Recent advances in generative artificial intelligence (AI), including large language models (LLMs), have opened new possibilities in education and in particular on scaling up personalization. One form of personalization that generative AI powers is interactive learning via tutoring dialogues between AI-powered tutors and students. These interactions have the potential to tailor instruction to each student's needs and progress, while offering personalized feedback, all in real time, in a scalable way. Given the widespread success of human tutors for improving student outcomes [29], many recent works have developed LLM-based tutors, showing promise across various educational domains [15, 25, 30, 32, 33, 39, 42, 50]. Many LLM-based tutors are even deployed in practice, such as Khan Academy's Khanmigo [21] and Carnegie Learning's LiveHint [4]. Several preliminary studies have shown that interacting with LLMs can increase student learning [52], although some have shown that students can develop an over-reliance on LLMs which negatively impacts their learning [23]. Many prior works have focused on improving LLMs' ability to follow effective tutoring principles, adapting them for the tutoring task that they are not pre-trained for. One approach, explored in [46], analyzes the decision-making process underlying human tutor utterances, showing that integrating expert decisions enhances LLM-based tutoring. Another study, [28], examines tutor moves in interactions with an LLM-powered simulated student agent, demonstrating that move annotation data contributes to better tutoring performance.

This paper introduces MalAlgoQA, a novel dataset designed to evaluate the counterfactual reasoning capabilities of Large Language Models (LLMs) through a pedagogical approach. The dataset comprises mathematics and reading comprehension questions, each accompanied by four answer choices and their corresponding rationales. We focus on the incorrect answer rationales, termed "malgorithms", which highlights flawed reasoning steps leading to incorrect answers and offers valuable insights into erroneous thought processes. We also propose the Malgorithm Identification task, where LLMs are assessed based on their ability to identify corresponding malgorithm given an incorrect answer choice. To evaluate the model performance, we introduce two metrics: Algorithm Identification Accuracy (AIA) for correct answer rationale identification, and Malgorithm Identification Accuracy (MIA) for incorrect answer rationale identification. The task is challenging since state-of-the-art LLMs exhibit significant drops in MIA as compared to AIA. Moreover, we find that the chain-of-thought prompting technique not only fails to consistently enhance MIA, but can also lead to underperformance compared to simple prompting. These findings hold significant implications for the development of more cognitively-inspired LLMs to improve their counterfactual reasoning abilities, particularly through a pedagogical perspective where understanding and rectifying student misconceptions are crucial.

Despite rapid advancements in large language models (LLMs), QG remains a challenging problem due to its complicated process, open-ended nature, and the diverse settings in which question generation occurs. A common approach to address these challenges involves fine-tuning smaller, custom models using datasets containing background context, question, and answer. However, obtaining suitable domain-specific datasets with appropriate context is often more difficult than acquiring question-answer pairs. In this paper, we investigate training QG models using synthetic contexts generated by LLMs from readily available question-answer pairs. We conduct a comprehensive study to answer critical research questions related to the performance of models trained on synthetic contexts and their potential impact on QG research and applications. Our empirical results reveal: 1) contexts are essential for QG tasks, even if they are synthetic; 2) fine-tuning smaller language models has the capability of achieving better performances as compared to prompting larger language models; and 3) synthetic context and real context could achieve comparable performances. These findings highlight the effectiveness of synthetic contexts in QG and paves the way for future advancements in the field.

We develop Scalable Latent Exploration Score (ScaLES) to mitigate overexploration in Latent Space Optimization (LSO), a popular method for solving black-box discrete optimization problems. LSO utilizes continuous optimization within the latent space of a Variational Autoencoder (VAE) and is known to be susceptible to over-exploration, which manifests in unrealistic solutions that reduce its practicality. ScaLES is an exact and theoretically motivated method leveraging the trained decoder's approximation of the data distribution. ScaLES can be calculated with any existing decoder, e.g. from a VAE, without additional training, architectural changes, or access to the training data. Our evaluation across five LSO benchmark tasks and three VAE architectures demonstrates that ScaLES enhances the quality of the solutions while maintaining high objective values, leading to improvements over existing solutions. We believe that new avenues to LSO will be opened by ScaLES' ability to identify out of distribution areas, differentiability, and computational tractability.

