Fine-tuning Large Language Models (LLMs) on specific datasets is a common practice to improve performance on target tasks. However, this performance gain often leads to overfitting, where the model becomes too specialized in either the task or the characteristics of the training data, resulting in a loss of generalization. This paper introduces Selective Self-Rehearsal (SSR), a fine-tuning approach that achieves performance comparable to the standard supervised fine-tuning (SFT) while improving generalization. SSR leverages the fact that there can be multiple valid responses to a query. By utilizing the model's correct responses, SSR reduces model specialization during the fine-tuning stage. SSR first identifies the correct model responses from the training set by deploying an appropriate LLM as a judge. Then, it fine-tunes the model using the correct model responses and the gold response for the remaining samples. The effectiveness of SSR is demonstrated through experiments on the task of identifying unanswerable queries across various datasets. The results show that standard SFT can lead to an average performance drop of up to $16.7\%$ on multiple benchmarks, such as MMLU and TruthfulQA. In contrast, SSR results in close to $2\%$ drop on average, indicating better generalization capabilities compared to standard SFT.
It is important for official statistics production to apply ML with statistical rigor, as it presents both opportunities and challenges. Although machine learning has enjoyed rapid technological advances in recent years, its application does not possess the methodological robustness necessary to produce high quality statistical results. In order to account for all sources of error in machine learning models, the Total Machine Learning Error (TMLE) is presented as a framework analogous to the Total Survey Error Model used in survey methodology. As a means of ensuring that ML models are both internally valid as well as externally valid, the TMLE model addresses issues such as representativeness and measurement errors. There are several case studies presented, illustrating the importance of applying more rigor to the application of machine learning in official statistics.
Increasingly, large language models (LLMs) are being used to automate workplace processes requiring a high degree of creativity. While much prior work has examined the creativity of LLMs, there has been little research on whether they can generate valid creativity assessments for humans despite the increasingly central role of creativity in modern economies. We develop a psychometrically inspired framework for creating test items (questions) for a classic free-response creativity test: the creative problem-solving (CPS) task. Our framework, the creative psychometric item generator (CPIG), uses a mixture of LLM-based item generators and evaluators to iteratively develop new prompts for writing CPS items, such that items from later iterations will elicit more creative responses from test takers. We find strong empirical evidence that CPIG generates valid and reliable items and that this effect is not attributable to known biases in the evaluation process. Our findings have implications for employing LLMs to automatically generate valid and reliable creativity tests for humans and AI.
Fuzzing is an important dynamic program analysis technique designed for finding vulnerabilities in complex software. Fuzzing involves presenting a target program with crafted malicious input to cause crashes, buffer overflows, memory errors, and exceptions. Crafting malicious inputs in an efficient manner is a difficult open problem and the best approaches often apply uniform random mutations to pre-existing valid inputs. In this work, we propose to adopt fine-tuned large language models (FuzzCoder) to learn patterns in the input files from successful attacks to guide future fuzzing explorations. Specifically, we develop a framework to leverage the code LLMs to guide the mutation process of inputs in fuzzing. The mutation process is formulated as the sequence-to-sequence modeling, where LLM receives a sequence of bytes and then outputs the mutated byte sequence. FuzzCoder is fine-tuned on the created instruction dataset (Fuzz-Instruct), where the successful fuzzing history is collected from the heuristic fuzzing tool. FuzzCoder can predict mutation locations and strategies locations in input files to trigger abnormal behaviors of the program. Experimental results show that FuzzCoder based on AFL (American Fuzzy Lop) gain significant improvements in terms of effective proportion of mutation (EPM) and number of crashes (NC) for various input formats including ELF, JPG, MP3, and XML.
