Assessing fairness in artificial intelligence (AI) typically involves AI experts who select protected features, fairness metrics, and set fairness thresholds to assess outcome fairness. However, little is known about how stakeholders, particularly those affected by AI outcomes but lacking AI expertise, assess fairness. To address this gap, we conducted a qualitative study with 26 stakeholders without AI expertise, representing potential decision subjects in a credit rating scenario, to examine how they assess fairness when placed in the role of deciding on features with priority, metrics, and thresholds. We reveal that stakeholders' fairness decisions are more complex than typical AI expert practices: they considered features far beyond legally protected features, tailored metrics for specific contexts, set diverse yet stricter fairness thresholds, and even preferred designing customized fairness. Our results extend the understanding of how stakeholders can meaningfully contribute to AI fairness governance and mitigation, underscoring the importance of incorporating stakeholders' nuanced fairness judgments.

Modern computer vision is converging on a closed loop in which perception, reasoning and generation mutually reinforce each other. However, this loop remains incomplete: the top-down influence of high-level reasoning on the foundational learning of low-level perceptual features is not yet underexplored. This paper addresses this gap by proposing a new paradigm for pretraining foundation models in downstream domains. We introduce Visual insTruction Pretraining (ViTP), a novel approach that directly leverages reasoning to enhance perception. ViTP embeds a Vision Transformer (ViT) backbone within a Vision-Language Model and pretrains it end-to-end using a rich corpus of visual instruction data curated from target downstream domains. ViTP is powered by our proposed Visual Robustness Learning (VRL), which compels the ViT to learn robust and domain-relevant features from a sparse set of visual tokens. Extensive experiments on 16 challenging remote sensing and medical imaging benchmarks demonstrate that ViTP establishes new state-of-the-art performance across a diverse range of downstream tasks. The code is available at https://github.com/zcablii/ViTP.

3D occupancy prediction is critical for comprehensive scene understanding in vision-centric autonomous driving. Recent advances have explored utilizing 3D semantic Gaussians to model occupancy while reducing computational overhead, but they remain constrained by insufficient multi-view spatial interaction and limited multi-frame temporal consistency. To overcome these issues, in this paper, we propose a novel Spatial-Temporal Gaussian Splatting (ST-GS) framework to enhance both spatial and temporal modeling in existing Gaussian-based pipelines. Specifically, we develop a guidance-informed spatial aggregation strategy within a dual-mode attention mechanism to strengthen spatial interaction in Gaussian representations. Furthermore, we introduce a geometry-aware temporal fusion scheme that effectively leverages historical context to improve temporal continuity in scene completion. Extensive experiments on the large-scale nuScenes occupancy prediction benchmark showcase that our proposed approach not only achieves state-of-the-art performance but also delivers markedly better temporal consistency compared to existing Gaussian-based methods.

Single-cell RNA-seq provides detailed molecular snapshots of individual cells but is notoriously noisy. Variability stems from biological differences and technical factors, such as amplification bias and limited RNA capture efficiency, making it challenging to adapt computational pipelines to heterogeneous datasets or evolving technologies. As a result, most studies still rely on principal component analysis (PCA) for dimensionality reduction, valued for its interpretability and robustness, in spite of its known bias in high dimensions. Here, we improve upon PCA with a Random Matrix Theory (RMT)-based approach that guides the inference of sparse principal components using existing sparse PCA algorithms. We first introduce a novel biwhitening algorithm which self-consistently estimates the magnitude of transcriptomic noise affecting each gene in individual cells, without assuming a specific noise distribution. This enables the use of an RMT-based criterion to automatically select the sparsity level, rendering sparse PCA nearly parameter-free. Our mathematically grounded approach retains the interpretability of PCA while enabling robust, hands-off inference of sparse principal components. Across seven single-cell RNA-seq technologies and four sparse PCA algorithms, we show that this method systematically improves the reconstruction of the principal subspace and consistently outperforms PCA-, autoencoder-, and diffusion-based methods in cell-type classification tasks.

