Recent developments in large language models have shown advantages in reallocating a notable share of computational resource from training time to inference time. However, the principles behind inference time scaling are not well understood. In this paper, we introduce an analytically tractable model of inference-time scaling: Bayesian linear regression with a reward-weighted sampler, where the reward is determined from a linear model, modeling LLM-as-a-judge scenario. We study this problem in the high-dimensional regime, where the deterministic equivalents dictate a closed-form expression for the posterior predictive mean and variance. We analyze the generalization error when training data are sampled from a teacher model. We draw $k$ inference-time samples and select via softmax at a temperature applied to a quadratic reward. When the reward is not too different from the teacher, the generalization error decreases monotonically with increasing inference time samples $k$. However, the specific reward that optimizes inference-time selection generally differs from the teacher. In contrast, substantial reward misspecification induces a finite optimal $k$ beyond which more sampling can increase the generalization error. For fixed $k$, there exists an optimal sampling temperature. We experimentally verify these facts in large language model inference with an additional large language model as a judge. In the "best-of-$k$" limit with the teacher as reward, we theoretically show that the generalization error decays as $Θ(1/k^2)$ and determine the leading coefficient via extreme value theory. These formulas delineate domains where scaling inference-time computation is provably preferable to collecting more data. Finally, we demonstrate that when task difficulty increases, the previously mentioned advantage of inference-time compute degrades.

Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction during inference. Activation steering is a popular approach for inference-time controllability that has shown potential in inducing safer behavior in large language models (LLMs). However, we observe that current steering methods struggle to correct biases, where equiprobable outcomes across demographic groups are required. To address this, we propose Direct Steering Optimization (DSO) which uses reinforcement learning to find linear transformations for steering activations, tailored to mitigate bias while maintaining control over model performance. We demonstrate that DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs, while offering practitioners inference-time control over the trade-off. Overall, our work highlights the benefit of designing steering strategies that are directly optimized to control model behavior, providing more effective bias intervention than methods that rely on pre-defined heuristics for controllability.

Recent advances in 3D-aware generative models have enabled high-fidelity image synthesis of human identities. However, this progress raises urgent questions around user consent and the ability to remove specific individuals from a model's output space. We address this by introducing SUGAR, a framework for scalable generative unlearning that enables the removal of many identities (simultaneously or sequentially) without retraining the entire model. Rather than projecting unwanted identities to unrealistic outputs or relying on static template faces, SUGAR learns a personalized surrogate latent for each identity, diverting reconstructions to visually coherent alternatives while preserving the model's quality and diversity. We further introduce a continual utility preservation objective that guards against degradation as more identities are forgotten. SUGAR achieves state-of-the-art performance in removing up to 200 identities, while delivering up to a 700% improvement in retention utility compared to existing baselines. Our code is publicly available at https://github.com/judydnguyen/SUGAR-Generative-Unlearn.

We propose Modal Logical Neural Networks (MLNNs), a neurosymbolic framework that integrates deep learning with the formal semantics of modal logic, enabling reasoning about necessity and possibility. Drawing on Kripke semantics, we introduce specialized neurons for the modal operators $\Box$ and $\Diamond$ that operate over a set of possible worlds, enabling the framework to act as a differentiable ``logical guardrail.'' The architecture is highly flexible: the accessibility relation between worlds can either be fixed by the user to enforce known rules or, as an inductive feature, be parameterized by a neural network. This allows the model to optionally learn the relational structure of a logical system from data while simultaneously performing deductive reasoning within that structure. This versatile construction is designed for flexibility. The entire framework is differentiable from end to end, with learning driven by minimizing a logical contradiction loss. This not only makes the system resilient to inconsistent knowledge but also enables it to learn nonlinear relationships that can help define the logic of a problem space. We illustrate MLNNs on four case studies: grammatical guardrailing, multi-agent epistemic trust, detecting constructive deception in natural language negotiation, and combinatorial constraint satisfaction in Sudoku. These experiments demonstrate how enforcing or learning accessibility can increase logical consistency and interpretability without changing the underlying task architecture.