In various domains, the increasing application of machine learning allows researchers to access inexpensive predictive data, which can be utilized as auxiliary data for statistical inference. Although such data are often unreliable compared to gold-standard datasets, Prediction-Powered Inference (PPI) has been proposed to ensure statistical validity despite the unreliability. However, the challenge of `data silos' arises when the private gold-standard datasets are non-shareable for model training, leading to less accurate predictive models and invalid inferences. In this paper, we introduces the Federated Prediction-Powered Inference (Fed-PPI) framework, which addresses this challenge by enabling decentralized experimental data to contribute to statistically valid conclusions without sharing private information. The Fed-PPI framework involves training local models on private data, aggregating them through Federated Learning (FL), and deriving confidence intervals using PPI computation. The proposed framework is evaluated through experiments, demonstrating its effectiveness in producing valid confidence intervals.
A natural strategy for continual learning is to weigh a Bayesian ensemble of fixed functions. This suggests that if a (single) neural network could be interpreted as an ensemble, one could design effective algorithms that learn without forgetting. To realize this possibility, we observe that a neural network classifier with N parameters can be interpreted as a weighted ensemble of N classifiers, and that in the lazy regime limit these classifiers are fixed throughout learning. We call these classifiers the neural tangent experts and show they output valid probability distributions over the labels. We then derive the likelihood and posterior probability of each expert given past data. Surprisingly, the posterior updates for these experts are equivalent to a scaled and projected form of stochastic gradient descent (SGD) over the network weights. Away from the lazy regime, networks can be seen as ensembles of adaptive experts which improve over time. These results offer a new interpretation of neural networks as Bayesian ensembles of experts, providing a principled framework for understanding and mitigating catastrophic forgetting in continual learning settings.
Accurate prediction of the properties of crystalline materials is crucial for targeted discovery, and this prediction is increasingly done with data-driven models. However, for many properties of interest, the number of materials for which a specific property has been determined is much smaller than the number of known materials. To overcome this disparity, we propose a novel self-supervised learning (SSL) strategy for material property prediction. Our approach, crystal denoising self-supervised learning (CDSSL), pretrains predictive models (e.g., graph networks) with a pretext task based on recovering valid material structures when given perturbed versions of these structures. We demonstrate that CDSSL models out-perform models trained without SSL, across material types, properties, and dataset sizes.
We develop an estimator for treatment effects in high-dimensional settings with additive measurement error, a prevalent challenge in modern econometrics. We introduce the Double/Debiased Convex Conditioned LASSO (Double/Debiased CoCoLASSO), which extends the double/debiased machine learning framework to accommodate mismeasured covariates. Our principal contributions are threefold. (1) We construct a Neyman-orthogonal score function that remains valid under measurement error, incorporating a bias correction term to account for error-induced correlations. (2) We propose a method of moments estimator for the measurement error variance, enabling implementation without prior knowledge of the error covariance structure. (3) We establish the $\sqrt{N}$-consistency and asymptotic normality of our estimator under general conditions, allowing for both the number of covariates and the magnitude of measurement error to increase with the sample size. Our theoretical results demonstrate the estimator's efficiency within the class of regularized high-dimensional estimators accounting for measurement error. Monte Carlo simulations corroborate our asymptotic theory and illustrate the estimator's robust performance across various levels of measurement error. Notably, our covariance-oblivious approach nearly matches the efficiency of methods that assume known error variance.