This work addresses the challenge of forecasting urban water dynamics by developing a multi-input, multi-output deep learning model that incorporates both endogenous variables (e.g., water height or discharge) and exogenous factors (e.g., precipitation history and forecast reports). Unlike conventional forecasting, the proposed model, AquaCast, captures both inter-variable and temporal dependencies across all inputs, while focusing forecast solely on endogenous variables. Exogenous inputs are fused via an embedding layer, eliminating the need to forecast them and enabling the model to attend to their short-term influences more effectively. We evaluate our approach on the LausanneCity dataset, which includes measurements from four urban drainage sensors, and demonstrate state-of-the-art performance when using only endogenous variables. Performance also improves with the inclusion of exogenous variables and forecast reports. To assess generalization and scalability, we additionally test the model on three large-scale synthesized datasets, generated from MeteoSwiss records, the Lorenz Attractors model, and the Random Fields model, each representing a different level of temporal complexity across 100 nodes. The results confirm that our model consistently outperforms existing baselines and maintains a robust and accurate forecast across both real and synthetic datasets.

In this study, we propose FHIR-RAG-MEDS system that aims to integrate Health Level 7 Fast Healthcare Interoperability Resources (HL7 FHIR) with a Retrieval-Augmented Generation (RAG)-based system to improve personalized medical decision support on evidence-based clinical guidelines, emphasizing the need for research in practical applications. In the evolving landscape of medical decision support systems, integrating advanced technologies such as RAG and HL7 FHIR can significantly enhance clinical decision-making processes. Despite the potential of these technologies, there is limited research on their integration in practical applications.

Multimodal large language models (MLLMs) are rapidly evolving, presenting increasingly complex safety challenges. However, current dataset construction methods, which are risk-oriented, fail to cover the growing complexity of real-world multimodal safety scenarios (RMS). And due to the lack of a unified evaluation metric, their overall effectiveness remains unproven. This paper introduces a novel image-oriented self-adaptive dataset construction method for RMS, which starts with images and end constructing paired text and guidance responses. Using the image-oriented method, we automatically generate an RMS dataset comprising 35k image-text pairs with guidance responses. Additionally, we introduce a standardized safety dataset evaluation metric: fine-tuning a safety judge model and evaluating its capabilities on other safety datasets.Extensive experiments on various tasks demonstrate the effectiveness of the proposed image-oriented pipeline. The results confirm the scalability and effectiveness of the image-oriented approach, offering a new perspective for the construction of real-world multimodal safety datasets. The dataset is presented at https://huggingface.co/datasets/NewCityLetter/RMS2/tree/main.

Typographic attacks exploit multi-modal systems by injecting text into images, leading to targeted misclassifications, malicious content generation and even Vision-Language Model jailbreaks. In this work, we analyze how CLIP vision encoders behave under typographic attacks, locating specialized attention heads in the latter half of the model's layers that causally extract and transmit typographic information to the cls token. Building on these insights, we introduce Dyslexify - a method to defend CLIP models against typographic attacks by selectively ablating a typographic circuit, consisting of attention heads. Without requiring finetuning, dyslexify improves performance by up to 22.06% on a typographic variant of ImageNet-100, while reducing standard ImageNet-100 accuracy by less than 1%, and demonstrate its utility in a medical foundation model for skin lesion diagnosis. Notably, our training-free approach remains competitive with current state-of-the-art typographic defenses that rely on finetuning. To this end, we release a family of dyslexic CLIP models which are significantly more robust against typographic attacks. These models serve as suitable drop-in replacements for a broad range of safety-critical applications, where the risks of text-based manipulation outweigh the utility of text recognition.

