Automatic Speech Recognition (ASR) systems, despite large multilingual training, struggle in low-resource scenarios where labeled data is scarce. We propose BEARD (BEST-RQ Encoder Adaptation with Re-training and Distillation), a novel framework designed to adapt Whisper's encoder with unlabeled data. Unlike traditional self-supervised learning methods, BEARD uniquely combines a BEST-RQ objective with knowledge distillation from a frozen teacher encoder, ensuring the encoder's complementarity with the pre-trained decoder. Our experiments focus on the ATCO2 corpus from the challenging Air Traffic Control (ATC) communications domain, characterized by non-native speech, noise, and specialized phraseology. Using about 5,000 hours of untranscribed speech for BEARD and 2 hours of transcribed speech for fine-tuning, the proposed approach significantly outperforms previous baseline and fine-tuned model, achieving a relative improvement of 12% compared to the fine-tuned model. To the best of our knowledge, this is the first work to use a self-supervised learning objective for domain adaptation of Whisper.

Large-scale multimodal models achieve strong results on tasks like Visual Question Answering (VQA), but they are often limited when queries require cultural and visual information, everyday knowledge, particularly in low-resource and underrepresented languages. We introduce OASIS, a large-scale culturally grounded multimodal QA dataset covering images, text, and speech. OASIS is built with EverydayMMQA, a scalable semi-automatic framework for creating localized spoken and visual QA resources, supported by multi-stage human-in-the-loop validation. OASIS contains approximately 0.92M real images and 14.8M QA pairs, including 3.7M spoken questions, with 383 hours of human-recorded speech, and 20K hours of voice-cloned speech, from 42 speakers. It supports four input settings: text-only, speech-only, text+image, and speech+image. The dataset focuses on English and Arabic varieties across 18 countries, covering Modern Standard Arabic (MSA) as well as dialectal Arabic. It is designed to evaluate models beyond object recognition, targeting pragmatic, commonsense, and culturally grounded reasoning in real-world scenarios. We benchmark four closed-source models, three open-source models, and one fine-tuned model on OASIS. The framework and dataset will be made publicly available to the community. https://huggingface.co/datasets/QCRI/OASIS

Recovering the past from present observations is an intriguing challenge with potential applications in forensics and scene analysis. Thermal imaging, operating in the infrared range, provides access to otherwise invisible information. Since humans are typically warmer (37 C -98.6 F) than their surroundings, interactions such as sitting, touching, or leaning leave residual heat traces. These fading imprints serve as passive temporal codes, allowing for the inference of recent events that exceed the capabilities of RGB cameras. This work proposes a time-reversed reconstruction framework that uses paired RGB and thermal images to recover scene states from a few seconds earlier. The proposed approach couples Visual-Language Models (VLMs) with a constrained diffusion process, where one VLM generates scene descriptions and another guides image reconstruction, ensuring semantic and structural consistency. The method is evaluated in three controlled scenarios, demonstrating the feasibility of reconstructing plausible past frames up to 120 seconds earlier, providing a first step toward time-reversed imaging from thermal traces.

Pass$@k$ is widely used to report the reasoning performance of LLMs, but it often produces unstable and potentially misleading rankings, especially when the number of trials (samples) is limited and computational resources are constrained. We present a principled Bayesian evaluation framework that replaces Pass$@k$ and average accuracy over $N$ trials (avg$@N$) with posterior estimates of a model's underlying success probability and credible intervals, yielding stable rankings and a transparent decision rule for differences. Evaluation outcomes are modeled as categorical (not just 0/1) with a Dirichlet prior, giving closed-form expressions for the posterior mean and uncertainty of any weighted rubric and enabling the use of prior evidence when appropriate. Theoretically, under a uniform prior, the Bayesian posterior mean is order-equivalent to average accuracy (Pass$@1$), explaining its empirical robustness while adding principled uncertainty. Empirically, in simulations with known ground-truth success rates and on AIME'24/'25, HMMT'25, and BrUMO'25, the posterior-based procedure achieves faster convergence and greater rank stability than Pass$@k$ and recent variants, enabling reliable comparisons at far smaller sample counts. The framework clarifies when observed gaps are statistically meaningful (non-overlapping credible intervals) versus noise, and it naturally extends to graded, rubric-based evaluations. Together, these results recommend replacing Pass$@k$ for LLM evaluation and ranking with a posterior-based, compute-efficient protocol that unifies binary and non-binary evaluation while making uncertainty explicit. Source code is available at https://github.com/mohsenhariri/scorio

