Large Language Models (LLMs) have made progress in various real-world tasks, which stimulates requirements for the evaluation of LLMs. Existing LLM evaluation methods are mainly supervised signal-based which depends on static datasets and cannot evaluate the ability of LLMs in dynamic real-world scenarios where deep interaction widely exists. Other LLM evaluation methods are human-based which are costly and time-consuming and are incapable of large-scale evaluation of LLMs. To address the issues above, we propose a novel Deep Interaction-based LLM-evaluation framework. In our proposed framework, LLMs' performances in real-world domains can be evaluated from their deep interaction with other LLMs in elaborately designed evaluation tasks. Furthermore, our proposed framework is a general evaluation method that can be applied to a host of real-world tasks such as machine translation and code generation. We demonstrate the effectiveness of our proposed method through extensive experiments on four elaborately designed evaluation tasks.

Dataset distillation methods have demonstrated remarkable performance for neural networks trained with very limited training data. However, a significant challenge arises in the form of architecture overfitting: the distilled training data synthesized by a specific network architecture (i.e., training network) generates poor performance when trained by other network architectures (i.e., test networks). This paper addresses this issue and proposes a series of approaches in both architecture designs and training schemes which can be adopted together to boost the generalization performance across different network architectures on the distilled training data. We conduct extensive experiments to demonstrate the effectiveness and generality of our methods. Particularly, across various scenarios involving different sizes of distilled data, our approaches achieve comparable or superior performance to existing methods when training on the distilled data using networks with larger capacities.

This paper explores deep latent variable models for semi-supervised paraphrase generation, where the missing target pair for unlabelled data is modelled as a latent paraphrase sequence. We present a novel unsupervised model named variational sequence auto-encoding reconstruction (VSAR), which performs latent sequence inference given an observed text. To leverage information from text pairs, we additionally introduce a novel supervised model we call dual directional learning (DDL), which is designed to integrate with our proposed VSAR model. Combining VSAR with DDL (DDL+VSAR) enables us to conduct semi-supervised learning. Still, the combined model suffers from a cold-start problem. To further combat this issue, we propose an improved weight initialisation solution, leading to a novel two-stage training scheme we call knowledge-reinforced-learning (KRL). Our empirical evaluations suggest that the combined model yields competitive performance against the state-of-the-art supervised baselines on complete data. Furthermore, in scenarios where only a fraction of the labelled pairs are available, our combined model consistently outperforms the strong supervised model baseline (DDL) by a significant margin (p <.05; Wilcoxon test). Our code is publicly available at "https://github.com/jialin-yu/latent-sequence-paraphrase".

Statistical postprocessing is used to translate ensembles of raw numerical weather forecasts into reliable probabilistic forecast distributions. In this study, we examine the use of permutation-invariant neural networks for this task. In contrast to previous approaches, which often operate on ensemble summary statistics and dismiss details of the ensemble distribution, we propose networks which treat forecast ensembles as a set of unordered member forecasts and learn link functions that are by design invariant to permutations of the member ordering. We evaluate the quality of the obtained forecast distributions in terms of calibration and sharpness, and compare the models against classical and neural network-based benchmark methods. In case studies addressing the postprocessing of surface temperature and wind gust forecasts, we demonstrate state-of-the-art prediction quality. To deepen the understanding of the learned inference process, we further propose a permutation-based importance analysis for ensemble-valued predictors, which highlights specific aspects of the ensemble forecast that are considered important by the trained postprocessing models. Our results suggest that most of the relevant information is contained in few ensemble-internal degrees of freedom, which may impact the design of future ensemble forecasting and postprocessing systems.

A fundamental question in theoretical machine learning is generalization. Over the past decades, the PAC-Bayesian approach has been established as a flexible framework to address the generalization capabilities of machine learning algorithms, and design new ones. Recently, it has garnered increased interest due to its potential applicability for a variety of learning algorithms, including deep neural networks. In parallel, an information-theoretic view of generalization has developed, wherein the relation between generalization and various information measures has been established. This framework is intimately connected to the PAC-Bayesian approach, and a number of results have been independently discovered in both strands. In this monograph, we highlight this strong connection and present a unified treatment of generalization. We present techniques and results that the two perspectives have in common, and discuss the approaches and interpretations that differ. In particular, we demonstrate how many proofs in the area share a modular structure, through which the underlying ideas can be intuited. We pay special attention to the conditional mutual information (CMI) framework; analytical studies of the information complexity of learning algorithms; and the application of the proposed methods to deep learning. This monograph is intended to provide a comprehensive introduction to information-theoretic generalization bounds and their connection to PAC-Bayes, serving as a foundation from which the most recent developments are accessible. It is aimed broadly towards researchers with an interest in generalization and theoretical machine learning.

