Large language models (LLMs) are increasingly deployed in real-world systems, yet they can produce toxic or biased outputs that undermine safety and trust. Post-hoc model repair provides a practical remedy, but the high cost of parameter updates motivates selective use of repair data. Despite extensive prior work on data selection for model training, it remains unclear which sampling criteria are most effective and efficient when applied specifically to behavioral repair of large generative models. Our study presents a systematic analysis of sample prioritization strategies for LLM repair. We evaluate five representative selection methods, including random sampling, K-Center, gradient-norm-based selection(GraNd), stratified coverage (CCS), and a Semantic-Aware Prioritized Sampling (SAPS) approach we proposed. Repair effectiveness and trade-offs are assessed through toxicity reduction, perplexity on WikiText-2 and LAMBADA, and three composite metrics: the Repair Proximity Score (RPS), the Overall Performance Score (OPS), and the Repair Efficiency Score (RES). Experimental results show that SAPS achieves the best balance between detoxification, utility preservation, and efficiency, delivering comparable or superior repair outcomes with substantially less data. Random sampling remains effective for large or robust models, while high-overhead methods such as CCS and GraNd provide limited benefit. The optimal data proportion depends on model scale and repair method, indicating that sample selection should be regarded as a tunable component of repair pipelines. Overall, these findings establish selection-based repair as an efficient and scalable paradigm for maintaining LLM reliability.

Large Language Models (LLMs) have become indispensable tools across various applications, making it more important than ever to ensure the quality and the trustworthiness of their outputs. This has led to growing interest in uncertainty quantification (UQ) methods for assessing the reliability of LLM outputs. Many existing UQ techniques rely on token probabilities, which inadvertently introduces a bias with respect to the length of the output. While some methods attempt to account for this, we demonstrate that such biases persist even in length-normalized approaches. To address the problem, here we propose UNCERTAINTY-LINE: (Length-INvariant Estimation), a simple debiasing procedure that regresses uncertainty scores on output length and uses the residuals as corrected, length-invariant estimates. Our method is post-hoc, model-agnostic, and applicable to a range of UQ measures. Through extensive evaluation on machine translation, summarization, and question-answering tasks, we demonstrate that UNCERTAINTY-LINE: consistently improves over even nominally length-normalized UQ methods uncertainty estimates across multiple metrics and models.

Proper hyperparameter tuning is essential for achieving optimal performance of modern machine learning (ML) methods in predictive tasks. While there is an extensive literature on tuning ML learners for prediction, there is only little guidance available on tuning ML learners for causal machine learning and how to select among different ML learners. In this paper, we empirically assess the relationship between the predictive performance of ML methods and the resulting causal estimation based on the Double Machine Learning (DML) approach by Chernozhukov et al. (2018). DML relies on estimating so-called nuisance parameters by treating them as supervised learning problems and using them as plug-in estimates to solve for the (causal) parameter. We conduct an extensive simulation study using data from the 2019 Atlantic Causal Inference Conference Data Challenge. We provide empirical insights on the role of hyperparameter tuning and other practical decisions for causal estimation with DML. First, we assess the importance of data splitting schemes for tuning ML learners within Double Machine Learning. Second, we investigate how the choice of ML methods and hyperparameters, including recent AutoML frameworks, impacts the estimation performance for a causal parameter of interest. Third, we assess to what extent the choice of a particular causal model, as characterized by incorporated parametric assumptions, can be based on predictive performance metrics.

Recent advances in generative artificial intelligence (AI) have captured worldwide attention. Tools such as Dalle-2 and ChatGPT suggest that tasks previously thought to be beyond the capabilities of AI may now augment the productivity of creative media in various new ways, including through the generation of synthetic video. This research paper explores the utility of using AI-generated synthetic video to create viable educational content for online educational settings. To date, there is limited research investigating the real-world educational value of AI-generated synthetic media. To address this gap, we examined the impact of using AI-generated synthetic video in an online learning platform on both learners content acquisition and learning experience. We took a mixed-method approach, randomly assigning adult learners (n=83) into one of two micro-learning conditions, collecting pre- and post-learning assessments, and surveying participants on their learning experience. The control condition included a traditionally produced instructor video, while the experimental condition included a synthetic video with a realistic AI-generated character. The results show that learners in both conditions demonstrated significant improvement from pre- to post-learning (p<.001), with no significant differences in gains between the two conditions (p=.80). In addition, no differences were observed in how learners perceived the traditional and synthetic videos. These findings suggest that AI-generated synthetic learning videos have the potential to be a viable substitute for videos produced via traditional methods in online educational settings, making high quality educational content more accessible across the globe.

