Incident management (IM) is central to the reliability of large-scale microservice systems. Yet manual IM, where on-call engineers examine metrics, logs, and traces is labor-intensive and error-prone in the face of massive and heterogeneous observability data. Existing automated IM approaches often struggle to generalize across systems, provide limited interpretability, and incur high deployment costs, which hinders adoption in practice. In this paper, we present OpsAgent, a lightweight, self-evolving multi-agent system for IM that employs a training-free data processor to convert heterogeneous observability data into structured textual descriptions, along with a multi-agent collaboration framework that makes diagnostic inference transparent and auditable. To support continual capability growth, OpsAgent also introduces a dual self-evolution mechanism that integrates internal model updates with external experience accumulation, thereby closing the deployment loop. Comprehensive experiments on the OPENRCA benchmark demonstrate state-of-the-art performance and show that OpsAgent is generalizable, interpretable, cost-efficient, and self-evolving, making it a practically deployable and sustainable solution for long-term operation in real-world microservice systems. Notably, its deployment in Lenovo's production environment further validates its effectiveness in real-world industrial settings.

Evaluating generative models is challenging because standard metrics often fail to reflect human preferences. Human evaluations are more reliable but costly and noisy, as participants vary in expertise, attention, and diligence. Pairwise comparisons improve consistency, yet aggregating them into overall quality scores requires careful modeling. Bradley-Terry-based methods update item scores from comparisons, but existing approaches either ignore rater variability or lack convergence guarantees, limiting robustness and interpretability. We introduce BBQ, a Bayesian Bradley-Terry variant that explicitly models rater quality, downweighting or removing unreliable participants, and provides guaranteed monotonic likelihood convergence through an Expectation-Maximization algorithm. Empirical results show that BBQ provides efficient inference, well-calibrated uncertainty estimates, and more robust, interpretable rankings compared to baseline Bradley-Terry models, even with noisy or crowdsourced raters. This framework enables more reliable and cost-effective human evaluation of generative models.

We present UGround, a \textbf{U}nified visual \textbf{Ground}ing paradigm that dynamically selects intermediate layers across \textbf{U}nrolled transformers as ``mask as prompt,'' diverging from the prevailing pipeline that leverages the fixed last hidden layer as ``\texttt{<SEG>} as prompt.'' UGround addresses two primary challenges posed by the prevailing paradigm: (1) its reliance on the fixed last hidden layer, which sequentially amplifies cumulative errors arising from layer-by-layer propagation without intermediate correction, and (2) its use of \texttt{<SEG>} as a prompt, which implicitly projects textual embeddings into visual space without explicit spatial cues (e.g., coordinates). Central to UGround is Policy-Prompted Masking, which comprises two key components: Stochastic Skip Connection (SSC) and Mask as Prompt (MasP). SSC is a reinforcement learning policy that, via stochastic sampling, allows each \texttt{<SEG>} token to slide across unrolled transformer layers, enabling dynamic layer selection at which it connects to the vision model (e.g., SAM) in a skip-connection fashion. Given the selected hidden layer, MasP uses the similarity map derived from the \texttt{<SEG>} token and image tokens as a soft logit mask to prompt SAM for mask generation, offering explicit spatial cues through its activation regions. To validate the effectiveness of UGround, we, for the first time, have unified visual grounding within a single framework from an attribute perspective, spanning from traditional refer expression segmentation to newly proposed reasoning segmentation, single-target to multi-target, positive query to false premise (empty target). All code and models are publicly available at https://github.com/rui-qian/UGround.