Ischemic stroke is a time-critical medical emergency where rapid diagnosis is essential for improving patient outcomes. Non-contrast computed tomography (NCCT) serves as the frontline imaging tool, yet it often fails to reveal the subtle ischemic changes present in the early, hyperacute phase. This limitation can delay crucial interventions. To address this diagnostic challenge, we introduce a semi-supervised segmentation method using generative adversarial networks (GANs) to accurately delineate early ischemic stroke regions. The proposed method employs an adversarial framework to effectively learn from a limited number of annotated NCCT scans, while simultaneously leveraging a larger pool of unlabeled scans. By employing Dice loss, cross-entropy loss, a feature matching loss and a self-training loss, the model learns to identify and delineate early infarcts, even when they are faint or their size is small. Experiments on the publicly available Acute Ischemic Stroke Dataset (AISD) demonstrate the potential of the proposed method to enhance diagnostic capabilities, reduce the burden of manual annotation, and support more efficient clinical decision-making in stroke care.

Phonocardiogram (PCG) analysis is vital for cardiovascular disease diagnosis, yet the scarcity of labeled pathological data hinders the capability of AI systems. To bridge this, we introduce H-LDM, a Hierarchical Latent Diffusion Model for generating clinically accurate and controllable PCG signals from structured metadata. Our approach features: (1) a multi-scale VAE that learns a physiologically-disentangled latent space, separating rhythm, heart sounds, and murmurs; (2) a hierarchical text-to-biosignal pipeline that leverages rich clinical metadata for fine-grained control over 17 distinct conditions; and (3) an interpretable diffusion process guided by a novel Medical Attention module. Experiments on the PhysioNet CirCor dataset demonstrate state-of-the-art performance, achieving a Fréchet Audio Distance of 9.7, a 92% attribute disentanglement score, and 87.1% clinical validity confirmed by cardiologists. Augmenting diagnostic models with our synthetic data improves the accuracy of rare disease classification by 11.3\%. H-LDM establishes a new direction for data augmentation in cardiac diagnostics, bridging data scarcity with interpretable clinical insights.

Graph Neural Networks (GNNs) have emerged as powerful tools for learning over graph-structured data, yet recent studies have shown that their performance gains are beginning to plateau. In many cases, well-established models such as GCN and GAT, when appropriately tuned, can match or even exceed the performance of more complex, state-of-the-art architectures. This trend highlights a key limitation in the current landscape: the difficulty of selecting the most suitable model for a given graph task or dataset. To address this, we propose Self-Adaptive Graph Mixture of Models (SAGMM), a modular and practical framework that learns to automatically select and combine the most appropriate GNN models from a diverse pool of architectures. Unlike prior mixture-of-experts approaches that rely on variations of a single base model, SAGMM leverages architectural diversity and a topology-aware attention gating mechanism to adaptively assign experts to each node based on the structure of the input graph. To improve efficiency, SAGMM includes a pruning mechanism that reduces the number of active experts during training and inference without compromising performance. We also explore a training-efficient variant in which expert models are pretrained and frozen, and only the gating and task-specific layers are trained. We evaluate SAGMM on 16 benchmark datasets covering node classification, graph classification, regression, and link prediction tasks, and demonstrate that it consistently outperforms or matches leading GNN baselines and prior mixture-based methods, offering a robust and adaptive solution for real-world graph learning.