Littlestone and Warmuth [51] introduced sample compression schemes as an abstraction of the underlying structure of learning algorithms. Roughly, the aim of a sample compression scheme is to compress samples of a concept class (i.e., of a set system) C as much as possible, such that data coherent with the original samples can be reconstructed from the compressed data. There are two types of sample compression schemes: labeled, see [35, 51] and unlabeled, see [7, 34, 49]. A labeled compression scheme of size k compresses every sample of C to a labeled subsample of size at most k and an unlabeled compression scheme of size k compresses every sample of C to a subset of size at most k of the domain of the sample (see the end of the introduction for precise definitions). The Vapnik-Chervonenkis dimension (VC-dimension) of a set system, was introduced by [69] as a complexity measure of set systems. VC-dimension is central in PAC-learning and plays an important role in combinatorics, algorithmics, discrete geometry, and combinatorial optimization. In particular, it coincides with the rank in the theory of (complexes of) oriented matroids. Furthermore, within machine learning and closely tied to the topic of this paper, the sample compression conjecture of [35] and [51] states that any set system of VC-dimension d has a labeled sample compression scheme of size O(d). This question remains one of the oldest open problems in computational learning theory.

For classification problems with significant class imbalance, subsampling can reduce computational costs at the price of inflated variance in estimating model parameters. We propose a method for subsampling efficiently for logistic regression by adjusting the class balance locally in feature space via an accept-reject scheme. Our method generalizes standard case-control sampling, using a pilot estimate to preferentially select examples whose responses are conditionally rare given their features. The biased subsampling is corrected by a post-hoc analytic adjustment to the parameters. The method is simple and requires one parallelizable scan over the full data set. Standard case-control sampling is inconsistent under model misspecification for the population risk-minimizing coefficients $\theta^*$. By contrast, our estimator is consistent for $\theta^*$ provided that the pilot estimate is. Moreover, under correct specification and with a consistent, independent pilot estimate, our estimator has exactly twice the asymptotic variance of the full-sample MLE - even if the selected subsample comprises a miniscule fraction of the full data set, as happens when the original data are severely imbalanced. The factor of two improves to $1+\frac{1}{c}$ if we multiply the baseline acceptance probabilities by $c>1$ (and weight points with acceptance probability greater than 1), taking roughly $\frac{1+c}{2}$ times as many data points into the subsample. Experiments on simulated and real data show that our method can substantially outperform standard case-control subsampling.

Inductive learning is based on inferring a general rule from a finite data set and using it to label new data. In transduction one attempts to solve the problem of using a labeled training set to label a set of unlabeled points, which are given to the learner prior to learning. Although transduction seems at the outset to be an easier task than induction, there have not been many provably useful algorithms for transduction. Moreover, the precise relation between induction and transduction has not yet been determined. The main theoretical developments related to transduction were presented by Vapnik more than twenty years ago. One of Vapnik's basic results is a rather tight error bound for transductive classification based on an exact computation of the hypergeometric tail. While tight, this bound is given implicitly via a computational routine. Our first contribution is a somewhat looser but explicit characterization of a slightly extended PAC-Bayesian version of Vapnik's transductive bound. This characterization is obtained using concentration inequalities for the tail of sums of random variables obtained by sampling without replacement. We then derive error bounds for compression schemes such as (transductive) support vector machines and for transduction algorithms based on clustering. The main observation used for deriving these new error bounds and algorithms is that the unlabeled test points, which in the transductive setting are known in advance, can be used in order to construct useful data dependent prior distributions over the hypothesis space.

This paper is concerned with transductive learning. Although transduction appears to be an easier task than induction, there have not been many provably useful algorithms and bounds for transduction. We present explicit error bounds for transduction and derive a general technique for devising bounds within this setting. The technique is applied to derive error bounds for compression schemes such as (transductive) SVMs and for transduction algorithms based on clustering.

This paper is concerned with transductive learning. Although transduction appears to be an easier task than induction, there have not been many provably useful algorithms and bounds for transduction. We present explicit error bounds for transduction and derive a general technique for devising bounds within this setting. The technique is applied to derive error bounds for compression schemes such as (transductive) SVMs and for transduction algorithms based on clustering.

This paper is concerned with transductive learning. Although transduction appears to be an easier task than induction, there have not been many provably useful algorithms and bounds for transduction. We present explicit error bounds for transduction and derive a general technique for devising bounds within this setting. The technique is applied to derive error bounds for compression schemes such as (transductive) SVMs and for transduction algorithms based on clustering.