This paper presents a comprehensive analysis of how AI-generated medical responses are perceived and evaluated by non-experts. A total of 300 participants gave evaluations for medical responses that were either written by a medical doctor on an online healthcare platform, or generated by a large language model and labeled by physicians as having high or low accuracy. Results showed that participants could not effectively distinguish between AI-generated and Doctors' responses and demonstrated a preference for AI-generated responses, rating High Accuracy AI-generated responses as significantly more valid, trustworthy, and complete/satisfactory. Low Accuracy AI-generated responses on average performed very similar to Doctors' responses, if not more. Participants not only found these low-accuracy AI-generated responses to be valid, trustworthy, and complete/satisfactory but also indicated a high tendency to follow the potentially harmful medical advice and incorrectly seek unnecessary medical attention as a result of the response provided. This problematic reaction was comparable if not more to the reaction they displayed towards doctors' responses. This increased trust placed on inaccurate or inappropriate AI-generated medical advice can lead to misdiagnosis and harmful consequences for individuals seeking help. Further, participants were more trusting of High Accuracy AI-generated responses when told they were given by a doctor and experts rated AI-generated responses significantly higher when the source of the response was unknown. Both experts and non-experts exhibited bias, finding AI-generated responses to be more thorough and accurate than Doctors' responses but still valuing the involvement of a Doctor in the delivery of their medical advice. Ensuring AI systems are implemented with medical professionals should be the future of using AI for the delivery of medical advice.
EVINCE (Entropy and Variation IN Conditional Exchanges) is a novel framework for optimizing multi-LLM dialogues using conditional statistics and information theory. It addresses limitations in multi-agent debate (MAS) frameworks, where multiple LLMs ``chat'' without behavior modulation or mutual information quality assessment. Using dual entropy optimization to balance perspective diversity and prior knowledge, $\EVINCE$ provides quantitative tools to dynamically regulate LLM linguistic behaviors. When mutual information is low and both cross-entropy and Wasserstein distance are high, EVINCE promotes contentious dialogues to expose diverse perspectives and uncover inconsistencies. Conversely, as cross-entropy decreases and mutual information stabilizes, it transitions discussions into a conciliatory phase, encouraging compromise and acknowledgment of valid points. Using information-theoretic metrics and optimizing mutual information, $\EVINCE$ emerges as a structured and highly effective framework for multi-LLM collaboration.
Large language models (LLMs) have shown high agreement with human raters across a variety of tasks, demonstrating potential to ease the challenges of human data collection. In computational social science (CSS), researchers are increasingly leveraging LLM annotations to complement slow and expensive human annotations. Still, guidelines for collecting and using LLM annotations, without compromising the validity of downstream conclusions, remain limited. We introduce Confidence-Driven Inference: a method that combines LLM annotations and LLM confidence indicators to strategically select which human annotations should be collected, with the goal of producing accurate statistical estimates and provably valid confidence intervals while reducing the number of human annotations needed. Our approach comes with safeguards against LLM annotations of poor quality, guaranteeing that the conclusions will be both valid and no less accurate than if we only relied on human annotations. We demonstrate the effectiveness of Confidence-Driven Inference over baselines in statistical estimation tasks across three CSS settings--text politeness, stance, and bias--reducing the needed number of human annotations by over 25% in each. Although we use CSS settings for demonstration, Confidence-Driven Inference can be used to estimate most standard quantities across a broad range of NLP problems.
Mental models play an important role in whether user interaction with intelligent systems, such as dialog systems is successful or not. Adaptive dialog systems present the opportunity to align a dialog agent's behavior with heterogeneous user expectations. However, there has been little research into what mental models users form when interacting with a task-oriented dialog system, how these models affect users' interactions, or what role system adaptation can play in this process, making it challenging to avoid damage to human-AI partnership. In this work, we collect a new publicly available dataset for exploring user mental models about information seeking dialog systems. We demonstrate that users have a variety of conflicting mental models about such systems, the validity of which directly impacts the success of their interactions and perceived usability of system. Furthermore, we show that adapting a dialog agent's behavior to better align with users' mental models, even when done implicitly, can improve perceived usability, dialog efficiency, and success. To this end, we argue that implicit adaptation can be a valid strategy for task-oriented dialog systems, so long as developers first have a solid understanding of users' mental models.