A novel methodology for short-term energy forecasting using an Extreme Learning Machine ($\mathtt{ELM}$) is proposed. Using six years of hourly data collected in Corsica (France) from multiple energy sources (solar, wind, hydro, thermal, bioenergy, and imported electricity), our approach predicts both individual energy outputs and total production (including imports, which closely follow energy demand, modulo losses) through a Multi-Input Multi-Output ($\mathtt{MIMO}$) architecture. To address non-stationarity and seasonal variability, sliding window techniques and cyclic time encoding are incorporated, enabling dynamic adaptation to fluctuations. The $\mathtt{ELM}$ model significantly outperforms persistence-based forecasting, particularly for solar and thermal energy, achieving an $\mathtt{nRMSE}$ of $17.9\%$ and $5.1\%$, respectively, with $\mathtt{R^2} > 0.98$ (1-hour horizon). The model maintains high accuracy up to five hours ahead, beyond which renewable energy sources become increasingly volatile. While $\mathtt{MIMO}$ provides marginal gains over Single-Input Single-Output ($\mathtt{SISO}$) architectures and offers key advantages over deep learning methods such as $\mathtt{LSTM}$, it provides a closed-form solution with lower computational demands, making it well-suited for real-time applications, including online learning. Beyond predictive accuracy, the proposed methodology is adaptable to various contexts and datasets, as it can be tuned to local constraints such as resource availability, grid characteristics, and market structures.

Text-to-video generation has advanced rapidly, but existing methods typically output only the final composited video and lack editable layered representations, limiting their use in professional workflows. We propose \textbf{LayerT2V}, a unified multi-layer video generation framework that produces multiple semantically consistent outputs in a single inference pass: the full video, an independent background layer, and multiple foreground RGB layers with corresponding alpha mattes. Our key insight is that recent video generation backbones use high compression in both time and space, enabling us to serialize multiple layer representations along the temporal dimension and jointly model them on a shared generation trajectory. This turns cross-layer consistency into an intrinsic objective, improving semantic alignment and temporal coherence. To mitigate layer ambiguity and conditional leakage, we augment a shared DiT backbone with LayerAdaLN and layer-aware cross-attention modulation. LayerT2V is trained in three stages: alpha mask VAE adaptation, joint multi-layer learning, and multi-foreground extension. We also introduce \textbf{VidLayer}, the first large-scale dataset for multi-layer video generation. Extensive experiments demonstrate that LayerT2V substantially outperforms prior methods in visual fidelity, temporal consistency, and cross-layer coherence.

Training deep generative models like Variational Autoencoders (VAEs) requires propagating gradients through stochastic latent variables, which introduces estimation variance that can slow convergence and degrade performance. In this paper, we explore an orthogonal direction, which we call Silent Gradients. Instead of designing improved stochastic estimators, we show that by restricting the decoder architecture in specific ways, the expected ELBO can be computed analytically. This yields gradients with zero estimation variance as we can directly compute the evidence lower-bound without resorting to Monte Carlo samples of the latent variables. We first provide a theoretical analysis in a controlled setting with a linear decoder and demonstrate improved optimization compared to standard estimators. To extend this idea to expressive nonlinear decoders, we introduce a training paradigm that uses the analytic gradient to guide early encoder learning before annealing to a standard stochastic estimator. Across multiple datasets, our approach consistently improves established baselines, including reparameterization, Gumbel-Softmax, and REINFORCE. These results suggest that architectural choices enabling analytic expectation computation can significantly stabilize the training of generative models with stochastic components.

Nonlinear Model Predictive Control (NMPC) is a powerful approach for controlling highly dynamic robotic systems, as it accounts for system dynamics and optimizes control inputs at each step. However, its high computational complexity makes implementation on resource-constrained microcontrollers impractical. While recent studies have demonstrated the feasibility of Model Predictive Control (MPC) with linearized dynamics on microcontrollers, applying full NMPC remains a significant challenge. This work presents an efficient solution for generating and deploying NMPC on microcontrollers (NMPCM) to control quadrotor UAVs. The proposed method optimizes computational efficiency while maintaining high control accuracy. Simulations in Gazebo/ROS and real-world experiments validate the effectiveness of the approach, demonstrating its capability to achieve high-frequency NMPC execution in real-time systems. The code is available at: https://github.com/aralab-unr/NMPCM.