Pose estimation refers to tracking a human's full body posture, including their head, torso, arms, and legs. The problem is challenging in practical settings where the number of body sensors are limited. Past work has shown promising results using conditional diffusion models, where the pose prediction is conditioned on both <location, rotation> measurements from the sensors. Unfortunately, nearly all these approaches generalize poorly across users, primarly because location measurements are highly influenced by the body size of the user. In this paper, we formulate pose estimation as an inverse problem and design an algorithm capable of zero-shot generalization. Our idea utilizes a pre-trained diffusion model and conditions it on rotational measurements alone; the priors from this model are then guided by a likelihood term, derived from the measured locations. Thus, given any user, our proposed InPose method generatively estimates the highly likely sequence of poses that best explains the sparse on-body measurements.

Contrastive vision-language models (VLMs) have made significant progress in binding visual and textual information, yet understanding long, compositional captions remains an open challenge. While these capabilities are often assumed to be closely related, the conditions under which they reinforce each other remain unclear. In this paper, we empirically analyze when compositional reasoning and long-caption understanding transfer across tasks, and when this relationship fails. Through controlled experiments across diverse training objectives, datasets, and architectural designs, we find a bidirectional but sensitive relationship between the two capabilities. Models trained on poorly grounded captions or with limited parameter updates fail to generalize, while high-quality long-caption data with strong visual grounding promotes both capabilities simultaneously. We further show that architectural choices aimed at preserving general alignment, such as frozen positional embeddings, can inadvertently limit compositional learning. Our analysis provides actionable guidelines for data selection and model design to improve VLM generalization.

Time series forecasting is essential in various domains. Compared to relying solely on endogenous variables (i.e., target variables), considering exogenous variables (i.e., covariates) provides additional predictive information and often leads to more accurate predictions. However, existing methods for time series forecasting with exogenous variables (TSF-X) have the following shortcomings: 1) they do not leverage future exogenous variables, 2) they fail to fully account for the correlation between endogenous and exogenous variables. In this study, to better leverage exogenous variables, especially future exogenous variables, we propose DAG, which utilizes Dual correlAtion network along both the temporal and channel dimensions for time series forecasting with exoGenous variables. Specifically, we propose two core components: the Temporal Correlation Module and the Channel Correlation Module. Both modules consist of a correlation discovery submodule and a correlation injection submodule. The former is designed to capture the correlation effects of historical exogenous variables on future exogenous variables and on historical endogenous variables, respectively. The latter injects the discovered correlation relationships into the processes of forecasting future endogenous variables based on historical endogenous variables and future exogenous variables.

This paper presents a motion planning and risk analysis framework for enhancing human-robot collaboration with a Multi-Rotor Aerial Vehicle. The proposed method employs Signal Temporal Logic to encode key mission objectives, including safety, temporal requirements, and human preferences, with particular emphasis on ergonomics and comfort. An optimization-based planner generates dynamically feasible trajectories while explicitly accounting for the vehicle's nonlinear dynamics and actuation constraints. To address the resulting non-convex and non-smooth optimization problem, smooth robustness approximations and gradient-based techniques are adopted. In addition, an uncertainty-aware risk analysis is introduced to quantify the likelihood of specification violations under human-pose uncertainty. A robustness-aware event-triggered replanning strategy further enables online recovery from disturbances and unforeseen events by preserving safety margins during execution. The framework is validated through MATLAB and Gazebo simulations on an object handover task inspired by power line maintenance scenarios. Results demonstrate the ability of the proposed method to achieve safe, efficient, and resilient human-robot collaboration under realistic operating conditions.