Diffusion language models, especially masked discrete diffusion models, have achieved great success recently. While there are some theoretical and primary empirical results showing the advantages of latent reasoning with looped transformers or continuous chain-of-thoughts, continuous diffusion models typically underperform their discrete counterparts. In this paper, we argue that diffusion language models do not necessarily need to be in the discrete space. In particular, we prove that continuous diffusion models have stronger expressivity than discrete diffusions and looped transformers. We attribute the contradiction between the theoretical expressiveness and empirical performance to their practical trainability: while continuous diffusion provides intermediate supervision that looped transformers lack, they introduce additional difficulty decoding tokens into the discrete token space from the continuous representation space. We therefore propose Coevolutionary Continuous Discrete Diffusion (CCDD), which defines a joint multimodal diffusion process on the union of a continuous representation space and a discrete token space, leveraging a single model to simultaneously denoise in the joint space. By combining two modalities, CCDD is expressive with rich semantics in the latent space, as well as good trainability and sample quality with the help of explicit discrete tokens. We also propose effective architectures and advanced training/sampling techniques for CCDD, which reveals strong empirical performance in extensive language modeling experiments on real-world tasks.

AdaBoost sequentially fits so-called weak learners to minimize an exponential loss, which penalizes misclassified data points more severely than other loss functions like cross-entropy. Paradoxically, AdaBoost generalizes well in practice as the number of weak learners grows. In the present work, we introduce Penalized Exponential Loss (PENEX), a new formulation of the multi-class exponential loss that is theoretically grounded and, in contrast to the existing formulation, amenable to optimization via first-order methods, making it a practical objective for training neural networks. We demonstrate that PENEX effectively increases margins of data points, which can be translated into a generalization bound. Empirically, across computer vision and language tasks, PENEX improves neural network generalization in low-data regimes, matching and in some settings outperforming established regularizers at comparable computational cost. Our results highlight the potential of the exponential loss beyond its application in AdaBoost.

We introduce DeepFHT, a survival-analysis framework that couples deep neural networks with first hitting time (FHT) distributions from stochastic process theory. Time to event is represented as the first passage of a latent diffusion process to an absorbing boundary. A neural network maps input variables to physically meaningful parameters including initial condition, drift, and diffusion, within a chosen FHT process such as Brownian motion, both with drift and driftless. This yields closed- form survival and hazard functions and captures time-varying risk without assuming proportional- hazards. We compare DeepFHT with Cox regression using synthetic and real-world datasets. The method achieves predictive accuracy on par with the state-of-the-art approach, while maintaining a physics- based interpretable parameterization that elucidates the relation between input features and risk. This combination of stochastic process theory and deep learning provides a principled avenue for modeling survival phenomena in complex systems

Chain of thought (CoT) elicits reasoning in large language models by explicitly generating intermediate tokens. In contrast, latent thought reasoning operates directly in the continuous latent space, enabling computation beyond discrete linguistic representations. While both approaches exploit iterative computation, their comparative capabilities remain underexplored. In this work, we present a formal analysis showing that latent thought admits more efficient parallel computation than inherently sequential CoT. In contrast, CoT enables approximate counting and sampling through stochastic decoding. These separations suggest the tasks for which depth-driven recursion is more suitable, thereby offering practical guidance for choosing between reasoning paradigms.

Image restoration (IR) aims to recover images degraded by unknown mixtures while preserving semanticsconditions under which discriminative restorers and UNet-based diffusion priors often oversmooth, hallucinate, or drift. We present LucidFlux, a caption-free IR framework that adapts a large diffusion transformer (Flux.1) without image captions. Our LucidFlux introduces a lightweight dual-branch conditioner that injects signals from the degraded input and a lightly restored proxy to respectively anchor geometry and suppress artifacts. Then, a timestep- and layer-adaptive modulation schedule is designed to route these cues across the backbones hierarchy, in order to yield coarse-to-fine and context-aware updates that protect the global structure while recovering texture. After that, to avoid the latency and instability of text prompts or Vision-Language Model (VLM) captions, we enforce caption-free semantic alignment via SigLIP features extracted from the proxy. A scalable curation pipeline further filters large-scale data for structure-rich supervision. Across synthetic and in-the-wild benchmarks, our LucidFlux consistently outperforms strong open-source and commercial baselines, and ablation studies verify the necessity of each component. LucidFlux shows that, for large DiTs, when, where, and what to condition onrather than adding parameters or relying on text promptsis the governing lever for robust and caption-free image restoration in the wild.