There are currently two main paradigms for evaluating large language models (LLMs), reference-based evaluation and preference-based evaluation. The first, carried over from the evaluation of machine learning models in general, relies on pre-defined task instances, for which reference task executions are available. The second, best exemplified by the LM-arena, relies on (often self-selected) users bringing their own intents to a site that routes these to several models in parallel, among whose responses the user then selects their most preferred one. The former paradigm hence excels at control over what is tested, while the latter comes with higher ecological validity, testing actual use cases interactively. Recently, a third complementary paradigm has emerged that combines some of the strengths of these approaches, offering control over multi-turn, reference-free, repeatable interactions, while stressing goal-directedness: dialogue game based evaluation. While the utility of this approach has been shown by several projects, its adoption has been held back by the lack of a mature, easily re-usable implementation. In this paper, we present clembench, which has been in continuous development since 2023 and has in its latest release been optimized for ease of general use. We describe how it can be used to benchmark one's own models (using a provided set of benchmark game instances in English), as well as how easily the benchmark itself can be extended with new, tailor-made targeted tests.

The evaluation of large language model (LLM) outputs is increasingly performed by other LLMs, a setup commonly known as "LLM-as-a-judge", or autograders. While autograders offer a scalable alternative to human evaluation, they have shown mixed reliability and may exhibit systematic biases, depending on response type, scoring methodology, domain specificity, or other factors. Here we propose a statistical framework based on Bayesian generalised linear models (GLMs) that enables researchers to simultaneously assess their autograders while addressing their primary research questions (e.g., LLM evaluation). Our approach models evaluation outcomes (e.g., scores or pairwise preferences) as a function of properties of the grader (e.g., human vs. autograder) and the evaluated item (e.g., response length or the LLM that generated it), allowing for explicit quantification of scoring differences and potential biases within a unified framework. In addition, our method can be used to augment traditional metrics such as inter-rater agreement, by providing uncertainty estimates and clarifying sources of disagreement. Overall, this approach contributes to more robust and interpretable use of autograders in LLM evaluation, enabling both performance analysis and bias detection.

Fairness in machine learning (ML) has become a rapidly growing area of research. But why, in the first place, is unfairness in ML wrong? And why should we care about improving fairness? Most fair-ML research implicitly appeals to distributive equality: the idea that desirable benefits and goods, such as opportunities (e.g., Barocas et al., 2023), should be equally distributed across society. Unfair ML models, then, are seen as wrong because they unequally distribute such benefits. This paper argues that this exclusive focus on distributive equality offers an incomplete and potentially misleading ethical foundation. Grounding ML fairness in egalitarianism--the view that equality is a fundamental moral and social ideal--requires challenging structural inequality: systematic, institutional, and durable arrangements that privilege some groups while disadvantaging others. Structural inequality manifests through ML systems in two primary forms: allocative harms (e.g., economic loss) and representational harms (e.g., stereotypes, erasure). While distributive equality helps address allocative harms, it fails to explain why representational harms are wrong--why it is wrong for ML systems to reinforce social hierarchies that stratify people into superior and inferior groups--and why ML systems should aim to foster a society where people relate as equals (i.e., relational equality). To address these limitations, the paper proposes a multifaceted egalitarian framework for ML fairness that integrates both distributive and relational equality. Drawing on critical social and political philosophy, this framework offers a more comprehensive ethical foundation for tackling the full spectrum of harms perpetuated by ML systems. The paper also outlines practical pathways for implementing the framework across the entire ML pipeline.

