We propose a new analysis framework for clustering M items into an unknown number of K distinct groups using noisy and actively collected responses. At each time step, an agent is allowed to query pairs of items and observe bandit binary feedback. If the pair of items belongs to the same (resp. different) cluster, the observed feedback is 1 with probability p>1/2 (resp. q<1/2). Leveraging the ubiquitous change-of-measure technique, we establish a fundamental lower bound on the expected number of queries needed to achieve a desired confidence in the clustering accuracy, formulated as a sup-inf optimization problem. Building on this theoretical foundation, we design an asymptotically optimal algorithm in which the stopping criterion involves an empirical version of the inner infimum – the Generalized Likelihood Ratio (GLR) statistic – being compared to a threshold. We develop a computationally feasible variant of the GLR statistic and show that its performance gap to the lower bound can be accurately empirically estimated and remains within a constant multiple of the lower bound.

Language models (LMs) are increasingly extended with new learnable vocabulary tokens for domain-specific tasks, such as Semantic-ID tokens in generative recommendation. The standard practice initializes these new tokens as the mean of existing vocabulary embeddings, then relies on supervised fine-tuning to learn their representations. We present a systematic analysis of this strategy: through spectral and geometric diagnostics, we show that mean initialization collapses all new tokens into a degenerate subspace, erasing inter-token distinctions that subsequent fine-tuning struggles to fully recover. These findings suggest that token initialization is a key bottleneck when extending LMs with new vocabularies. Motivated by this diagnosis, we propose the Grounded Token Initialization Hypothesis: linguistically grounding novel tokens in the pretrained embedding space before fine-tuning better enables the model to leverage its general-purpose knowledge for novel-token domains. We operationalize this hypothesis as GTI (Grounded Token Initialization), a lightweight grounding stage that, prior to fine-tuning, maps new tokens to distinct, semantically meaningful locations in the pretrained embedding space using only paired linguistic supervision. Despite its simplicity, GTI outperforms both mean initialization and existing auxiliary-task adaptation methods in the majority of evaluation settings across multiple generative recommendation benchmarks, including industry-scale and public datasets. Further analyses show that grounded embeddings produce richer inter-token structure that persists through fine-tuning, corroborating the hypothesis that initialization quality is a key bottleneck in vocabulary extension.

Large language models (LLMs) frequently hallucinate, limiting their reliability in knowledge-intensive applications. Retrieval-augmented generation (RAG) and conformal factuality have emerged as potential ways to address this limitation. While RAG aims to ground responses in retrieved evidence, it provides no statistical guarantee that the final output is correct. Conformal factuality filtering offers distribution-free statistical reliability by scoring and filtering atomic claims using a threshold calibrated on held-out data, however, the informativeness of the final output is not guaranteed. We systematically analyze the reliability and usefulness of conformal factuality for RAG-based LLMs across generation, scoring, calibration, robustness, and efficiency. We propose novel informativeness-aware metrics that better reflect task utility under conformal filtering. Across three benchmarks and multiple model families, we find that (i) conformal filtering suffers from low usefulness at high factuality levels due to vacuous outputs, (ii) conformal factuality guarantee is not robust to distribution shifts and distractors, highlighting the limitation that requires calibration data to closely match deployment conditions, and (iii) lightweight entailment-based verifiers match or outperform LLM-based model confidence scorers while requiring over 100× fewer FLOPs. Overall, our results expose factuality-informativeness trade-offs and fragility of conformal filtering framework under distribution shifts and distractors, highlighting the need for new approaches for reliability with robustness and usefulness as key metrics, and provide actionable guidance for building RAG pipelines that are both reliable and computationally efficient.

The advent of always-on personal AI assistants, enabled by all-day wearable devices such as smart glasses, demands a new level of contextual understanding, one that goes beyond short, isolated events to encompass the continuous, longitudinal stream of egocentric video. Achieving this vision requires advances in long-horizon video understanding, where systems must interpret and recall visual and audio information spanning days or even weeks. Existing methods, including large language models and retrieval-augmented generation, are constrained by limited context windows and lack the ability to perform compositional, multi-hop reasoning over very long video streams. In this work, we address these challenges through EGAgent, an enhanced agentic framework centered on entity scene graphs, which represent people, places, objects, and their relationships over time. Our system equips a planning agent with tools for structured search and reasoning over these graphs, as well as hybrid visual and audio search capabilities, enabling detailed, cross-modal, and temporally coherent reasoning. Experiments on the EgoLifeQA and Video-MME (Long) datasets show that our method achieves state-of-the-art performance on EgoLifeQA (57.5%) and competitive performance on Video-MME (Long) (74.1%) for complex longitudinal video understanding tasks.