While Soft Actor-Critic (SAC) is highly effective in continuous control, its discrete counterpart (DSAC) performs poorly on challenging discrete-action domains such as Atari. Consequently, starting from DSAC, we revisit the design of actor-critic methods in this setting. First, we determine that the coupling between the actor and critic entropy is the primary reason behind the poor performance of DSAC. We demonstrate that by merely decoupling these components, DSAC's performance significantly improves. Motivated by this insight, we introduce a flexible off-policy actor-critic framework that subsumes DSAC as a special case and yields novel objectives. Our framework allows using an m-step Bellman operator for the critic update, and instantiates the actor objective by combining standard policy optimization methods with entropy regularization. Theoretically, we prove that the proposed methods can guarantee convergence to the optimal regularized value function in the tabular setting, generalizing the results in prior work. Empirically, we evaluate the proposed objectives on standard Atari games. Our ablations indicate that, unlike DSAC, these objectives, including novel ones, perform robustly even without entropy regularization or explicit exploration mechanisms.

We develop a theory of intelligent agency grounded in probabilistic modeling for neural models. Agents are represented as outcome distributions with epistemic utility given by log score, and compositions are defined through weighted logarithmic pooling that strictly improves every member's welfare. We prove that strict unanimity is impossible under linear pooling or in binary outcome spaces, but possible with three or more outcomes. Our framework admits recursive structure via cloning invariance, continuity, and openness, while tilt-based analysis rules out trivial duplication. Finally, we formalize an agentic alignment phenomenon in LLMs using our theory: eliciting a benevolent persona ("Luigi'") induces an antagonistic counterpart ("Waluigi"), while a manifest-then-suppress Waluigi strategy yields strictly larger first-order misalignment reduction than pure Luigi reinforcement alone. These results clarify how developing a principled mathematical framework for how subagents can coalesce into coherent higher-level entities provides novel implications for alignment in agentic AI systems.

Many estimation problems in aerospace navigation and robotics involve measurements that depend on prior states. A prominent example is odometry, which measures the relative change between states over time. Accurately handling these delayed-state measurements requires capturing their correlations with prior state estimates, and a widely used approach is stochastic cloning (SC), which augments the state vector to account for these correlations. This work revisits a long-established but often overlooked alternative--the delayed-state Kalman filter--and demonstrates that a properly derived filter yields exactly the same state and covariance update as SC, without requiring state augmentation. Moreover, two equivalent formulations of the delayed-state Kalman filter (DSKF) are presented, providing complementary perspectives on how the prior-state measurement correlations can be handled within the generalized Kalman filter. These formulations are shown to be comparable to SC in asymptotic computational and memory complexity, while one DSKF formulation can offer reduced arithmetic and storage costs for certain problem dimensions. Our findings clarify a common misconception that Kalman filter variants are inherently unable to handle correlated delayed-state measurements, demonstrating that an alternative formulation achieves the same results without state augmentation.

Transformer-based models are at the forefront in long time-series forecasting (LTSF). While in many cases, these models are able to achieve state of the art results, they suffer from a bias toward low-frequencies in the data and high computational and memory requirements. Recent work has established that learnable frequency filters can be an integral part of a deep forecasting model by enhancing the model's spectral utilization. These works choose to use a multilayer perceptron to process their filtered signals and thus do not solve the issues found with transformer-based models. In this paper, we establish that adding a filter to the beginning of transformer-based models enhances their performance in long time-series forecasting. We add learnable filters, which only add an additional $\approx 1000$ parameters to several transformer-based models and observe in multiple instances 5-10 \% relative improvement in forecasting performance. Additionally, we find that with filters added, we are able to decrease the embedding dimension of our models, resulting in transformer-based architectures that are both smaller and more effective than their non-filtering base models. We also conduct synthetic experiments to analyze how the filters enable Transformer-based models to better utilize the full spectrum for forecasting.