Concept Bottleneck Models (CBMs) enable interpretable image classification by structuring predictions around human-understandable concepts, but extending this paradigm to video remains challenging due to the difficulty of extracting concepts and modeling them over time. In this paper, we introduce MoTIF (Moving Temporal Interpretable Framework), a transformer-based concept architecture that operates on sequences of temporally grounded concept activations, by employing per-concept temporal self-attention to model when individual concepts recur and how their temporal patterns contribute to predictions. Central to the framework is a class-conditioned VLM-based concept discovery module that extracts object- and action-centric textual concepts from training videos, yielding temporally expressive concept sets without manual concept annotation. Across multiple video benchmarks, this combination improves over global concept bottlenecks and remains competitive within the interpretable concept-bottleneck setting, while narrowing the gap to strong black-box video baselines that we report as contextual references. Code available at github.com/patrick-knab/MoTIF.

We propose a novel mixture of experts framework for field-of-view enhancement in binaural signal matching. Our approach enables dynamic spatial audio rendering that adapts to continuous talker motion, allowing users to emphasize or suppress sounds from selected directions while preserving natural binaural cues. Unlike traditional methods that rely on explicit direction-of-arrival estimation or operate in the Ambisonics domain, our signal-dependent framework combines multiple binaural filters in an online manner using implicit localization. This allows for real-time tracking and enhancement of moving sound sources, supporting applications such as speech focus, noise reduction, and world-locked audio in augmented and virtual reality. The method is agnostic to array geometry offering a flexible solution for spatial audio capture and personalized playback in next-generation consumer audio devices.

The quality of a transtibial prosthetic socket depends on the prosthetist's skills and expertise, as the fitting is performed manually. This study investigates multiple artificial intelligence (AI) approaches to help standardize transtibial prosthetic socket design. Data from 118 patients were collected by prosthetists working in the Dutch healthcare system. This data consists of a three-dimensional (3D) scan of the residual limb and a corresponding 3D model of the prosthetist-designed socket. Multiple data pre-processing steps are performed for alignment, standardization and optionally compression using Morphable Models and Principal Component Analysis. Afterward, three different algorithms - a 3D neural network, Feedforward neural network, and random forest - are developed to either predict 1) the final socket shape or 2) the adaptations performed by a prosthetist to predict the socket shape based on the 3D scan of the residual limb. Each algorithm's performance was evaluated by comparing the prosthetist-designed socket with the AI-generated socket, using two metrics in combination with the error location. First, we measure the surface-to-surface distance to assess the overall surface error between the AI-generated socket and the prosthetist-designed socket. Second, distance maps between the AI-generated and prosthetist sockets are utilized to analyze the error's location. For all algorithms, estimating the required adaptations outperformed direct prediction of the final socket shape. The random forest model applied to adaptation prediction yields the lowest error with a median surface-to-surface distance of 1.24 millimeters, a first quartile of 1.03 millimeters, and a third quartile of 1.54 millimeters.

Good storytelling involves surprise -- unpredictability in how the story unfolds -- and sense-making, the requirement that the story forms a coherent sequence. However, to date, these two qualities have largely been addressed in isolation. We formalize these qualities and their relationship in an information-theoretic framework, using detective fiction as a paradigm case of narratives in which a hidden truth is discovered through reasoning. Our central theoretical result shows that surprise and coherence must trade off for any *single* reader model, but can coexist when two reader modes are distinguished: a pre-revelation mode that forms expectations while the ending is unknown, and a post-resolution hindsight mode that re-evaluates the story after the culprit is revealed. The balance of these two dimensions is realized in the common requirement of *fair play*, giving the reader a chance to solve the mystery while maintaining a challenge. We operationalize the framework using large language models as simulated readers, and define reference-less evaluation metrics for surprise, coherence, and fair play. Experiments on LLM-generated stories validate our theoretical predictions: while models generally succeed in creating surprise or coherence, achieving fair play poses a challenge even for strong models. Moreover, surprise and coherence do not positively correlate across stories, resisting reduction to a single latent quality. A human study validates the metrics, confirming they capture aspects of narrative quality that matter to readers. Our metrics also reproduce established literary intuitions, finding Christie's stories more surprising and more fair-playing than Conan Doyle's.