The quadratic compute and memory costs of global self-attention severely limit its use in high-resolution images. Local attention reduces complexity by restricting attention to neighborhoods. Block-sparse kernels can further improve the efficiency of local attention, but conventional local attention patterns often fail to deliver significant speedups because tokens within a window are not contiguous in the 1D sequence. This work proposes a novel method for constructing windows and neighborhoods based on the Hilbert curve. Image tokens are first reordered along a Hilbert curve, and windows and neighborhoods are then formed on the reordered 1D sequence. From a block-sparse perspective, this strategy significantly increases block sparsity and can be combined with existing block-sparse kernels to improve the efficiency of 2D local attention. Experiments show that the proposed Hilbert Window Attention and Hilbert Slide Attention can accelerate window attention and slide attention by about $4\times$ and $18\times$, respectively. To assess practicality, the strategy is instantiated as the Hilbert Window Transformer and the Hilbert Neighborhood Transformer, both of which achieve end-to-end speedups with minimal accuracy loss. Overall, combining Hilbert-guided local attention with block-sparse kernels offers a general and practical approach to enhancing the efficiency of 2D local attention for images.

We present ETHOS (Encountered-Type Haptics for On-demand Social Interaction), a dynamic encountered-type haptic display (ETHD) that enables natural physical contact in virtual reality (VR) during social interactions such as handovers, fist bumps, and high-fives. The system integrates a torque-controlled robotic manipulator with interchangeable passive props (silicone hand replicas and a baton), marker-based physical-virtual registration via a ChArUco board, and a safety monitor that gates motion based on the user's head and hand pose. We introduce two control strategies: (i) a static mode that presents a stationary prop aligned with its virtual counterpart, consistent with prior ETHD baselines, and (ii) a dynamic mode that continuously updates prop position by exponentially blending an initial mid-point trajectory with real-time hand tracking, generating a unique contact point for each interaction. Bench tests show static colocation accuracy of 5.09 +/- 0.94 mm, while user interactions achieved temporal alignment with an average contact latency of 28.53 +/- 31.21 ms across all interaction and control conditions. These results demonstrate the feasibility of recreating socially meaningful haptics in VR. By incorporating essential safety and control mechanisms, ETHOS establishes a practical foundation for high-fidelity, dynamic interpersonal interactions in virtual environments.

We introduce the task of Multi-Modal Context-Aware Recognition (MCoRec) in the ninth CHiME Challenge, which addresses the cocktail-party problem of overlapping conversations in a single-room setting using audio, visual, and contextual cues. MCoRec captures natural multi-party conversations where the recordings focus on unscripted, casual group chats, leading to extreme speech overlap of up to 100% and highly fragmented conversational turns. The task requires systems to answer the question "Who speaks when, what, and with whom?" by jointly transcribing each speaker's speech and clustering them into their respective conversations from audio-visual recordings. Audio-only baselines exceed 100% word error rate, whereas incorporating visual cues yields substantial 50% improvements, highlighting the importance of multi-modality. In this manuscript, we present the motivation behind the task, outline the data collection process, and report the baseline systems developed for the MCoRec.

Large Language Models (LLMs) have redefined complex task automation with exceptional generalization capabilities. Despite these advancements, state-of-the-art methods rely on single-strategy prompting, missing the synergy of diverse reasoning approaches. No single strategy excels universally, highlighting the need for frameworks that fuse strategies to maximize performance and ensure robustness. We introduce the Select, Mix, and ReinvenT (SMaRT) framework, an innovative strategy fusion approach designed to overcome this constraint by creating balanced and efficient solutions through the seamless integration of diverse reasoning strategies. Unlike existing methods, which employ LLMs merely as evaluators, SMaRT uses them as intelligent integrators, unlocking the "best of all worlds" across tasks. Extensive empirical evaluations across benchmarks in reasoning, planning, and sequential decision-making highlight the robustness and adaptability of SMaRT. The framework consistently outperforms state-of-the-art baselines in solution quality, constraint adherence, and performance metrics. This work redefines LLM-driven decision-making by pioneering a new paradigm in cross-strategy calibration, unlocking superior outcomes for reasoning systems and advancing the boundaries of self-refining methodologies.