Approximately 283 million people worldwide live with visual impairments, motivating increasing research into leveraging Visual Language Models (VLMs) to develop effective walking assistance systems for blind and low vision individuals. However, existing VLMs in walking assistant task often have outputs that contain considerable redundancy and extraneous details, adversely affecting users' ability to accurately assess their surroundings. Moreover, these models typically lack the capability to proactively assess environmental risks and adaptively trigger reminders based on the appropriate scene, leading to excessive temporal redundancy. To mitigate output and temporal redundancy, we propose WalkVLM-LR, a walking assistance model with less redundancy. To reduce output redundancy, we introduce four human-preference-based custom reward functions within the GRPO-based reasoning framework to optimize the output in terms of conciseness, fluency, keyword density, and accuracy, thereby producing more informative and streamlined outputs. To minimize temporal redundancy, we incorporate an environment awareness discriminator, which shares the visual encoder with the VLMs to reduce redundant computations and enhance discriminative efficiency, to make WalkVLM-LR assess scene risk levels and minimize unnecessary reminders. Experimental results demonstrate that our method achieves state-of-the-art performance across all evaluation metrics compared with other models, particularly in output conciseness and less temporal redundancy.

Compression-based dissimilarities (CD) offer a flexible and domain-agnostic means of measuring similarity by identifying implicit information through redundancies between data objects. However, as similarity features are derived from the data, rather than defined as an input, it often proves difficult to align with the task at hand, particularly in complex clustering or classification settings. To address this issue, we introduce "context steering", a novel methodology that actively guides the feature-shaping process. Instead of passively accepting the emergent data structure (typically a hierarchy derived from clustering CDs), our approach "steers" the process by systematically analyzing how each object influences the relational context within a clustering framework. This process generates a custom-tailored embedding that isolates and amplifies class-distinctive information. We validate this supervised context-steering strategy using Normalized Compression Distance (NCD) and Relative Compression Distance (NRC) combined with hierarchical clustering, and evaluate the learned embeddings through both classification performance and cluster-quality metrics. Experiments on heterogeneous datasets-from text to real-world audio-show that the proposed approach yields robust task-oriented embeddings from compression dissimilarities, moving from traditional transductive uses of distance matrices to an inductive representation that can be applied to unseen data.

Large Language Models (LLMs) show remarkable potential for few-shot information extraction (IE), yet their performance is highly sensitive to the choice of in-context examples. Conventional selection strategies often fail to provide informative guidance, as they overlook a key source of model fallibility: confusion stemming not just from semantic content, but also from the generation of well-structured formats required by IE tasks. To address this, we introduce Active Prompting for Information Extraction (APIE), a novel active prompting framework guided by a principle we term introspective confusion. Our method empowers an LLM to assess its own confusion through a dual-component uncertainty metric that uniquely quantifies both Format Uncertainty (difficulty in generating correct syntax) and Content Uncertainty (inconsistency in extracted semantics). By ranking unlabeled data with this comprehensive score, our framework actively selects the most challenging and informative samples to serve as few-shot exemplars. Extensive experiments on four benchmarks show that our approach consistently outperforms strong baselines, yielding significant improvements in both extraction accuracy and robustness. Our work highlights the critical importance of a fine-grained, dual-level view of model uncertainty when it comes to building effective and reliable structured generation systems.