We study the Logistic Contextual Slate Bandit problem, where, at each round, an agent selects a slate of $N$ items from an exponentially large set (of size $2^{Ω(N)}$) of candidate slates provided by the environment. A single binary reward, determined by a logistic model, is observed for the chosen slate. Our objective is to develop algorithms that maximize cumulative reward over $T$ rounds while maintaining low per-round computational costs. We propose two algorithms, Slate-GLM-OFU and Slate-GLM-TS, that accomplish this goal. These algorithms achieve $N^{O(1)}$ per-round time complexity via local planning (independent slot selections), and low regret through global learning (joint parameter estimation). We provide theoretical and empirical evidence supporting these claims. Under a well-studied diversity assumption, we prove that Slate-GLM-OFU incurs only $\tilde{O}(\sqrt{T})$ regret. Extensive experiments across a wide range of synthetic settings demonstrate that our algorithms consistently outperform state-of-the-art baselines, achieving both the lowest regret and the fastest runtime. Furthermore, we apply our algorithm to select in-context examples in prompts of Language Models for solving binary classification tasks such as sentiment analysis. Our approach achieves competitive test accuracy, making it a viable alternative in practical scenarios.

Machine Learning systems are increasingly deployed in decision-making settings that shape user behavior and, in turn, the data on which future decisions are based. Performative Prediction (PP) formalizes this feedback loop by modeling how deployed models induce distributional shifts. It studies how to learn robust and well-performing models under such dynamics. However, existing PP frameworks typically assume that the model governing these decisions is the same model observed by users (therefore, to which they respond). In practice, deployer institutions may instead disclose curated models, while internally relying on distinct opaque models. We introduce Decoupled Performative Prediction (DPP), a framework that explicitly models mismatches between the model governing institutional decisions and the model that shapes user behavior. By analyzing the resulting optimization landscape, we show that DPP admits new different solutions that provably achieve lower risk for the institution than those under classical PP. We further propose an algorithm with provable convergence guarantees under standard assumptions, demonstrating how easy institutions can benefit from strategically deceptive deployment when they control model disclosure and users lack countervailing power. To capture the implications of such behavior, we introduce the deception cost, a quantitative measure of the degree of deception experienced by users. We study settings in which institutions incorporate this cost into the optimization process, motivated by reputational concerns or potential user abandonment, and show that such self-imposed constraints are insufficient to protect users. Overall, our results demonstrate that model disclosure is not merely an ethical consideration but a core technical design decision, underscoring the need for regulations that hold institutions accountable for deceptive deployment practices.

We address the problem of generating simulated, yet realistic, time-series data from a causal model with the same observational and interventional distributions as a given real dataset (probabilistic causal digital twin). While non-causal models (e.g., GANs) also strive to simulate realistic data, causal models are fundamentally more powerful, able to simulate the effect of interventions (what-if scenarios), optimize decisions, perform root-cause analysis, and counterfactual causal reasoning. We introduce the Adversarial Causal Tuning (ACT) methodology, which outputs the optimal causal model that fits the data, along with a quantification of the goodness-of-fit. The returned causal model can then be employed to simulate new data or to perform other causal reasoning tasks. ACT adopts ideas from Generative Adversarial Network training and AutoML to search for optimal causal pipelines and discriminators that detect deviations between the distributions of real and simulated data. It also adapts a permutation testing procedure from established causal tuning methods to penalize models for complexity. Through extensive experiments on real, semi-synthetic, and synthetic datasets, we show that (a) employing multiple optimized discriminators is paramount for selecting the optimal causal models and quantifying goodness-of-fit, (b) ACT selects the optimal causal model in synthetic datasets while avoiding overfitting, generating data indistinguishable from the true data distribution (c) all state-of-the-art generative and causal simulation methods, exhibit room for improvement in reproducing real data distributions; generating realistic temporal data is still an open research challenge.