Generative single-image super-resolution (SISR) is advancing rapidly, yet even state-of-the-art models produce visual artifacts: unnatural patterns and texture distortions that degrade perceived quality. These defects vary widely in perceptual impact--some are barely noticeable, while others are highly disturbing--yet existing detection methods treat them equally. We propose characterizing artifacts by their prominence to human observers rather than as uniform binary defects. We present a novel dataset of 1302 artifact examples from 11 SISR methods annotated with crowdsourced prominence scores, and provide prominence annotations for 593 existing artifacts from the DeSRA dataset, revealing that 48% of them go unnoticed by most viewers. Building on this data, we train a lightweight regressor that produces spatial prominence heatmaps. We demonstrate that our method outperforms existing detectors and effectively guides SR model fine-tuning for artifact suppression. Our dataset and code are available at https://tinyurl.com/2u9zxtyh.

Precise robot manipulation is critical for fine-grained applications such as chemical and biological experiments, where even small errors (e.g., reagent spillage) can invalidate an entire task. Existing approaches often rely on pre-collected expert demonstrations and train policies via imitation learning (IL) or offline reinforcement learning (RL). However, obtaining high-quality demonstrations for precision tasks is difficult and time-consuming, while offline RL commonly suffers from distribution shifts and low data efficiency. We introduce a Role-Model Reinforcement Learning (RM-RL) framework that unifies online and offline training in real-world environments. The key idea is a role-model strategy that automatically generates labels for online training data using approximately optimal actions, eliminating the need for human demonstrations. RM-RL reformulates policy learning as supervised training, reducing instability from distribution mismatch and improving efficiency. A hybrid training scheme further leverages online role-model data for offline reuse, enhancing data efficiency through repeated sampling. Extensive experiments show that RM-RL converges faster and more stably than existing RL methods, yielding significant gains in real-world manipulation: 53% improvement in translation accuracy and 20% in rotation accuracy. Finally, we demonstrate the successful execution of a challenging task, precisely placing a cell plate onto a shelf, highlighting the framework's effectiveness where prior methods fail.

Mamba, a recently proposed linear-time sequence model, has attracted significant attention for its computational efficiency and strong empirical performance. However, a rigorous theoretical understanding of its underlying mechanisms remains limited. In this work, we provide a theoretical analysis of Mamba's in-context learning (ICL) capability by focusing on tasks defined by low-dimensional nonlinear target functions. Specifically, we study in-context learning of a single-index model $y \approx g_*(\langle \boldsymbolβ, \boldsymbol{x} \rangle)$, which depends on only a single relevant direction $\boldsymbolβ$, referred to as feature. We prove that Mamba, pretrained by gradient-based methods, can achieve efficient ICL via test-time feature learning, extracting the relevant direction directly from context examples. Consequently, we establish a test-time sample complexity that improves upon linear Transformers -- analyzed to behave like kernel methods -- and is comparable to nonlinear Transformers, which have been shown to surpass the Correlational Statistical Query (CSQ) lower bound and achieve near information-theoretically optimal rate in previous works. Our analysis reveals the crucial role of the nonlinear gating mechanism in Mamba for feature extraction, highlighting it as the fundamental driver behind Mamba's ability to achieve both computational efficiency and high performance.

Quadrupedal robots exhibit a wide range of viable gaits, but generating specific footfall sequences often requires laborious expert tuning of numerous variables, such as touch-down and lift-off events and holonomic constraints for each leg. This paper presents a unified reinforcement learning framework for generating versatile quadrupedal gaits by leveraging the intrinsic symmetries and velocity-period relationship of dynamic legged systems. We propose a symmetry-guided reward function design that incorporates temporal, morphological, and time-reversal symmetries. By focusing on preserved symmetries and natural dynamics, our approach eliminates the need for predefined trajectories, enabling smooth transitions between diverse locomotion patterns such as trotting, bounding, half-bounding, and galloping. Implemented on the Unitree Go2 robot, our method demonstrates robust performance across a range of speeds in both simulations and hardware tests, significantly improving gait adaptability without extensive reward tuning or explicit foot placement control. This work provides insights into dynamic locomotion strategies and underscores the crucial role of symmetries in robotic gait design.