We consider the problem of online regret minimization in linear bandits with access to prior observations (offline data) from the underlying bandit model. There are numerous applications where extensive offline data is often available, such as in recommendation systems, online advertising. Consequently, this problem has been studied intensively in recent literature. Our algorithm, Offline-Online Phased Elimination (OOPE), effectively incorporates the offline data to substantially reduce the online regret compared to prior work. To leverage offline information prudently, OOPE uses an extended D-optimal design within each exploration phase. OOPE achieves an online regret is $\tilde{O}(\sqrt{\deff T \log \left(|\mathcal{A}|T\right)}+d^2)$. $\deff \leq d)$ is the effective problem dimension which measures the number of poorly explored directions in offline data and depends on the eigen-spectrum $(λ_k)_{k \in [d]}$ of the Gram matrix of the offline data. The eigen-spectrum $(λ_k)_{k \in [d]}$ is a quantitative measure of the \emph{quality} of offline data. If the offline data is poorly explored ($\deff \approx d$), we recover the established regret bounds for purely online setting while, when offline data is abundant ($\Toff >> T$) and well-explored ($\deff = o(1) $), the online regret reduces substantially. Additionally, we provide the first known minimax regret lower bounds in this setting that depend explicitly on the quality of the offline data. These lower bounds establish the optimality of our algorithm in regimes where offline data is either well-explored or poorly explored. Finally, by using a Frank-Wolfe approximation to the extended optimal design we further improve the $O(d^{2})$ term to $O\left(\frac{d^{2}}{\deff} \min \{ \deff,1\} \right)$, which can be substantial in high dimensions with moderate quality of offline data $\deff = Ω(1)$.

Understanding dynamic outdoor environments requires capturing complex object interactions and their evolution over time. LiDAR-based 4D point clouds provide precise spatial geometry and rich temporal cues, making them ideal for representing real-world scenes. However, despite their potential, 4D LiDAR remains underexplored in the context of Multimodal Large Language Models (MLLMs) due to the absence of high-quality, modality-specific annotations and the lack of MLLM architectures capable of processing its high-dimensional composition. To address these challenges, we introduce B4DL, a new benchmark specifically designed for training and evaluating MLLMs on 4D LiDAR understanding. In addition, we propose a scalable data generation pipeline and an MLLM model that, for the first time, directly processes raw 4D LiDAR by bridging it with language understanding. Combined with our dataset and benchmark, our model offers a unified solution for spatio-temporal reasoning in dynamic outdoor environments. We provide rendered 4D LiDAR videos, generated dataset, and inference outputs on diverse scenarios at: https://github.com/ccho4702/B4DL

Large language models (LLMs) have proven effective in artificial intelligence, where the multi-agent system (MAS) holds considerable promise for healthcare development by achieving the collaboration of LLMs. However, the absence of a systematic pipeline for agent construction and the rigidity of static collaboration patterns render current MAS-based models vulnerable to collaboration failures, resulting in substantial performance degradation in medical decision-making scenarios. To this end, we propose a novel Masked Agent Collaboration (MAC) framework that harnesses Pareto-optimal agent construction and cross-consistency maximization mechanisms to achieve adaptive progressive propagation of collaborative information, boosting the medical decision-making capacity. Specifically, we first conduct a Pareto-frontier factors analysis towards the LLMs pool to consider their key factors, including the model size, inference time, diversity score, and throughput ratio, where we calculate the similarity between pairwise outputs within an LLM to derive its diversity score. Beyond this analysis, we enable the identification of Pareto-optimal models that balance efficiency and capability, which are subsequently selected as collaborative agents to consider the fundamental trade-offs inherent in practical LLM deployment. Afterward, we measure the pairwise similarity between the outputs from collaborative agents to determine their cross-consistency values, subsequently masking out the agent with the lowest cross-consistency value to eliminate the output that is likely semantically inconsistent. Finally, we conduct collaboration of agents by achieving adaptive progressive propagation, where each agent aggregates the outputs of unmasked agents from the previous layer as its input to generate the corresponding output via prompt engineering.

Information retrieval and question answering from safety regulations are essential for automated construction compliance checking but are hindered by the linguistic and structural complexity of regulatory text. Many queries are multi-hop, requiring synthesis across interlinked clauses. To address the challenge, this paper introduces BifrostRAG, a dual-graph retrieval-augmented generation (RAG) system that models both linguistic relationships and document structure. The proposed architecture supports a hybrid retrieval mechanism that combines graph traversal with vector-based semantic search, enabling large language models to reason over both the content and the structure of the text. On a multi-hop question dataset, BifrostRAG achieves 92.8% precision, 85.5% recall, and an F1 score of 87.3%. These results significantly outperform vector-only and graph-only RAG baselines, establishing BifrostRAG as a robust knowledge engine for LLM-driven compliance checking. The dual-graph, hybrid retrieval mechanism presented in this paper offers a transferable blueprint for navigating complex technical documents across knowledge-intensive engineering domains.