Being able to solve a task in diverse ways makes agents more robust to task variations and less prone to local optima. In this context, constrained diversity optimization has become a useful reinforcement learning (RL) framework for training a set of diverse agents in parallel. However, existing constrained-diversity RL methods often under-explore in complex tasks such as robot manipulation, resulting in limited behavioral diversity. We address this with a two-stage curriculum that introduces a spline-based trajectory prior as an inductive bias to produce diverse, high-reward behaviors in an initial stage, and then distills these behaviors into reactive, step-wise policies in a second stage. In our empirical evaluation, we provide novel insights into challenges of diversity-targeted training and show that our curriculum increases the diversity of learned skills while maintaining high task performance.

Applying Transformers to irregular time-series typically requires specializations to their baseline architecture, which can result in additional computational overhead and increased method complexity. We present the Rotary Masked Autoencoder (RoMAE), which utilizes the popular Rotary Positional Embedding (RoPE) method for continuous positions. RoMAE is an extension to the Masked Autoencoder (MAE) that enables interpolation and representation learning with multidimensional continuous positional information while avoiding any time-series-specific architectural specializations. We showcase RoMAE's performance on a variety of modalities including irregular and multivariate time-series, images, and audio, demonstrating that RoMAE surpasses specialized time-series architectures on difficult datasets such as the DESC ELAsTiCC Challenge while maintaining MAE's usual performance across other modalities. In addition, we investigate RoMAE's ability to reconstruct the embedded continuous positions, demonstrating that including learned embeddings in the input sequence breaks RoPE's relative position property.

Achieving versatile humanoid locomotion with a single policy presents a critical scalability challenge. Prevailing methods often rely on distilling multiple terrain-specific teacher policies into a unified student policy. However, while such distillation captures basic locomotion primitives, it struggles to organically compose these skills to adapt to complex environments, resulting in poor generalization to novel composite terrains unseen during training. To overcome this, we present DreamPolicy, a unified framework that integrates offline data with a diffusion-based world model, enabling a single policy to master both known and unseen terrains. Central to our approach is a terrain-aware world model, driven by an autoregressive diffusion world model trained on aggregated rollouts from specialized policies. This model synthesizes physically plausible future trajectories, which serve as dynamic objectives for a conditioned policy, thereby bypassing manual reward engineering. Unlike distillation, our world model captures generalizable locomotion skills, allowing for robust zero-shot transfer to unseen composite terrains. DreamPolicy naturally scales with data availability. As the offline dataset expands, the diffusion world model continuously acquires richer skills. Experiments demonstrate that DreamPolicy outperforms the strongest baseline by up to 27\% on unseen terrains and 38\% on combined terrains. By unifying world model-based planning and policy learning, DreamPolicy breaks the "one task, one policy" bottleneck and establishes a scalable, data-driven paradigm for generalist humanoid control.

Graph Contrastive Learning (GCL) relies on semantically consistent graph augmentations, but common local perturbations provide limited control over global structural consistency, motivating a more principled global augmentation strategy. We therefore propose Fractal Graph Contrastive Learning (FractalGCL), a theory-motivated framework that constructs a renormalisation-based augmented graph and introduces a fractal-dimension-aware contrastive loss that penalises unreliable positive views and reweights negative-pair repulsion by finite-scale box-counting discrepancies. However, computing these discrepancies introduces substantial overhead, so we derive and justify a Gaussian surrogate that avoids repeated box-counting on renormalised graphs, yielding about a $61\%$ runtime reduction. Experiments show that FractalGCL serves as an effective frozen-pretraining tool on MalNet-Tiny, achieves strong performance on the standard TUDataset benchmarks, and outperforms the next-best method on real-world urban traffic tasks by $4.51$ percentage points in average accuracy. Code is available at https://anonymous.4open.science/r/FractalGCL-0511/.