This paper investigates metric learning in a Reproducing Kernel Hilbert Space (RKHS) based on a set of random triplet comparisons in the form of "Do you think item h is more similar to item i or item j?" indicating similarity and differences between various items. The goal is to learn a metric in the RKHS that reflects the comparisons. Nonlinear metric learning using kernel methods and neural networks has shown great empirical promise. While previous works have addressed certain aspects of this problem, there is little or no theoretical understanding of such methods. The exception is the special (linear) case in which the RKHS is the standard d-dimensional Euclidean space; there is a comprehensive theory for metric learning in the d-dimensional Euclidean space. This paper develops a general RKHS framework for metric learning and provides novel generalization guarantees and sample complexity bounds. We validate our findings through a set of simulations and experiments on real datasets. Our code is publicly available at https://github.com/RamyaLab/ metric-learning-RKHS.

Crowdsourced clustering leverages human input to group items into clusters. The design of tasks for crowdworkers, specifically the number of items presented per query, impacts answer quality and cognitive load. This work investigates the trade-off between query size and answer accuracy, revealing diminishing returns beyond 4-5 items per query. Crucially, we identify contextual bias in crowdworker responses – the likelihood of grouping items depends not only on their similarity but also on the other items present in the query. This structured noise contradicts assumptions made in existing noise models. Our findings underscore the need for more nuanced noise models that account for the complex interplay between items and query context in crowdsourced clustering tasks.

In lifelong learning, a learner faces a sequence of tasks with shared structure and aims to identify and leverage it to accelerate learning. We study the setting where such structure is captured by a common representation of data. Unlike multi-task learning or learning-to-learn, where tasks are available upfront to learn the representation, lifelong learning requires the learner to make use of its existing knowledge while continually gathering partial information in an online fashion. In this paper, we consider a generalized framework of lifelong representation learning. We propose a simple algorithm that uses multi-task empirical risk minimization as a subroutine and establish a sample complexity bound based on a new notion we introduce–the task-eluder dimension. Our result applies to a wide range of learning problems involving general function classes. As concrete examples, we instantiate our result on classification and regression tasks under noise.

Machine Learning (ML) models are trained on in-distribution (ID) data but often encounter out-of-distribution (OOD) inputs during deployment—posing serious risks in safety-critical domains. Recent works have focused on designing scoring functions to quantify OOD uncertainty, with score thresholds typically set based solely on ID data to achieve a target true positive rate (TPR), since OOD data is limited before deployment. However, these TPR-based thresholds leave false positive rates (FPR) uncontrolled, often resulting in high FPRs where OOD points are misclassified as ID. Moreover, fixed scoring functions and thresholds lack the adaptivity needed to handle newly observed, evolving OOD inputs, leading to sub-optimal performance. To address these challenges, we propose ASAT, a human-in-the-loop framework that safely updates both scoring functions and thresholds on the fly based on real-world OOD inputs. ASAT maximizes TPR while controlling FPR at all times under stationary conditions, even as the system adapts over time. Under nonstationary conditions, the method adapts to distribution shifts with only transient FPR violations during the adaptation period. We provide theoretical guarantees for FPR control under stationary conditions and present extensive empirical evaluations on OpenOOD benchmarks to demonstrate that our approach outperforms existing methods by achieving higher TPRs while maintaining FPR control.

Popular text-to-image (T2I) systems are trained on web-scraped data, which is heavily Amero and Euro-centric, underrepresenting the cultures of the Global South. To analyze these biases, we introduce CuRe, a novel and scalable benchmarking and scoring suite for cultural representativeness that leverages the marginal utility of attribute specification to T2I systems as a proxy for human judgments. Our CuRe benchmark dataset has a novel categorical hierarchy built from the crowdsourced Wikimedia knowledge graph, with 300 cultural artifacts across 32 cultural subcategories grouped into six broad cultural axes (food, art, fashion, architecture, celebrations, and people). Our dataset's categorical hierarchy enables CuRe scorers to evaluate T2I systems by analyzing their response to increasing the informativeness of text conditioning, enabling fine-grained cultural comparisons. We empirically observe much stronger correlations of our class of scorers to human judgments of perceptual similarity, image-text alignment, and cultural diversity across image encoders (SigLIP 2, AIMV2 and DINOv2), vision-language models (OpenCLIP, SigLIP 2, Gemini 2.0 Flash) and state-of-the-art text-to-image systems, including three variants of Stable Diffusion (1.5, XL, 3.5 Large), FLUX.1 [dev], Ideogram 2.0, and DALL-E 3. The code and dataset is open-sourced and available at https://aniketrege.github.io/cure/.