Social interactions play a crucial role in shaping human behavior, relationships, and societies. It encompasses various forms of communication, such as verbal conversation, non-verbal gestures, facial expressions, and body language. In this work, we develop a novel computational approach to detect face-to-face verbal conversations, a foundational aspect of human social interactions. We leverage multimodal data captured by a commodity smartwatch, specifically synchronizing microphone audio with 6-axis inertial signals (accelerometer and gyroscope). We design, train, and evaluate convolutional and attention-based neural networks using three different fusion methods to integrate the audio and motion modalities. To validate this framework, we conduct a lab study with 11 participants and a semi-naturalistic study with 24 participants. Our comprehensive evaluation demonstrates that fusing inertial data with audio significantly improves detection performance by capturing non-verbal conversational dynamics. Overall, our framework achieved 82.0$\pm$3.0% macro F1-score when detecting conversations in the lab and 77.2$\pm$1.8% in the semi-naturalistic setting. Lastly, we demonstrate real-time conversation detection by deploying our trained model to a user application running on a commercial smartwatch.

Accurate short-term state forecasting is essential for efficient and stable operation of modern power systems, especially in the context of increasing variability introduced by renewable and distributed energy resources. As these systems evolve rapidly, it becomes increasingly important to reliably predict their states in the short term to ensure operational stability, support control decisions, and enable interpretable monitoring of sensor and machine behavior. Modern power systems often span multiple physical domains - including electrical, mechanical, hydraulic, and thermal - posing significant challenges for modeling and prediction. Graph Neural Networks (GNNs) have emerged as a promising data-driven framework for system state estimation and state forecasting in such settings. By leveraging the topological structure of sensor networks, GNNs can implicitly learn inter-sensor relationships and propagate information across the network. However, most existing GNN-based methods are designed under the assumption of homogeneous sensor relationships and are typically constrained to a single physical domain. This limitation restricts their ability to integrate and reason over heterogeneous sensor data commonly encountered in real-world energy systems, such as those used in energy conversion infrastructure. In this work, we propose the use of Heterogeneous Graph Attention Networks to address these limitations. Our approach models both homogeneous intra-domain and heterogeneous inter-domain relationships among sensor data from two distinct physical domains - hydraulic and electrical - which exhibit fundamentally different temporal dynamics. Experimental results demonstrate that our method significantly outperforms conventional baselines on average by 35.5% in terms of normalized root mean square error, confirming its effectiveness in multi-domain, multi-rate power system state forecasting.

Accurate individual treatment-effect estimation demands not only reliable point predictions but also uncertainty measures that help practitioners \emph{locate} the source of model failure. We introduce a layer-wise variance decomposition for deep twin-network models: by toggling Monte Carlo Dropout independently in the shared encoder and the outcome heads, we split total predictive variance into an \emph{encoder component} ($σ_{\mathrm{enc}}^2$) and a \emph{head component} ($σ_{\mathrm{head}}^2$), with $σ_{\mathrm{enc}}^2 + σ_{\mathrm{head}}^2 \approx σ_{\mathrm{tot}}^2$ by the law of total variance. Across three synthetic covariate-shift regimes, the encoder component dominates under distributional shift ($ρ_{\mathrm{enc}}=0.53$) while the head component becomes informative only once encoder uncertainty is controlled. On a real-world twins cohort with induced multivariate shift, only $σ_{\mathrm{enc}}^2$ spikes on out-of-distribution samples and becomes the primary error predictor ($ρ_{\mathrm{enc}}\!\approx\!0.89$), while $σ_{\mathrm{head}}^2$ remains flat. The decomposition adds negligible cost over standard MC Dropout and provides a practical diagnostic for deciding whether to collect more diverse covariates or more outcome data.