Machine Unlearning (MU) aims to remove the information of specific training data from a trained model, ensuring compliance with privacy regulations and user requests. While one line of existing MU methods relies on linear parameter updates via task arithmetic, they suffer from weight entanglement. In this work, we propose a novel MU framework called Mode Connectivity Unlearning (MCU) that leverages mode connectivity to find an unlearning pathway in a nonlinear manner. To further enhance performance and efficiency, we introduce a parameter mask strategy that not only improves unlearning effectiveness but also reduces computational overhead. Moreover, we propose an adaptive adjustment strategy for our unlearning penalty coefficient to adaptively balance forgetting quality and predictive performance during training, eliminating the need for empirical hyperparameter tuning. Unlike traditional MU methods that identify only a single unlearning model, MCU uncovers a spectrum of unlearning models along the pathway. Overall, MCU serves as a plug-and-play framework that seamlessly integrates with any existing MU methods, consistently improving unlearning efficacy. Extensive experiments on the image classification task demonstrate that MCU achieves superior performance. The codes are available at https://github.com/TIML-Group/Mode-Connectivity-Unlearning.

Language modeling has shifted in recent years from a distribution over strings to prediction models with textual inputs and outputs for general-purpose tasks. This position paper highlights the often overlooked implications of this shift for the use of large language models (LLMs) as probability estimators, especially for world probabilities. In light of the theoretical distinction between distribution estimation and response prediction, we examine LLM training phases and common use cases for LLM output probabilities. We show that the different settings lead to distinct, potentially conflicting, desired output distributions. This lack of clarity leads to pitfalls when using output probabilities as event probabilities. Our position advocates for second-order prediction -- incorporating probabilities explicitly as part of the output -- as a theoretically sound method, in contrast to using token logprobs. We conclude with suggestions for potential directions to improve the probabilistic soundness of this method.

Clinical case reports and discharge summaries may be the most complete and accurate summarization of patient encounters, yet they are finalized, i.e., timestamped after the encounter. Complementary structured data streams become available sooner but suffer from incompleteness. To train models and algorithms on more complete and temporally fine-grained data, we construct a pipeline to phenotype, extract, and annotate time-localized findings within case reports using large language models. We apply our pipeline to generate an open-access textual time series corpus for Sepsis-3 comprising 2,139 case reports from the PubMed-Open Access (PMOA) Subset. To validate our system, we apply it to PMOA and timeline annotations from i2b2/MIMIC-IV and compare the results to physician-expert annotations. We show high recovery rates of clinical findings (event match rates: GPT-5--0.93, Llama 3.3 70B Instruct--0.76) and strong temporal ordering (concordance: GPT-5--0.965, Llama 3.3 70B Instruct--0.908). Our work characterizes the ability of LLMs to time-localize clinical findings in text, illustrating the limitations of LLM use for temporal reconstruction and providing several potential avenues of improvement via multimodal integration.

Pairwise LLM-as-a-judge evaluation asks the judge to identify the \emph{better} of two candidate answers. We study a one-line modification that asks for the \emph{worse} answer instead and recovers the preference by elimination, a procedure we call Goal-Reversed Prompting (GRP). GRP introduces no extra inference rounds, composes with any prompt template (direct, chain-of-thought, or Arena-Hard SOP), and leaves the rest of the evaluation pipeline untouched. Two observations motivate the reversal. Reverse reasoning is a recurring strategy in human problem solving, and modern instruction-tuned judges exhibit a positive-leaning bias that asking for the worse answer can counteract. On JudgeBench under a strict consistency protocol that counts a judgment as correct only when both response orderings agree with the gold preference, GRP improves all three closed-source judges we test across both response-pair sources. With GPT-4o-generated pairs, the Arena-Hard SOP baseline improves from 61.71\% to 66.23\% for GPT-4o (+4.52) and from 60.00\% to 66.00\% for Claude-3.5-Sonnet (+6.00), with the largest absolute gains on Reasoning and Mathematics. The lift persists when response pairs come from Claude-3.5-Sonnet and when the SOP scaffolding is stripped to a minimal direct-prompting template, suggesting that goal reversal acts on the underlying judging behavior rather than on a particular rubric. Stronger judges benefit more than weaker ones, suggesting that goal reversal exposes additional reasoning capacity rather than compensating for its absence.