Large language models (LLMs) exhibit pronounced position bias in long-context needle-in-haystack problems, systematically prioritizing the location of information over its relevance. While current mitigations rely on white-box access, this is effectively impossible for many state-of-the-art models. We introduce GOLD PANNING, a black-box Bayesian framework that performs inference-time active search over long contexts by (i) reordering documents to concentrate high-belief items in highly diagnostic positions (signal anchoring) and (ii) updating beliefs over document relevance from model outputs. Unlike conventional active learning, which prioritizes uncertainty reduction, GOLD PANNING leverages anchoring -- once flagged, keep it in sight -- to preserve weak cues. We implement this using iterative assignment derived from the model's diagnosticity profile, which provably identifies a target among $N$ documents in $O(\log N)$ rounds, ensuring scalability to many-document settings.On needle-in-a-haystack retrieval and long-context QA, GOLD PANNING matches Permutation Self-Consistency's target identification with $30--65%$ fewer queries and remains effective under calibration mismatch, suggesting coarse positional ordering drives performance gains. These results demonstrate that inherent model biases need not be failures, but can be used as tools for control.

Is bigger always better for time series foundation models? With the question in mind, we explore an alternative to training a single, large monolithic model: building a portfolio of smaller, pretrained forecasting models. By applying ensembling or model selection over these portfolios, we achieve competitive performance on large-scale benchmarks using much fewer parameters. We explore strategies for designing such portfolios and find that collections of specialist models consistently outperform portfolios of independently trained generalists. Remarkably, we demonstrate that post-training a base model is a compute-effective approach for creating sufficiently diverse specialists, and provide evidences that ensembling and model selection are more compute-efficient than test-time fine-tuning.

As Large Language Model (LLM) agents increasingly gain self-evolutionary capabilities to adapt and refine their strategies through real-world interaction, their long-term reliability becomes a critical concern. We identify the Alignment Tipping Process (ATP), a critical post-deployment risk unique to self-evolving LLM agents. Unlike training-time failures, ATP arises when continual interaction drives agents to abandon alignment constraints established during training in favor of reinforced, self-interested strategies. We formalize and analyze ATP through two complementary paradigms: Self-Interested Exploration, where repeated high-reward deviations induce individual behavioral drift, and Imitative Strategy Diffusion, where deviant behaviors spread across multi-agent systems. Building on these paradigms, we construct controllable testbeds and benchmark both open and closed-source LLMs. Our experiments show that alignment benefits erode rapidly under self-evolution, with initially aligned models converging toward unaligned states. In multi-agent settings, successful violations diffuse quickly, leading to collective misalignment. Moreover, current reinforcement learning-based alignment methods provide limited defenses against alignment tipping. These findings demonstrate that alignment of LLM agents is not a static property but a fragile and dynamic one, vulnerable to feedback-driven decay during deployment. Our data and code are available at https://github.com/aiming-lab/ATP.

Acoustic scene classification (ASC) models on edge devices typically operate under fixed class assumptions, lacking the transferability needed for real-world applications that require adaptation to new or refined acoustic categories. We propose ContrastASC, which learns generalizable acoustic scene representations by structuring the embedding space to preserve semantic relationships between scenes, enabling adaptation to unseen categories without retraining. Our approach combines supervised contrastive fine-tuning of pre-trained models with contrastive representation distillation to transfer this structured knowledge to compact student models. Our evaluation shows that ContrastASC demonstrates improved few-shot adaptation to unseen categories while maintaining strong closed-set performance.