The adoption of mobile robotic platforms in complex environments, such as agricultural settings, requires these systems to exhibit a flexible yet effective architecture that integrates perception and control. In such scenarios, several tasks need to be accomplished simultaneously, ranging from managing robot limits to performing operational tasks and handling human inputs. The purpose of this paper is to present a comprehensive control architecture for achieving complex tasks such as robotized harvesting in vineyards within the framework of the European project CANOPIES. In detail, a 16-DOF dual-arm mobile robot is employed, controlled via a Hierarchical Quadratic Programming (HQP) approach capable of handling both equality and inequality constraints at various priorities to harvest grape bunches selected by the perception system developed within the project. Furthermore, given the complexity of the scenario and the uncertainty in the perception system, which could potentially lead to collisions with the environment, the handling of interaction forces is necessary. Remarkably, this was achieved using the same HQP framework. This feature is further leveraged to enable semi-autonomous operations, allowing a human operator to assist the robotic counterpart in completing harvesting tasks. Finally, the obtained results are validated through extensive testing conducted first in a laboratory environment to prove individual functionalities, then in a real vineyard, encompassing both autonomous and semi-autonomous grape harvesting operations.

Low-rank adaptation is effective partly because downstream updates lie in a low-dimensional subspace, but the latent rank coordinates of LoRA are not identifiable: any invertible reparameterization of the adapter factors leaves the weight update unchanged. We show that variational inference with a diagonal rank-wise posterior turns this non-identifiability into a useful inductive bias. By breaking LoRA's rotational gauge symmetry, the variational objective selects a preferred basis in rank space, enabling automatic relevance determination over rank directions. This yields Low-Rank Variational Dropout (LRVD), a Bayesian framework that performs inference directly in the low-rank adaptation space rather than the ambient weight space. As an instantiation, BayesLoRA jointly learns effective adapter rank and predictive uncertainty with only $\mathcal{O}(r)$ additional parameters. Empirically, BayesLoRA induces stable rank structure aligned with the dominant singular directions of learned updates, yields compact predictive calibration and matches or exceeds strong low-rank sparsification baselines at comparable training cost.

Cooperative multi-agent reinforcement learning (MARL) under sparse rewards remains fundamentally challenging because agents often fail to concentrate their influence, leading to insufficiently coordinated exploration. To address this, we propose the Focusing Influence Mechanism (FIM), a framework that encourages agents to focus their influence on under-explored parts of the state space through an entropy-based criterion, while leveraging eligibility traces to enable multiple agents to consistently align and sustain their influence on the same parts of the state space when beneficial, thereby promoting coordinated and persistent joint behavior. By emphasizing under-explored regions of the state space, FIM facilitates more efficient and structured exploration even under extremely sparse rewards. Across diverse MARL benchmarks, FIM consistently improves cooperative performance over strong baselines.

This paper reformulates complementarity-based time-stepping for frictionless nonsmooth contact between smooth rigid bodies as a recursively generated linear complementarity problem (ReLCP), involving a sequence of LCPs of increasing dimension. Starting from a classical single-constraint shared-normal signed-distance (SNSD) LCP, the method adds unilateral constraints only when the discrete-time update predicted by the current contact set would violate nonpenetration of the underlying smooth surfaces. The resulting procedure acts directly on smooth geometry, enforces nonpenetration to a prescribed tolerance, and avoids the oversampling inherent to proxy-surface contact models such as tessellations or multi-sphere decompositions, for which improved geometric fidelity can drive rapid growth in constraint count and cost. For strictly convex bodies, we prove that an initially overlap free configuration with sufficiently small timestep sizes, imply finite termination of the adaptive augmentation, and yield a unique discrete-time velocity update. In the small timestep limit and for any fixed overlap-free discrete state with a fixed geometric overlap tolerance, we prove that the recursion terminates after the initial solve, reducing the method to the classical single-constraint SNSD LCP and retaining the usual consistency of complementarity time-stepping with the underlying differential variational inequality. Numerical tests on colliding ellipsoids, compacting ellipsoid suspensions, growing bacterial colonies, and taut chainmail networks demonstrate stable large-timestep behavior, bounded interpenetration without discretization-induced surface roughness, and substantial reductions in both active constraint counts and runtime relative to representative discrete-surface complementarity formulations.