Decades of research in machine learning have given us powerful tools for making accurate predictions. But when used in social settings and on human inputs, better accuracy does not immediately translate to better social outcomes. To effectively promote social well-being through machine learning, this position article advocates for the wide adoption of \emph{social welfare} as a guiding principle. The field of welfare economics asks: how should we allocate limited resources to self-interested agents in a way that maximizes social benefit? We argue that this perspective applies to many modern applications of machine learning in social contexts. As such, we propose that welfare serves as an additional core criterion in the design, study, and use of learning algorithms, complementing the conventional pillars of optimization, generalization, and expressivity, and as a compass guiding both theory and practice.

We introduce the problem of best arm identification (BAI) with post-action context, a new BAI problem in a stochastic multi-armed bandit environment and the fixed-confidence setting. The problem addresses the scenarios in which the learner receives a post-action context in addition to the reward after playing each action. This post-action context provides additional information that can significantly facilitate the decision process. We analyze two different types of the post-action context: (i) separator, where the reward depends solely on the context, and (ii) non-separator, where the reward depends on both the action and the context. For both cases, we derive instance-dependent lower bounds on the sample complexity and propose algorithms that asymptotically achieve the optimal sample complexity. For the separator setting, we propose a novel sampling rule called G-tracking, which uses the geometry of the context space to directly track the contexts rather than the actions. For the non-separator setting, we do so by demonstrating that the Track-and-Stop algorithm can be extended to this setting. Moreover, in both settings, we theoretically and empirically show that algorithms that ignore the post-action context are sub-optimal. Finally, our empirical results showcase the advantage of our approaches compared to the state of the art.

Machine unlearning seeks to remove the influence of specified data from a trained model. While the unlearning accuracy provides a widely used metric for assessing unlearning performance, it falls short in assessing the reliability of forgetting. In this paper, we find that the forgetting data points misclassified by unlearning accuracy still have their ground truth labels included in the conformal prediction set from the uncertainty quantification perspective, leading to a phenomenon we term fake forgetting. To address this issue, we propose a novel metric CR, inspired by conformal prediction, that offers a more reliable assessment of forgetting quality. Building on these insights, we further propose an unlearning framework CPU that incorporates conformal prediction into the Carlini & Wagner adversarial attack loss, enabling the ground truth label to be effectively removed from the conformal prediction set. Through extensive experiments on image classification tasks, we demonstrate both the effectiveness of our proposed metric and the superior forgetting quality achieved by our framework. Code is available at https://github.com/TIML-Group/Conformal-Prediction-Unlearning.

Machine learning models are often evaluated using point estimates of performance metrics such as accuracy, F1 score, or mean squared error. Such summaries fail to capture the inherent variability induced by stochastic elements of the training process, including data splitting, initialization, and hyperparameter optimization. This work proposes a distributional perspective on model evaluation by treating performance metrics as random quantities rather than fixed values. Instead of focusing solely on aggregate measures, empirical distributions of performance metrics are analyzed using quantiles and corresponding confidence intervals. The study investigates point and interval estimation of quantiles based on real-data use cases for classification and regression tasks, complemented by simulation studies for validation. Special emphasis is placed on small sample sizes, reflecting practical constraints in machine learning, where repeated training is computationally expensive. The results show that meaningful statistical inference on the underlying performance distribution is feasible even with sample sizes in the range of 10-25, while standard nonparametric confidence interval remain applicable under these conditions. The proposed approach provides a more detailed characterization of variability and uncertainty compared to mean-based evaluation and enables a more differentiated comparison of models. In particular, it supports a risk-oriented interpretation of model performance, which is relevant in applications where reliability is critical. The presented methods are easy to implement and broadly applicable, making them a practical extension to standard performance evaluation procedures in machine learning.