Large language models (LLMs) are increasingly applied in clinical decision support, yet current evaluations rarely reveal whether their outputs reflect genuine medical reasoning or superficial correlations. We introduce DeVisE (Demographics and Vital signs Evaluation), a behavioral testing framework that probes fine-grained clinical understanding through controlled counterfactuals. Using intensive care unit (ICU) discharge notes from MIMIC-IV, we construct both raw (real-world) and template-based (synthetic) variants with single-variable perturbations in demographic (age, gender, ethnicity) and vital sign attributes. We evaluate eight LLMs, spanning general-purpose and medical variants, under zero-shot setting. Model behavior is analyzed through (1) input-level sensitivity, capturing how counterfactuals alter perplexity, and (2) downstream reasoning, measuring their effect on predicted ICU length-of-stay and mortality. Overall, our results show that standard task metrics obscure clinically relevant differences in model behavior, with models differing substantially in how consistently and proportionally they adjust predictions to counterfactual perturbations.

Animals flexibly recombine a finite set of core motor motifs to meet diverse task demands, but existing behavior segmentation methods oversimplify this process by imposing discrete syllables under restrictive generative assumptions. To better capture the continuous structure of behavior generation, we introduce motif-based continuous dynamics (MCD) discovery, a framework that (1) uncovers interpretable motif sets as latent basis functions of behavior by leveraging representations of behavioral transition structure, and (2) models behavioral dynamics as continuously evolving mixtures of these motifs. We validate MCD on a multi-task gridworld, a labyrinth navigation task, and freely moving animal behavior. Across settings, it identifies reusable motif components, captures continuous compositional dynamics, and generates realistic trajectories beyond the capabilities of traditional discrete segmentation models. By providing a generative account of how complex animal behaviors emerge from dynamic combinations of fundamental motor motifs, our approach advances the quantitative study of natural behavior.

World model based planning has significantly improved decision-making in complex environments by enabling agents to simulate future states and make informed choices. This computational burden is particularly restrictive in robotics, where resources are severely constrained. To address this limitation, we propose a Sparse Imagination for Efficient Visual World Model Planning, which enhances computational efficiency by reducing the number of tokens processed during forward prediction. Our method leverages a sparsely trained vision-based world model based on transformers with randomized grouped attention strategy, allowing the model to flexibly adjust the number of tokens processed based on the computational resource. By enabling sparse imagination during latent rollout, our approach significantly accelerates planning while maintaining high control fidelity. Experimental results demonstrate that sparse imagination preserves task performance while dramatically improving inference efficiency. This general technique for visual planning is applicable from simple test-time trajectory optimization to complex real-world tasks with the latest VLAs, enabling the deployment of world models in real-time scenarios.

In order to behave intelligently both humans and machines have to represent their knowledge adequately for how it is used. Humans often use analogies to transfer their knowledge to new domains, or help others with this transfer via explanations. Hence, an important question is: What representation can be used to construct, find, and evaluate analogies? In this paper, we study features of a domain that are important for constructing analogies. We do so by formalizing knowledge domains as categories. We use the well-known example of the analogy between the solar system and the hydrogen atom to demonstrate how to construct domain categories. We also show how functors, pullbacks, and pushouts can be used to define an analogy, describe its core and a corresponding blend of the underlying domains.

We present a novel geometric perspective on the latent space of diffusion models. We first show that the standard pullback approach, utilizing the deterministic probability flow ODE decoder, is fundamentally flawed. It provably forces geodesics to decode as straight segments in data space, effectively ignoring any intrinsic data geometry beyond the ambient Euclidean space. Complementing this view, diffusion also admits a stochastic decoder via the reverse SDE, which enables an information geometric treatment with the Fisher-Rao metric. However, a choice of $x_T$ as the latent representation collapses this metric due to memorylessness. We address this by introducing a latent spacetime $z=(x_t,t)$ that indexes the family of denoising distributions $p(x_0 | x_t)$ across all noise scales, yielding a nontrivial geometric structure. We prove these distributions form an exponential family and derive simulation-free estimators for curve lengths, enabling efficient geodesic computation. The resulting structure induces a principled Diffusion Edit Distance, where geodesics trace minimal sequences of noise and denoise edits between data. We also demonstrate benefits for transition path sampling in molecular systems, including constrained variants such as low-variance transitions and region avoidance. Code is available at: https://github.com/rafalkarczewski/spacetime-geometry.