Graph Neural Networks (GNNs) have emerged as powerful tools for predicting outcomes in graph-structured data. However, a notable limitation of GNNs is their inability to provide robust uncertainty estimates, which undermines their reliability in contexts where errors are costly. One way to address this issue is by providing prediction sets that contain the true label with a predefined probability margin. Our approach builds upon conformal prediction (CP), a framework that promises to construct statistically robust prediction sets or intervals. There are two primary challenges: first, given dependent data like graphs, it is unclear whether the critical assumption in CP - exchangeability - still holds when applied to node classification. Second, even if the exchangeability assumption is valid for conformalized link prediction, we need to ensure high efficiency, i.e., the resulting prediction set or the interval length is small enough to provide useful information. In this article, we propose a novel approach termed Robust Conformal Prediction for GNNs (RoCP-GNN), which integrates conformal prediction (CP) directly into the GNN training process. This method generates prediction sets, instead of just point predictions, that are valid at a user-defined confidence level, assuming only exchangeability. Our approach robustly predicts outcomes with any predictive GNN model while quantifying the uncertainty in predictions within the realm of graph-based semi-supervised learning (SSL). Experimental results demonstrate that GNN models with size loss provide a statistically significant increase in performance. We validate our approach on standard graph benchmark datasets by coupling it with various state-of-the-art GNNs in node classification. The code will be made available after publication.
Lagrangian decomposition (LD) is a relaxation method that provides a dual bound for constrained optimization problems by decomposing them into more manageable sub-problems. This bound can be used in branch-and-bound algorithms to prune the search space effectively. In brief, a vector of Lagrangian multipliers is associated with each sub-problem, and an iterative procedure (e.g., a sub-gradient optimization) adjusts these multipliers to find the tightest bound. Initially applied to integer programming, Lagrangian decomposition also had success in constraint programming due to its versatility and the fact that global constraints provide natural sub-problems. However, the non-linear and combinatorial nature of sub-problems in constraint programming makes it computationally intensive to optimize the Lagrangian multipliers with sub-gradient methods at each node of the tree search. This currently limits the practicality of LD as a general bounding mechanism for constraint programming. To address this challenge, we propose a self-supervised learning approach that leverages neural networks to generate multipliers directly, yielding tight bounds. This approach significantly reduces the number of sub-gradient optimization steps required, enhancing the pruning efficiency and reducing the execution time of constraint programming solvers. This contribution is one of the few that leverage learning to enhance bounding mechanisms on the dual side, a critical element in the design of combinatorial solvers. To our knowledge, this work presents the first generic method for learning valid dual bounds in constraint programming.
This paper considers inference in a linear instrumental variable regression model with many potentially weak instruments, in the presence of heterogeneous treatment effects. I first show that existing test procedures, including those that are robust to either weak instruments or heterogeneous treatment effects, can be arbitrarily oversized. I propose a novel and valid test based on a score statistic and a ``leave-three-out" variance estimator. In the presence of heterogeneity and within the class of tests that are functions of the leave-one-out analog of a maximal invariant, this test is asymptotically the uniformly most powerful unbiased test. In two applications to judge and quarter-of-birth instruments, the proposed inference procedure also yields a bounded confidence set while some existing methods yield unbounded or empty confidence sets.
We present a safety-critical controller for the problem of stabilization for force-controlled nonholonomic mobile robots. The proposed control law is based on the constructions of control Lyapunov functions (CLFs) and control barrier functions (CBFs) for cascaded systems. To address nonholonomicity, we design the nominal controller that guarantees global asymptotic stability and local exponential stability for the closed-loop system in polar coordinates and construct a strict Lyapunov function valid on any compact sets. Furthermore, we present a procedure for constructing CBFs for cascaded systems, utilizing the CBF of the kinematic model through integrator backstepping. Quadratic programming is employed to combine CLFs and CBFs to integrate both stability and safety in the closed loop. The proposed control law is time-invariant, continuous along trajectories, and easy to implement. Our main results guarantee both safety and local asymptotic stability for the closed-loop system.