Large-scale pre-training has fundamentally changed how machine learning research is done today: large foundation models are trained once, and then can be used by anyone in the community (including those without data or compute resources to train a model from scratch) to adapt and fine-tune to specific tasks. Applying this same framework to reinforcement learning (RL) is appealing because it offers compelling avenues for addressing core challenges in RL, including sample efficiency and robustness. However, there remains a fundamental challenge to pre-train large models in the context of RL: actions have long-term dependencies, so training a foundation model that reasons across time is important. Recent advances in generative AI have provided new tools for modeling highly complex distributions. In this paper, we build a probabilistic model to predict which states an agent will visit in the temporally distant future (i.e., an occupancy measure) using flow matching. As large datasets are often constructed by many distinct users performing distinct tasks, we include in our model a latent variable capturing the user intention. This intention increases the expressivity of our model, and enables adaptation with generalized policy improvement. We call our proposed method intention-conditioned flow occupancy models (InFOM). Comparing with alternative methods for pre-training, our experiments on $36$ state-based and $4$ image-based benchmark tasks demonstrate that the proposed method achieves $1.8 \times$ median improvement in returns and increases success rates by $36\%$. Website: https://chongyi-zheng.github.io/infom Code: https://github.com/chongyi-zheng/infom

Collision avoidance -- involving a rapid threat detection and quick execution of the appropriate evasive maneuver -- is a critical aspect of driving. However, existing models of human collision avoidance behavior are fragmented, focusing on specific scenarios or only describing certain aspects of the avoidance behavior, such as response times. This paper addresses these gaps by proposing a novel computational cognitive model of human collision avoidance behavior based on active inference. Active inference provides a unified approach to modeling human behavior: the minimization of free energy. Building on prior active inference work, our model incorporates established cognitive mechanisms such as evidence accumulation to simulate human responses in two distinct collision avoidance scenarios: front-to-rear lead vehicle braking and lateral incursion by an oncoming vehicle. We demonstrate that our model explains a wide range of previous empirical findings on human collision avoidance behavior. Specifically, the model closely reproduces both aggregate results from meta-analyses previously reported in the literature and detailed, scenario-specific effects observed in a recent driving simulator study, including response timing, maneuver selection, and execution. Our results highlight the potential of active inference as a unified framework for understanding and modeling human behavior in complex real-life driving tasks.

While the performance of machine learning systems has experienced significant improvement in recent years, relatively little attention has been paid to the fundamental question: to what extent can we improve our models? This paper provides a means of answering this question in the setting of binary classification, which is practical and theoretically supported. We extend a previous work that utilizes soft labels for estimating the Bayes error, the optimal error rate, in two important ways. First, we theoretically investigate the properties of the bias of the hard-label-based estimator discussed in the original work. We reveal that the decay rate of the bias is adaptive to how well the two class-conditional distributions are separated, and it can decay significantly faster than the previous result suggested as the number of hard labels per instance grows. Second, we tackle a more challenging problem setting: estimation with corrupted soft labels. One might be tempted to use calibrated soft labels instead of clean ones. However, we reveal that calibration guarantee is not enough, that is, even perfectly calibrated soft labels can result in a substantially inaccurate estimate. Then, we show that isotonic calibration can provide a statistically consistent estimator under an assumption weaker than that of the previous work. Our method is instance-free, i.e., we do not assume access to any input instances. This feature allows it to be adopted in practical scenarios where the instances are not available due to privacy issues. Experiments with synthetic and real-world datasets show the validity of our methods and theory. The code is available at https://github.com/RyotaUshio/bayes-error-estimation.

We present a fine-grained theoretical analysis of the performance gap between two-stage reinforcement learning from human feedback~(RLHF) and direct preference optimization~(DPO). Our study decomposes this gap into two sources: the explicit representation gap under exact optimization and the implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model, highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.