Regarding intelligent transportation systems, low-bitrate transmission via lossy point cloud compression is vital for facilitating real-time collaborative perception among connected agents, such as vehicles and infrastructures, under restricted bandwidth. In existing compression transmission systems, the sender lossily compresses point coordinates and reflectance to generate a transmission code stream, which faces transmission burdens from reflectance encoding and limited detection robustness due to information loss. To address these issues, this paper proposes a 3D object detection framework with reflectance prediction-based knowledge distillation (RPKD). We compress point coordinates while discarding reflectance during low-bitrate transmission, and feed the decoded non-reflectance compressed point clouds into a student detector. The discarded reflectance is then reconstructed by a geometry-based reflectance prediction (RP) module within the student detector for precise detection. A teacher detector with the same structure as the student detector is designed for performing reflectance knowledge distillation (RKD) and detection knowledge distillation (DKD) from raw to compressed point clouds. Our cross-source distillation training strategy (CDTS) equips the student detector with robustness to low-quality compressed data while preserving the accuracy benefits of raw data through transferred distillation knowledge. Experimental results on the KITTI and DAIR-V2X-V datasets demonstrate that our method can boost detection accuracy for compressed point clouds across multiple code rates. We will release the code publicly at https://github.com/HaoJing-SX/RPKD.

Large language models (LLMs) are increasingly explored as substitutes for human participants in cognitive tasks, but their ability to simulate human behavioral variability remains unclear. This study examines whether LLMs can approximate individual differences in the phonemic fluency task, where participants generate words beginning with a target letter. We evaluated 34 distinct models across 45 configurations from major closed-source and open-source providers, and compared outputs to responses from 106 human participants. While some models, especially Claude 3.7 Sonnet, approximated human averages and lexical preferences, none reproduced the scope of human variability. LLM outputs were consistently less diverse, with newer models and thinking-enabled modes often reducing rather than increasing variability. Network analysis further revealed fundamental differences in retrieval structure between humans and the most human-like model. Ensemble simulations combining outputs from diverse models also failed to recover human-level diversity, likely due to high vocabulary overlap across models. These results highlight key limitations in using LLMs to simulate human cognition and behavior.

We address the problem of inferring the causal direction between a continuous variable $X$ and a discrete variable $Y$ from observational data. For the model $X \to Y$, we adopt the threshold model used in prior work. For the model $Y \to X$, we consider two cases: (1) the conditional distributions of $X$ given different values of $Y$ form a location-shift family, and (2) they are mixtures of generalized normal distributions with independently parameterized components. We establish identifiability of the causal direction through three theoretical results. First, we prove that under $X \to Y$, the density ratio of $X$ conditioned on different values of $Y$ is monotonic. Second, we establish that under $Y \to X$ with non-location-shift conditionals, monotonicity of the density ratio holds only on a set of Lebesgue measure zero in the parameter space. Third, we show that under $X \to Y$, the conditional distributions forming a location-shift family requires a precise coordination between the causal mechanism and input distribution, which is non-generic under the principle of independent mechanisms. Together, these results imply that monotonicity of the density ratio characterizes the direction $X \to Y$, whereas non-monotonicity or location-shift conditionals characterizes $Y \to X$. Based on this, we propose Density Ratio-based Causal Discovery (DRCD), a method that determines causal direction by testing for location-shift conditionals and monotonicity of the estimated density ratio. Experiments on synthetic and real-world datasets demonstrate that DRCD outperforms existing methods.