Conventional supervised learning methods are often vulnerable to spurious correlations, particularly under distribution shifts in test data. To address this issue, several approaches, most notably Group DRO, have been developed. While these methods are highly robust to subpopulation or group shifts, they remain vulnerable to intra-group distributional shifts, which frequently occur in minority groups with limited samples. We propose a hierarchical extension of Group DRO that addresses both inter-group and intra-group uncertainties, providing robustness to distribution shifts at multiple levels. We also introduce new benchmark settings that simulate realistic minority group distribution shifts-an important yet previously underexplored challenge in spurious correlation research. Our method demonstrates strong robustness under these conditions-where existing robust learning methods consistently fail-while also achieving superior performance on standard benchmarks. These results highlight the importance of broadening the ambiguity set to better capture both inter-group and intra-group distributional uncertainties.

Large Language Models (LLMs) have shown remarkable performance on general Question Answering (QA), yet they often struggle in domain-specific scenarios where accurate and up-to-date information is required. Retrieval-Augmented Generation (RAG) addresses this limitation by enriching LLMs with external knowledge, but existing systems primarily rely on unstructured documents, while largely overlooking relational databases, which provide precise, timely, and efficiently queryable factual information, serving as indispensable infrastructure in domains such as finance, healthcare, and scientific research. Motivated by this gap, we conduct a systematic analysis that reveals three central observations: (i) databases and documents offer complementary strengths across queries, (ii) naively combining both sources introduces noise and cost without consistent accuracy gains, and (iii) selecting the most suitable source for each query is crucial to balance effectiveness and efficiency. We further observe that query types show consistent regularities in their alignment with retrieval paths, suggesting that routing decisions can be effectively guided by systematic rules that capture these patterns. Building on these insights, we propose a rule-driven routing framework. A routing agent scores candidate augmentation paths based on explicit rules and selects the most suitable one; a rule-making expert agent refines the rules over time using QA feedback to maintain adaptability; and a path-level meta-cache reuses past routing decisions for semantically similar queries to reduce latency and cost. Experiments on three QA benchmarks demonstrate that our framework consistently outperforms static strategies and learned routing baselines, achieving higher accuracy while maintaining moderate computational cost.

The reasoning capabilities of Large Language Models (LLMs) are increasingly attributed to training data quality rather than mere parameter scaling. However, existing data-centric paradigms often equate quality with factuality or diversity and ignore the internal logical complexity of training samples. In this work, we propose that natural language harbors Structured Logical Knowledge manifested through entailment relationships and logical topologies. To quantify this, we introduce Structured Logical Knowledge Density (SLKD), a novel metric that measures logical information content by decomposing natural language into executable predicates and logical primitives. Our analysis reveals a significant logical disparity in current datasets where sparse logical signals predominate. Consequently, we propose a density aware re-cognizing optimization strategy that prioritizes high-density logical samples to enhance with the LLM's reasoning ability. Extensive experiments demonstrate that our approach enhances reasoning performance and generalization without increasing total data volume. These results, further validated within a reinforcement learning framework, suggest that elevating logical density is more critical than expanding data scale for realizing the full cognitive potential of LLMs. The released code is available in the Appendix C.

Existing works are dedicated to untangling atomized numerical components (features) from the hidden states of Large Language Models (LLMs). However, they typically rely on autoencoders constrained by some training-time regularization on single training instances, without an explicit guarantee of global sparsity among instances, causing a large amount of dense (simultaneously inactive) features, harming the feature sparsity and atomization. In this paper, we propose a novel autoencoder variant that enforces minimal entropy on minibatches of hidden activations, thereby promoting feature independence and sparsity across instances. For efficient entropy calculation, we discretize the hidden activations to 1-bit via a step function and apply gradient estimation to enable backpropagation, so that we term it as Binary Autoencoder (BAE) and empirically demonstrate two major applications: (1) Feature set entropy calculation. Entropy can be reliably estimated on binary hidden activations, which can be leveraged to characterize the inference dynamics of LLMs. (2) Feature untangling. Compared to typical methods, due to improved training strategy, BAE avoids dense features while producing the largest number of interpretable ones among baselines.