Online continual learning aims to continuously train neural networks from a continuous data stream with a single pass-through data. As the most effective approach, the rehearsal-based methods replay part of previous data. Commonly used predictors in existing methods tend to generate biased dot-product logits that prefer to the classes of current data, which is known as a bias issue and a phenomenon of forgetting. Many approaches have been proposed to overcome the forgetting problem by correcting the bias; however, they still need to be improved in online fashion. In this paper, we try to address the bias issue by a more straightforward and more efficient method. By decomposing the dot-product logits into an angle factor and a norm factor, we empirically find that the bias problem mainly occurs in the angle factor, which can be used to learn novel knowledge as cosine logits. On the contrary, the norm factor abandoned by existing methods helps remember historical knowledge. Based on this observation, we intuitively propose to leverage the norm factor to balance the new and old knowledge for addressing the bias. To this end, we develop a heuristic approach called unbias experience replay (UER). UER learns current samples only by the angle factor and further replays previous samples by both the norm and angle factors. Extensive experiments on three datasets show that UER achieves superior performance over various state-of-the-art methods. The code is in https://github.com/FelixHuiweiLin/UER.

Despite their impressive capabilities, large language models (LLMs) are prone to hallucinations, i.e., generating content that deviates from facts seen during pretraining. We propose a simple decoding strategy for reducing hallucinations with pretrained LLMs that does not require conditioning on retrieved external knowledge nor additional fine-tuning. Our approach obtains the next-token distribution by contrasting the differences in logits obtained from projecting the later layers versus earlier layers to the vocabulary space, exploiting the fact that factual knowledge in an LLMs has generally been shown to be localized to particular transformer layers. We find that this Decoding by Contrasting Layers (DoLa) approach is able to better surface factual knowledge and reduce the generation of incorrect facts. DoLa consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving the performance of LLaMA family models on TruthfulQA by 12-17% absolute points, demonstrating its potential in making LLMs reliably generate truthful facts.

Envision an intelligent agent capable of assisting users in conducting forecasting tasks through intuitive, natural conversations, without requiring in-depth knowledge of the underlying machine learning (ML) processes. A significant challenge for the agent in this endeavor is to accurately comprehend the user's prediction goals and, consequently, formulate precise ML tasks. In this paper, we take a pioneering step towards this ambitious goal by introducing a new concept called Forecast Utterance and then focus on the automatic and accurate interpretation of users' prediction goals from these utterances. Specifically, we frame the task as a slot-filling problem, where each slot corresponds to a specific aspect of the goal prediction task. We then employ two zero-shot methods for solving the slot-filling task, namely: 1) Entity Extraction (EE), and 2) Question-Answering (QA) techniques. Our experiments, conducted with three meticulously crafted data sets, validate the viability of our ambitious goal and demonstrate the effectiveness of both EE and QA techniques in interpreting Forecast Utterances.

Instruction-tuned Large Language Models (LLMs) have recently showcased remarkable ability to generate fitting responses to natural language instructions. However, an open research question concerns the inherent biases of trained models and their responses. For instance, if the data used to tune an LLM is dominantly written by persons with a specific political bias, we might expect generated answers to share this bias. Current research work seeks to de-bias such models, or suppress potentially biased answers. With this demonstration, we take a different view on biases in instruction-tuning: Rather than aiming to suppress them, we aim to make them explicit and transparent. To this end, we present OpinionGPT, a web demo in which users can ask questions and select all biases they wish to investigate. The demo will answer this question using a model fine-tuned on text representing each of the selected biases, allowing side-by-side comparison. To train the underlying model, we identified 11 different biases (political, geographic, gender, age) and derived an instruction-tuning corpus in which each answer was written by members of one of these demographics. This paper presents OpinionGPT, illustrates how we trained the bias-aware model and showcases the web application (available at https://opiniongpt.informatik.hu-berlin.de).

Over the last decades, Stochastic Gradient Descent (SGD) has been intensively studied by the Machine Learning community. Despite its versatility and excellent performance, the optimization of large models via SGD still is a time-consuming task. To reduce training time, it is common to distribute the training process across multiple devices. Recently, it has been shown that the convergence of asynchronous SGD (ASGD) will always be faster than mini-batch SGD. However, despite these improvements in the theoretical bounds, most ASGD convergence-rate proofs still rely on a centralized parameter server, which is prone to become a bottleneck when scaling out the gradient computations across many distributed processes. In this paper, we present a novel convergence-rate analysis for decentralized and asynchronous SGD (DASGD) which does not require partial synchronization among nodes nor restrictive network topologies. Specifically, we provide a bound of $\mathcal{O}(\sigma\epsilon^{-2}) + \mathcal{O}(QS_{avg}\epsilon^{-3/2}) + \mathcal{O}(S_{avg}\epsilon^{-1})$ for the convergence rate of DASGD, where $S_{avg}$ is the average staleness between models, $Q$ is a constant that bounds the norm of the gradients, and $\epsilon$ is a (small) error that is allowed within the bound. Furthermore, when gradients are not bounded, we prove the convergence rate of DASGD to be $\mathcal{O}(\sigma\epsilon^{-2}) + \mathcal{O}(\sqrt{\hat{S}_{avg}\hat{S}_{max}}\epsilon^{-1})$, with $\hat{S}_{max}$ and $\hat{S}_{avg}$ representing a loose version of the average and maximum staleness, respectively. Our convergence proof holds for a fixed stepsize and any non-convex, homogeneous, and L-smooth objective function. We anticipate that our results will be of high relevance for the adoption of DASGD by a broad community of researchers and developers.