Deep neural networks (DNNs) remain challenged by distribution shifts in complex open-world domains like automated driving (AD): Robustness against yet unknown novel objects (semantic shift) or styles like lighting conditions (covariate shift) cannot be guaranteed. Hence, reliable operation-time monitors for identification of out-of-training-data-distribution (OOD) scenarios are imperative. Current approaches for OOD classification are untested for complex domains like AD, are limited in the kinds of shifts they detect, or even require supervision with OOD samples. To prepare for unanticipated shifts, we instead establish a framework around a principled, unsupervised and model-agnostic method that unifies detection of semantic and covariate shifts: Find a full model of the training data's feature distribution, to then use its density at new points as in-distribution (ID) score. To implement this, we propose to combine Vision Foundation Models (VFMs) as feature extractors with density modeling techniques. Through a comprehensive benchmark of 4 VFMs with different backbone architectures and 5 density-modeling techniques against established baselines, we provide the first systematic evaluation of OOD classification capabilities of VFMs across diverse conditions. A comparison with state-of-the-art binary OOD classification methods reveals that VFM embeddings with density estimation outperform existing approaches in identifying OOD inputs. Additionally, we show that our method detects high-risk inputs likely to cause errors in downstream tasks, thereby improving overall performance. Overall, VFMs, when coupled with robust density modeling techniques, are promising to realize model-agnostic, unsupervised, reliable safety monitors in complex vision tasks

In recent years, vision language models (VLMs) have made significant advancements in video understanding. However, a crucial capability - fine-grained motion comprehension - remains under-explored in current benchmarks. To address this gap, we propose MotionBench, a comprehensive evaluation benchmark designed to assess the fine-grained motion comprehension of video understanding models. MotionBench evaluates models' motion-level perception through six primary categories of motion-oriented question types and includes data collected from diverse sources, ensuring a broad representation of real-world video content. Experimental results reveal that existing VLMs perform poorly in understanding fine-grained motions. To enhance VLM's ability to perceive fine-grained motion within a limited sequence length of LLM, we conduct extensive experiments reviewing VLM architectures optimized for video feature compression and propose a novel and efficient Through-Encoder (TE) Fusion method. Experiments show that higher frame rate inputs and TE Fusion yield improvements in motion understanding, yet there is still substantial room for enhancement. Our benchmark aims to guide and motivate the development of more capable video understanding models, emphasizing the importance of fine-grained motion comprehension. Project page: https://motion-bench.github.io .

Gaussian graphical model selection is usually studied under independent sampling, but in many applications observations arise from dependent dynamics. We study structure learning when the data consist of a single trajectory of Gaussian Glauber dynamics. We develop two complementary approaches. The first is a local edge-testing estimator based on an appropriately designed correlation test that reveals edges. This estimator does not require waiting for the chain to mix and admits an embarrassingly parallel edgewise implementation. The second is a burn-in/thinning reduction: under a Dobrushin contraction condition, we prove that a suitably subsampled Gaussian Gibbs trajectory is close in total variation to an i.i.d. product sample, allowing standard i.i.d. Gaussian graphical model learners to be used as black boxes. The key technical ingredient, which may be of independent interest, is a high-dimensional total-variation bound for random-scan Gaussian Gibbs samplers, obtained by combining Wasserstein contraction with an approximate Lipschitz smoothing argument. We prove finite-sample recovery guarantees for both approaches, establish information-theoretic lower bounds on the observation time, and empirically compare the resulting sample-computation tradeoffs.