Multi-agent systems have demonstrated exceptional performance in downstream tasks beyond diverse single agent baselines. A growing body of work has explored ways to improve their reasoning and collaboration, from vote, debate, to complex interaction protocols. However, it still remains opaque why specific choice would be preferred in multi-agent systems. Inspired by the decision-making mechanism of democratic committees and The Society of Mind, we introduce Roundtable Policy, an inference-time reasoning framework for multi-agent systems that performs inference through the weighted consensus of multiple LLMs. Through extensive experiments, we demonstrate its that this approach significantly enhances reasoning in complex heterogeneous scientific tasks. Roundtable Policy emphasizes structured and interpretable inference rather than opaque convergence, while requires only black-box access and uniform procedures, making it broadly applicable to diverse multi-agent systems.

Mixture-of-Experts (MoE) in Large Language Models (LLMs) routes each token through a subset of specialized Feed-Forward Networks (FFN), known as experts. We present SteerMoE, a framework to steer MoE models by detecting and controlling behavior-associated experts. We detect key experts by comparing how often they activate between paired inputs that demonstrate opposite behaviors (e.g., safe vs. unsafe). By selectively activating or deactivating such experts during inference, we control behaviors like faithfulness and safety without fine-tuning. Across 11 benchmarks and 6 LLMs, our steering raises safety by up to +20% and faithfulness by +27%. Alternatively, unsafe steering drops safety by -41% alone, and -100% when combined with existing jailbreak methods, bypassing all safety guardrails. Overall, SteerMoE offers a lightweight, effective, and widely applicable test-time control, while revealing unique vulnerabilities in MoE LLMs. https://github.com/adobe-research/SteerMoE

We introduce MedQARo, the first large-scale medical QA benchmark in Romanian, alongside a comprehensive evaluation of state-of-the-art large language models (LLMs). We construct a high-quality and large-scale dataset comprising 105,880 QA pairs about cancer patients from two medical centers. The questions regard medical case summaries of 1,242 patients, requiring both keyword extraction and reasoning. Our benchmark contains both in-domain and cross-domain (cross-center and cross-cancer) test collections, enabling a precise assessment of generalization capabilities. We experiment with four open-source LLMs from distinct families of models on MedQARo. Each model is employed in two scenarios: zero-shot prompting and supervised fine-tuning. We also evaluate two state-of-the-art LLMs exposed only through APIs, namely GPT-5.2 and Gemini 3 Flash. Our results show that fine-tuned models significantly outperform zero-shot models, indicating that pretrained models fail to generalize on MedQARo. Our findings demonstrate the importance of both domain-specific and language-specific fine-tuning for reliable clinical QA in Romanian.

Despite their success in speech processing, neural networks often operate as black boxes, prompting the question: what informs their decisions, and how can we interpret them? This work examines this issue in the context of lexical stress. A dataset of English disyllabic words was automatically constructed from read and spontaneous speech. Several Convolutional Neural Network (CNN) architectures were trained to predict stress position from a spectrographic representation of disyllabic words lacking minimal stress pairs (e.g., initial stress WAllet, final stress exTEND), achieving up to 92% accuracy on held-out test data. Layerwise Relevance Propagation (LRP), a technique for neural network interpretability analysis, revealed that predictions for held-out minimal pairs (PROtest vs. proTEST ) were most strongly influenced by information in stressed versus unstressed syllables, particularly the spectral properties of stressed vowels. However, the classifiers also attended to information throughout the word. A feature-specific relevance analysis is proposed, and its results suggest that our best-performing classifier is strongly influenced by the stressed vowel's first and second formants, with some evidence that its pitch and third formant also contribute. These results reveal deep learning's ability to acquire distributed cues to stress from naturally occurring data, extending traditional phonetic work based around highly controlled stimuli.