We study metric learning from preference comparisons under the ideal point model, in which a user prefers an item over another if it is closer to their latent ideal item. These items are embedded into $\mathbb{R}^d$ equipped with an unknown Mahalanobis distance shared across users. While recent work shows that it is possible to simultaneously recover the metric and ideal items given $\mathcal{O}(d)$ pairwise comparisons per user, in practice we often have a limited budget of $o(d)$ comparisons. We study whether the metric can still be recovered, even though learning individual ideal items is now no longer possible. We show that, on the one hand, $o(d)$ comparisons may not reveal any information about the metric, even with infinitely many users. On the other hand, when comparisons are made over items that exhibit low-dimensional structure, each user can contribute to learning the metric restricted to a low-dimensional subspace so that the metric can be jointly identified. We present a divide-and-conquer approach that achieves this, and provide theoretical recovery guarantees and empirical validation.

This article investigates the theoretical and empirical performance of Fisher-Pitman-type permutation tests for assessing the equality of unknown Poisson mixture distributions. Building on nonparametric maximum likelihood estimators (NPMLEs) of the mixing distribution, these tests are theoretically shown to be able to adapt to complicated unspecified structures of count data and also consistent against their corresponding ANOVA-type alternatives; the latter is a result in parallel to classic claims made by Robinson. The studied methods are then applied to a single-cell RNA-seq data obtained from different cell types from brain samples of autism subjects and healthy controls; empirically, they unveil genes that are differentially expressed between autism and control subjects yet are missed using common tests. For justifying their use, rate optimality of NPMLEs is also established in settings similar to nonparametric Gaussian (Wu and Yang) and binomial mixtures (Tian, Kong, and Valiant; Vinayak et al.). Supplementary materials for this article are available online.

Robustness to out-of-distribution (OOD) samples is crucial for safely deploying machine learning models in the open world. Recent works have focused on designing scoring functions to quantify OOD uncertainty. Setting appropriate thresholds for these scoring functions for OOD detection is challenging as OOD samples are often unavailable up front. Typically, thresholds are set to achieve a desired true positive rate (TPR), e.g., 95% TPR. However, this can lead to very high false positive rates (FPR), ranging from 60 to 96%, as observed in the Open-OOD benchmark. In safety-critical real-life applications, e.g., medical diagnosis, controlling the FPR is essential when dealing with various OOD samples dynamically. To address these challenges, we propose a mathematically grounded OOD detection framework that leverages expert feedback to safely update the threshold on the fly. We provide theoretical results showing that it is guaranteed to meet the FPR constraint at all times while minimizing the use of human feedback. Another key feature of our framework is that it can work with any scoring function for OOD uncertainty quantification. Empirical evaluation of our system on synthetic and benchmark OOD datasets shows that our method can maintain FPR at most 5% while maximizing TPR.

Face recognition (FR) systems are widely deployed in critical applications, making their reliability and robustness across diverse populations and conditions essential. Standard evaluation of FR systems typically relies on datasets such as LFW to estimate average recognition accuracy. Some benchmarks also capture coarse-grained intra-identity variations such as aging, pose, and lighting. However, human faces undergo more fine-grained changes, including appearance changes such as hairstyles and makeup, that are underrepresented in existing benchmarks. Counterfactual evaluation provides a method to assess FR robustness under such fine-grained variations. Existing counterfactual face datasets synthesized with image generators, however, are limited in attribute coverage due to the use of humans for verification in the pipeline. We propose CounterFace, a new counterfactual evaluation dataset comprising 20 facial attributes and 8 demographic factors, exceeding prior synthetic face datasets by 14 attributes and 2 demographics. The dataset is generated using a fully automated pipeline based on off-the-shelf image generators with custom verifiers, removing human need for verification. CounterFace contains 11,821 counterfactual face pairs, and a post-hoc user study confirms the faithfulness of the generated counterfactuals. We evaluate two commercial and four open-source FR systems (AWS Rekognition, Face++, AdaFace, MagFace, ArcFace, FaceNet) across 160 attribute-demographic combinations. Our dataset helps in the isolation of precise failure modes for individual systems unlike standard evaluation benchmarks. Results indicate that the performance degradation varies across attributes and demographics for all six systems and occluding attributes (e.g., facemask and facial hair) universally degrade performance.