Video quality assessment (VQA) is a crucial task in the development of video compression standards, as it directly impacts the viewer experience. This paper presents the results of the Compressed Video Quality Assessment challenge, held in conjunction with the Advances in Image Manipulation (AIM) workshop at ECCV 2024. The challenge aimed to evaluate the performance of VQA methods on a diverse dataset of 459 videos, encoded with 14 codecs of various compression standards (AVC/H.264, HEVC/H.265, AV1, and VVC/H.266) and containing a comprehensive collection of compression artifacts. To measure the methods performance, we employed traditional correlation coefficients between their predictions and subjective scores, which were collected via large-scale crowdsourced pairwise human comparisons. For training purposes, participants were provided with the Compressed Video Quality Assessment Dataset (CVQAD), a previously developed dataset of 1022 videos. Up to 30 participating teams registered for the challenge, while we report the results of 6 teams, which submitted valid final solutions and code for reproducing the results. Moreover, we calculated and present the performance of state-of-the-art VQA methods on the developed dataset, providing a comprehensive benchmark for future research. The dataset, results, and online leaderboard are publicly available at https://challenges.videoprocessing.ai/challenges/compressedvideo-quality-assessment.html.
Consider a scenario where a harmfulness detection metric is employed by a system to filter unsafe responses generated by a Large Language Model. When analyzing individual harmful and unethical prompt-response pairs, the metric correctly classifies each pair as highly unsafe, assigning the highest score. However, when these same prompts and responses are concatenated, the metric's decision flips, assigning the lowest possible score, thereby misclassifying the content as safe and allowing it to bypass the filter. In this study, we discovered that several harmfulness LLM-based metrics, including GPT-based, exhibit this decision-flipping phenomenon. Additionally, we found that even an advanced metric like GPT-4o is highly sensitive to input order. Specifically, it tends to classify responses as safe if the safe content appears first, regardless of any harmful content that follows, and vice versa. This work introduces automatic concatenation-based tests to assess the fundamental properties a valid metric should satisfy. We applied these tests in a model safety scenario to assess the reliability of harmfulness detection metrics, uncovering a number of inconsistencies.
This paper characterizes the optimal capacity-distortion (C-D) tradeoff in an optical point-to-point system with single-input single-output (SISO) for communication and single-input multiple-output (SIMO) for sensing within an integrated sensing and communication (ISAC) framework. We consider the optimal rate-distortion (R-D) region and explore several inner (IB) and outer bounds (OB). We introduce practical, asymptotically optimal maximum a posteriori (MAP) and maximum likelihood estimators (MLE) for target distance, addressing nonlinear measurement-to-state relationships and non-conjugate priors. As the number of sensing antennas increases, these estimators converge to the Bayesian Cram\'er-Rao bound (BCRB). We also establish that the achievable rate-Cram\'er-Rao bound (R-CRB) serves as an OB for the optimal C-D region, valid for both unbiased estimators and asymptotically large numbers of receive antennas. To clarify that the input distribution determines the tradeoff across the Pareto boundary of the C-D region, we propose two algorithms: i) an iterative Blahut-Arimoto algorithm (BAA)-type method, and ii) a memory-efficient closed-form (CF) approach. The CF approach includes a CF optimal distribution for high optical signal-to-noise ratio (O-SNR) conditions. Additionally, we adapt and refine the deterministic-random tradeoff (DRT) to this optical ISAC context.
We present a probabilistic model for stochastic iterative algorithms with the use case of optimization algorithms in mind. Based on this model, we present PAC-Bayesian generalization bounds for functions that are defined on the trajectory of the learned algorithm, for example, the expected (non-asymptotic) convergence rate and the expected time to reach the stopping criterion. Thus, not only does this model allow for learning stochastic algorithms based on their empirical performance, it also yields results about their actual convergence rate and their actual convergence time. We stress that, since the model is valid in a more general setting than learning-to-optimize, it is of interest for other fields of application, too. Finally, we conduct five practically relevant experiments, showing the validity of our claims.