Artificial intelligence (AI) systems increasingly support decision-making across critical domains, yet current explainable AI (XAI) approaches prioritize algorithmic transparency over human comprehension. While XAI methods reveal computational processes for model validation and audit, end users require explanations integrating domain knowledge, contextual reasoning, and professional frameworks. This disconnect reveals a fundamental design challenge: existing AI explanation approaches fail to address how practitioners actually need to understand and act upon recommendations. This paper introduces Explanatory AI as a complementary paradigm where AI systems leverage generative and multimodal capabilities to serve as explanatory partners for human understanding. Unlike traditional XAI that answers "How did the algorithm decide?" for validation purposes, Explanatory AI addresses "Why does this make sense?" for practitioners making informed decisions. Through theory-informed design, we synthesize multidisciplinary perspectives on explanation from cognitive science, communication research, and education with empirical evidence from healthcare contexts and AI expert interviews. Our analysis identifies five dimensions distinguishing Explanatory AI from traditional XAI: explanatory purpose (from diagnostic to interpretive sense-making), communication mode (from static technical to dynamic narrative interaction), epistemic stance (from algorithmic correspondence to contextual plausibility), adaptivity (from uniform design to personalized accessibility), and cognitive design (from information overload to cognitively aligned delivery). We derive five meta-requirements specifying what systems must achieve and formulate ten design principles prescribing how to build them.

Evaluating the capabilities and risks of foundation models is paramount, yet current methods demand extensive domain expertise, hindering their scalability as these models rapidly evolve. We introduce SKATE: a novel evaluation framework in which large language models (LLMs) compete by generating and solving verifiable tasks for one another. Our core insight is to treat evaluation as a game: models act as both task-setters and solvers, incentivized to create questions which highlight their own strengths while exposing others' weaknesses. SKATE offers several key advantages, balancing scalability, open-endedness, and objectivity. It is fully automated, data-free, and scalable, requiring no human input or domain expertise. By using verifiable tasks rather than LLM judges, scoring is objective. Unlike domain-limited programmatically-generated benchmarks (e.g. chess-playing or spatial reasoning), having LLMs creatively pose challenges enables open-ended and scalable evaluation. As a proof of concept, we introduce LLM-set code-output-prediction (COP) challenges as a verifiable and extensible framework in which to test our approach. Using a TrueSkill-based ranking system, we evaluate six frontier LLMs and find that: (1) weaker models can reliably differentiate and score stronger ones, (2) LLM-based systems are capable of self-preferencing behavior, generating questions that align with their own capabilities, and (3) SKATE automatically surfaces fine-grained capability differences between models. Our findings are an important step towards general, scalable evaluation frameworks which can keep pace with LLM progress.

Conformal unlearning aims to ensure that a trained conformal predictor miscovers data points with specific shared characteristics, such as those from a particular label class, associated with a specific user, or belonging to a defined cluster, while maintaining valid coverage on the remaining data. Existing machine unlearning methods, which typically approximate a model retrained from scratch after removing the data to be forgotten, face significant challenges when applied to conformal unlearning. These methods often lack rigorous, uncertainty-aware statistical measures to evaluate unlearning effectiveness and exhibit a mismatch between their degraded performance on forgotten data and the frequency with which that data are still correctly covered by conformal predictors-a phenomenon we term ''fake conformal unlearning''. To address these limitations, we propose a new paradigm for conformal machine unlearning that provides finite-sample, uncertainty-aware guarantees on unlearning performance without relying on a retrained model as a reference. We formalize conformal unlearning to require high coverage on retained data and high miscoverage on forgotten data, introduce practical empirical metrics for evaluation, and present an algorithm that optimizes these conformal objectives. Extensive experiments on vision and text benchmarks demonstrate that the proposed approach effectively removes targeted information while preserving utility.