Auto-labeling is an important family of techniques that produce labeled training sets with minimum manual annotation. A prominent variant, threshold-based auto-labeling (TBAL), works by finding thresholds on a model's confidence scores above which it can accurately automatically label unlabeled data. However, many models are known to produce overconfident scores, leading to poor TBAL performance. While a natural idea is to apply off-the-shelf calibration methods to alleviate the overconfidence issue, we show that such methods fall short. Rather than experimenting with ad-hoc choices of confidence functions, we propose a framework for studying the optimal TBAL confidence function. We develop a tractable version of the framework to obtain Colander (Confidence functions for Efficient and Reliable Auto-labeling), a new post-hoc method specifically designed to maximize performance in TBAL systems. We perform an extensive empirical evaluation of Colander and compare it against methods designed for calibration. Colander achieves up to 60% improvement on coverage over the baselines while maintaining error level below 5% and using the same amount of labeled data.

Text-to-image diffusion models have achieved widespread popularity due to their unprecedented image generation capability. In particular, their ability to synthesize and modify human faces has spurred research into using generated face images in both training data augmentation and model performance assessments. In this paper, we study the efficacy and shortcomings of generative models in the context of face generation. Utilizing a combination of qualitative and quantitative measures, including embedding-based metrics and user studies, we present a framework to audit the characteristics of generated faces conditioned on a set of social attributes. We applied our framework on faces generated through state-of-the-art text-to-image diffusion models. We identify several limitations of face image generation that include faithfulness to the text prompt, demographic disparities, and distributional shifts. Furthermore, we present an analytical model that provides insights into how training data selection contributes to the performance of generative models. Our survey data and analytics code can be found online at https://github.com/wi-pi/Limitations of Face Generation

Using triplet comparison queries of the form “Do you think item a is more similar to item b or item c?” to learn a positive definite matrix to capture a distance metric in Rd has been a popular approach to capture how human perceive similarity and differences between various objects/concepts. Most of the existing works focus on learning a single metric using data from all people in the dataset. However, people can systematically differ in their notions of similarity over a set of objects due to their diverse backgrounds. Therefore, using a single metric for everyone has limited capacity in capturing the heterogeneity while modeling how people perceive objects in populations that have diverse subgroups. The subgroup structure is often salient and difficult to know a priori. We propose to learn the subgroup structure from the answers to triplet queries by clustering the user-triplet observation matrix. By modeling the problem of metric clustering as a low-rank matrix recovery problem, we leverage convex optimization based approach to perform clustering. We provide analysis for two cluster case that sheds light on when the approach succeeds and fails as function of distance between the metrics, size of the clusters, number of triplet queries answered per person and the noise level in the answers obtained. We validate our results through extensive simulations. Furthermore, we also provide analysis that shows how an outlier impacts the discovery of cluster structure.

Large foundation models pretrained on raw web-scale data are not readily deployable without additional step of extensive alignment to human preferences. Such alignment is typically done by collecting large amounts of pairwise comparisons from humans ("Do you prefer output A or B?") and learning a reward model or a policy with the Bradley-Terry-Luce (BTL) model as a proxy for a human's underlying implicit preferences. These methods generally suffer from assuming a universal preference shared by all humans, which lacks the flexibility of adapting to plurality of opinions and preferences. In this work, we propose PAL, a framework to model human preference complementary to existing pretraining strategies, which incorporates plurality from the ground up. We propose using the ideal point model as a lens to view alignment using preference comparisons. Together with our novel reformulation and using mixture modeling, our framework captures the plurality of population preferences while simultaneously learning a common preference latent space across different preferences, which can few-shot generalize to new, unseen users. Our approach enables us to use the penultimate-layer representation of large foundation models and simple MLP layers to learn reward functions that are on-par with the existing large state-of-the-art reward models, thereby enhancing efficiency of reward modeling significantly. We show that PAL achieves competitive reward model accuracy compared to strong baselines on 1) Language models with Summary dataset ; 2) Image Generative models with Pick-a-Pic dataset ; 3) A new semisynthetic heterogeneous dataset generated using Anthropic Personas. Finally, our experiments also highlight the shortcoming of current preference datasets that are created using rigid rubrics which wash away heterogeneity, and call for more nuanced data collection approaches.