalization of models trained with current optimizers such as SGD and AdamW on various datasets and networks. Furthermore, we show that GAM can help SAM find flatter minima and achieve better generalization.

Improving Question Answering with Generation of NQ-like Questions

Question Answering (QA) systems require a large amount of annotated data which is costly and time-consuming to gather. Converting datasets of existing QA benchmarks are challenging due to different formats and complexities. To address these issues, we propose an algorithm to automatically generate shorter questions resembling day-to-day human communication in the Natural Questions (NQ) dataset from longer trivia questions in Quizbowl (QB) dataset by leveraging conversion in style among the datasets. This provides an automated way to generate more data for our QA systems. To ensure quality as well as quantity of data, we detect and remove ill-formed questions using a neural classifier. We demonstrate that in a low resource setting, using the generated data improves the QA performance over the baseline system on both NQ and QB data. Our algorithm improves the scalability of training data while maintaining quality of data for QA systems.

CausPref: Causal Preference Learning for Out-of-Distribution Recommendation

In spite of the tremendous development of recommender system owing to the progressive capability of machine learning recently, the current recommender system is still vulnerable to the distribution shift of users and items in realistic scenarios, leading to the sharp decline of performance in testing environments. It is even more severe in many common applications where only the implicit feedback from sparse data is available. Hence, it is crucial to promote the performance stability of recommendation method in different environments. In this work, we first make a thorough analysis of implicit recommendation problem from the viewpoint of out-of-distribution (OOD) generalization. Then under the guidance of our theoretical analysis, we propose to incorporate the recommendation-specific DAG learner into a novel causal preference-based recommendation framework named CausPref, mainly consisting of causal learning of invariant user preference and anti-preference negative sampling to deal with implicit feedback. Extensive experimental results from real-world datasets clearly demonstrate that our approach surpasses the benchmark models significantly under types of out-of-distribution settings, and show its impressive interpretability.