Detecting AI-synthetic faces presents a critical challenge: it is hard to capture consistent structural relationships between facial regions across diverse generation techniques. Current methods, which focus on specific artifacts rather than fundamental inconsistencies, often fail when confronted with novel generative models. To address this limitation, we introduce Layer-aware Mask Modulation Vision Transformer (LAMM-ViT), a Vision Transformer designed for robust facial forgery detection. This model integrates distinct Region-Guided Multi-Head Attention (RG-MHA) and Layer-aware Mask Modulation (LAMM) components within each layer. RG-MHA utilizes facial landmarks to create regional attention masks, guiding the model to scrutinize architectural inconsistencies across different facial areas. Crucially, the separate LAMM module dynamically generates layer-specific parameters, including mask weights and gating values, based on network context. These parameters then modulate the behavior of RG-MHA, enabling adaptive adjustment of regional focus across network depths. This architecture facilitates the capture of subtle, hierarchical forgery cues ubiquitous among diverse generation techniques, such as GANs and Diffusion Models. In cross-model generalization tests, LAMM-ViT demonstrates superior performance, achieving 94.09% mean ACC (a +5.45% improvement over SoTA) and 98.62% mean AP (a +3.09% improvement). These results demonstrate LAMM-ViT's exceptional ability to generalize and its potential for reliable deployment against evolving synthetic media threats.

We introduce a framework for robust uncertainty quantification in situations where labeled training data are corrupted, through noisy or missing labels. We build on conformal prediction, a statistical tool for generating prediction sets that cover the test label with a pre-specified probability. The validity of conformal prediction, however, holds under the i.i.d assumption, which does not hold in our setting due to the corruptions in the data. To account for this distribution shift, the privileged conformal prediction (PCP) method proposed leveraging privileged information (PI) -- additional features available only during training -- to re-weight the data distribution, yielding valid prediction sets under the assumption that the weights are accurate. In this work, we analyze the robustness of PCP to inaccuracies in the weights. Our analysis indicates that PCP can still yield valid uncertainty estimates even when the weights are poorly estimated. Furthermore, we introduce uncertain imputation (UI), a new conformal method that does not rely on weight estimation. Instead, we impute corrupted labels in a way that preserves their uncertainty. Our approach is supported by theoretical guarantees and validated empirically on both synthetic and real benchmarks. Finally, we show that these techniques can be integrated into a triply robust framework, ensuring statistically valid predictions as long as at least one underlying method is valid.

In this paper, we tackle the problem of learning to play 3v3 multi-drone volleyball, a new embodied competitive task that requires both high-level strategic coordination and low-level agile control. The task is turn-based, multi-agent, and physically grounded, posing significant challenges due to its long-horizon dependencies, tight inter-agent coupling, and the underactuated dynamics of quadrotors. To address this, we propose Hierarchical Co-Self-Play (HCSP), a hierarchical reinforcement learning framework that separates centralized high-level strategic decision-making from decentralized low-level motion control. We design a three-stage population-based training pipeline to enable both strategy and skill to emerge from scratch without expert demonstrations: (I) training diverse low-level skills, (II) learning high-level strategy via self-play with fixed low-level skills, and (III) joint fine-tuning through co-self-play. Experiments show that HCSP achieves superior performance, outperforming non-hierarchical self-play and rule-based hierarchical baselines with an average 82.9% win rate and a 71.5% win rate against the two-stage variant. Moreover, co-self-play leads to emergent team behaviors such as role switching and coordinated formations, demonstrating the effectiveness of our hierarchical design and training scheme. The project page is at https://hi-co-self-play.github.io.