Grokking, or delayed generalization, is a phenomenon where generalization in a deep neural network (DNN) occurs long after achieving near zero training error. Previous studies have reported the occurrence of grokking in specific controlled settings, such as DNNs initialized with large-norm parameters or transformers trained on algorithmic datasets. We demonstrate that grokking is actually much more widespread and materializes in a wide range of practical settings, such as training of a convolutional neural network (CNN) on CIFAR10 or a Resnet on Imagenette. We introduce the new concept of delayed robustness, whereby a DNN groks adversarial examples and becomes robust, long after interpolation and/or generalization. We develop an analytical explanation for the emergence of both delayed generalization and delayed robustness based on the local complexity of a DNN's input-output mapping. Our local complexity measures the density of so-called linear regions (aka, spline partition regions) that tile the DNN input space and serves as a utile progress measure for training. We provide the first evidence that, for classification problems, the linear regions undergo a phase transition during training whereafter they migrate away from the training samples (making the DNN mapping smoother there) and towards the decision boundary (making the DNN mapping less smooth there). Grokking occurs post phase transition as a robust partition of the input space thanks to the linearization of the DNN mapping around the training points. Website: https://bit.ly/grok-adversarial

We study overparameterization in generative adversarial networks (GANs) that can interpolate the training data. We show that overparameterization can improve generalization performance and accelerate the training process. We study the generalization error as a function of latent space dimension and identify two main behaviors, depending on the learning setting. First, we show that overparameterized generative models that learn distributions by minimizing a metric or $f$-divergence do not exhibit double descent in generalization errors; specifically, all the interpolating solutions achieve the same generalization error. Second, we develop a novel pseudo-supervised learning approach for GANs where the training utilizes pairs of fabricated (noise) inputs in conjunction with real output samples. Our pseudo-supervised setting exhibits double descent (and in some cases, triple descent) of generalization errors. We combine pseudo-supervision with overparameterization (i.e., overly large latent space dimension) to accelerate training while matching or even surpassing generalization performance without pseudo-supervision. While our analysis focuses mostly on linear models, we also apply important insights for improving generalization of nonlinear, multilayer GANs.

The goal of this series is to chronicle opinions and issues in the field of machine learning as they stand today and as they change over time. The plan is to host this survey periodically until the AI singularity paperclip-frenzy-driven doomsday, keeping an updated list of topical questions and interviewing new community members for each edition.

The study of Deep Network (DN) training dynamics has largely focused on the evolution of the loss function, evaluated on or around train and test set data points. In fact, many DN phenomenon were first introduced in literature with that respect, e.g., double descent, grokking. In this study, we look at the training dynamics of the input space partition or linear regions formed by continuous piecewise affine DNs, e.g., networks with (leaky)ReLU nonlinearities. First, we present a novel statistic that encompasses the local complexity (LC) of the DN based on the concentration of linear regions inside arbitrary dimensional neighborhoods around data points. We observe that during training, the LC around data points undergoes a number of phases, starting with a decreasing trend after initialization, followed by an ascent and ending with a final descending trend. Using exact visualization methods, we come across the perplexing observation that during the final LC descent phase of training, linear regions migrate away from training and test samples towards the decision boundary, making the DN input-output nearly linear everywhere else. We also observe that the different LC phases are closely related to the memorization and generalization performance of the DN, especially during grokking.

We study the new problem of automatic question generation (QG) from multi-modal sources containing images and texts, significantly expanding the scope of most of the existing work that focuses exclusively on QG from only textual sources. We propose a simple solution for our new problem, called MultiQG-TI, which enables a text-only question generator to process visual input in addition to textual input. Specifically, we leverage an image-to-text model and an optical character recognition model to obtain the textual description of the image and extract any texts in the image, respectively, and then feed them together with the input texts to the question generator. We only fine-tune the question generator while keeping the other components fixed. On the challenging ScienceQA dataset, we demonstrate that MultiQG-TI significantly outperforms ChatGPT with few-shot prompting, despite having hundred-times less trainable parameters. Additional analyses empirically confirm the necessity of both visual and textual signals for QG and show the impact of various modeling choices.

Automated scoring of open-ended student responses has the potential to significantly reduce human grader effort. Recent advances in automated scoring often leverage textual representations based on pre-trained language models such as BERT and GPT as input to scoring models. Most existing approaches train a separate model for each item/question, which is suitable for scenarios such as essay scoring where items can be quite different from one another. However, these approaches have two limitations: 1) they fail to leverage item linkage for scenarios such as reading comprehension where multiple items may share a reading passage; 2) they are not scalable since storing one model per item becomes difficult when models have a large number of parameters. In this paper, we report our (grand prize-winning) solution to the National Assessment of Education Progress (NAEP) automated scoring challenge for reading comprehension. Our approach, in-context BERT fine-tuning, produces a single shared scoring model for all items with a carefully-designed input structure to provide contextual information on each item. We demonstrate the effectiveness of our approach via local evaluations using the training dataset provided by the challenge. We also discuss the biases, common error types, and limitations of our approach.