Low-rank and sparse composite approximation is a natural idea to compress Large Language Models (LLMs). However, such an idea faces two primary challenges that adversely affect the performance of existing methods. The first challenge relates to the interaction and cooperation between low-rank and sparse matrices, while the second involves determining weight allocation across different layers, as redundancy varies considerably among them. To address these challenges, we propose a novel two-stage LLM compression method with the capability of global resource allocation for rank and sparsity. It is noteworthy that the overall optimization space is vast, making comprehensive optimization computationally prohibitive. Therefore, to reduce the optimization space, our first stage utilizes robust principal component analysis to decompose the weight matrices of LLMs into low-rank and sparse components, which span the low dimensional and sparse spaces containing the resultant low-rank and sparse matrices, respectively. In the second stage, we propose a probabilistic global allocation strategy to jointly identify the low-rank and sparse structures within the above two spaces. The appealing feature of our approach is its ability to automatically detect the redundancy across different layers and to manage the interaction between the sparse and low-rank components. Extensive experimental results indicate that our method significantly surpasses state-of-the-art techniques for sparsification and composite approximation.

Recent work suggests that preference-tuning techniques -- such as Reinforcement Learning from Human Feedback (RLHF) methods like PPO and GRPO, as well as alternatives like DPO -- reduce diversity, creating a dilemma given that these models are widely deployed in applications requiring varied outputs. We argue that diversity without consideration of quality has limited practical value. To address this issue, we introduce a framework for measuring effective semantic diversity -- diversity among outputs that meet quality thresholds -- which better reflects the practical utility of large language models (LLMs). Using open-ended tasks that require no human intervention, we find counterintuitive results: when using diversity metrics that do not explicitly consider quality, preference-tuned models -- particularly those trained via RL -- often produce outputs with lower diversity; however, these same preference-tuned models generate greater effective semantic diversity than supervised fine-tuned (SFT) or base models. Our analysis further shows another trend: while larger models may exhibit greater effective semantic diversity than smaller models, the smaller models are consistently more parameter-efficient at producing unique content within a fixed sampling budget. These findings have practical implications for applications that require diverse yet high-quality outputs, from creative assistance to synthetic data generation.

Systematic generalization refers to the capacity to understand and generate novel combinations from known components. Despite recent progress by large language models (LLMs) across various domains, these models often fail to extend their knowledge to novel compositional scenarios, revealing notable limitations in systematic generalization. There has been an ongoing debate about whether neural networks possess the capacity for systematic generalization, with recent studies suggesting that meta-learning approaches designed for compositionality can significantly enhance this ability. However, these insights have largely been confined to linguistic problems, leaving their applicability to other tasks an open question. In this study, we extend meta-learning for compositionality to the domain of abstract spatial reasoning. To this end, we introduce $\textit{Compositional-ARC}\unicode{x2014}$a dataset designed to evaluate the capacity of models to systematically generalize from known geometric transformations (e.g., translation, rotation) of abstract two-dimensional objects to novel combinations of these transformations (e.g., translation+rotation). Our results show that a small transformer-based encoder-decoder model, trained via meta-learning for compositionality, can systematically generalize to previously unseen transformation compositions. Notably, despite having only 5.7M parameters, this model significantly outperforms state-of-the-art LLMs$\unicode{x2014}$including o3-mini, GPT-4o, and Gemini 2.0 Flash, which fail to exhibit similar systematic behavior$\unicode{x2014}$and performs on par with the winning model of the ARC prize 2024, an 8B-parameter LLM trained via test-time training. Our findings highlight the effectiveness of meta-learning in promoting systematicity beyond linguistic tasks, suggesting a promising direction toward more robust and generalizable models.

Gating mechanisms have emerged as an effective strategy integrated into model designs beyond recurrent neural networks for addressing long-range dependency problems. In a broad understanding, it provides adaptive control over the information flow while maintaining computational efficiency. However, there is a lack of theoretical analysis on how the gating mechanism works in neural networks. In this paper, inspired by the \textit{convolution theorem}, we systematically explore the effect of gating mechanisms on the training dynamics of neural networks from a frequency perspective. We investigate the interact between the element-wise product and activation functions in managing the responses to different frequency components. Leveraging these insights, we propose a Gating Mechanism Network (GmNet), a lightweight model designed to efficiently utilize the information of various frequency components. It minimizes the low-frequency bias present in existing lightweight models. GmNet achieves impressive performance in terms of both effectiveness and efficiency in the image classification task.