CyberSec Research

Flow matching has emerged as a compelling generative modeling approach that is widely used across domains. To generate data via a flow matching model, an ordinary differential equation (ODE) is numerically solved via forward integration of the modeled velocity field. To better capture the multi-modality that is inherent in typical velocity fields, hierarchical flow matching was recently introduced. It uses a hierarchy of ODEs that are numerically integrated when generating data. This hierarchy of ODEs captures the multi-modal velocity distribution just like vanilla flow matching is capable of modeling a multi-modal data distribution. While this hierarchy enables to model multi-modal velocity distributions, the complexity of the modeled distribution remains identical across levels of the hierarchy. In this paper, we study how to gradually adjust the complexity of the distributions across different levels of the hierarchy via mini-batch couplings. We show the benefits of mini-batch couplings in hierarchical rectified flow matching via compelling results on synthetic and imaging data. Code is available at https://riccizz.github.io/HRF_coupling.

Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreover, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. Our code is available at https://github.com/dvlab-research/VisionThink.

Learning visuomotor policies via imitation has proven effective across a wide range of robotic domains. However, the performance of these policies is heavily dependent on the number of training demonstrations, which requires expensive data collection in the real world. In this work, we aim to reduce data collection efforts when learning visuomotor robot policies by leveraging existing or cost-effective data from a wide range of embodiments, such as public robot datasets and the datasets of humans playing with objects (human data from play). Our approach leverages two key insights. First, we use optic flow as an embodiment-agnostic action representation to train a World Model (WM) across multi-embodiment datasets, and finetune it on a small amount of robot data from the target embodiment. Second, we develop a method, Latent Policy Steering (LPS), to improve the output of a behavior-cloned policy by searching in the latent space of the WM for better action sequences. In real world experiments, we observe significant improvements in the performance of policies trained with a small amount of data (over 50% relative improvement with 30 demonstrations and over 20% relative improvement with 50 demonstrations) by combining the policy with a WM pretrained on two thousand episodes sampled from the existing Open X-embodiment dataset across different robots or a cost-effective human dataset from play.

Neural networks are often highly sensitive to input and weight perturbations. This sensitivity has been linked to pathologies such as vulnerability to adversarial examples, divergent training, and overfitting. To combat these problems, past research has looked at building neural networks entirely from Lipschitz components. However, these techniques have not matured to the point where researchers have trained a modern architecture such as a transformer with a Lipschitz certificate enforced beyond initialization. To explore this gap, we begin by developing and benchmarking novel, computationally-efficient tools for maintaining norm-constrained weight matrices. Applying these tools, we are able to train transformer models with Lipschitz bounds enforced throughout training. We find that optimizer dynamics matter: switching from AdamW to Muon improves standard methods -- weight decay and spectral normalization -- allowing models to reach equal performance with a lower Lipschitz bound. Inspired by Muon's update having a fixed spectral norm, we co-design a weight constraint method that improves the Lipschitz vs. performance tradeoff on MLPs and 2M parameter transformers. Our 2-Lipschitz transformer on Shakespeare text reaches validation accuracy 60%. Scaling to 145M parameters, our 10-Lipschitz transformer reaches 21% accuracy on internet text. However, to match the NanoGPT baseline validation accuracy of 39.4%, our Lipschitz upper bound increases to 10^264. Nonetheless, our Lipschitz transformers train without stability measures such as layer norm, QK norm, and logit tanh softcapping.

Socio-economic indicators like regional GDP, population, and education levels, are crucial to shaping policy decisions and fostering sustainable development. This research introduces GeoReg a regression model that integrates diverse data sources, including satellite imagery and web-based geospatial information, to estimate these indicators even for data-scarce regions such as developing countries. Our approach leverages the prior knowledge of large language model (LLM) to address the scarcity of labeled data, with the LLM functioning as a data engineer by extracting informative features to enable effective estimation in few-shot settings. Specifically, our model obtains contextual relationships between data features and the target indicator, categorizing their correlations as positive, negative, mixed, or irrelevant. These features are then fed into the linear estimator with tailored weight constraints for each category. To capture nonlinear patterns, the model also identifies meaningful feature interactions and integrates them, along with nonlinear transformations. Experiments across three countries at different stages of development demonstrate that our model outperforms baselines in estimating socio-economic indicators, even for low-income countries with limited data availability.

Team modeling remains a fundamental challenge at the intersection of Artificial Intelligence and the Social Sciences. Social Science research emphasizes the need to jointly model dynamics and relations, while practical applications demand unified models capable of inferring multiple team constructs simultaneously, providing interpretable insights and actionable recommendations to enhance team performance. However, existing works do not meet these practical demands. To bridge this gap, we present TRENN, a novel tempo-relational architecture that integrates: (i) an automatic temporal graph extractor, (ii) a tempo-relational encoder, (iii) a decoder for team construct prediction, and (iv) two complementary explainability modules. TRENN jointly captures relational and temporal team dynamics, providing a solid foundation for MT-TRENN, which extends TReNN by replacing the decoder with a multi-task head, enabling the model to learn shared Social Embeddings and simultaneously predict multiple team constructs, including Emergent Leadership, Leadership Style, and Teamwork components. Experimental results demonstrate that our approach significantly outperforms approaches that rely exclusively on temporal or relational information. Additionally, experimental evaluation has shown that the explainability modules integrated in MT-TRENN yield interpretable insights and actionable suggestions to support team improvement. These capabilities make our approach particularly well-suited for Human-Centered AI applications, such as intelligent decision-support systems in high-stakes collaborative environments.

Temporal distribution shifts pose a key challenge for machine learning models trained and deployed in dynamically evolving environments. This paper introduces RIDER (RIsk minimization under Dynamically Evolving Regimes) which derives optimally-weighted empirical risk minimization procedures under temporal distribution shifts. Our approach is theoretically grounded in the random distribution shift model, where random shifts arise as a superposition of numerous unpredictable changes in the data-generating process. We show that common weighting schemes, such as pooling all data, exponentially weighting data, and using only the most recent data, emerge naturally as special cases in our framework. We demonstrate that RIDER consistently improves out-of-sample predictive performance when applied as a fine-tuning step on the Yearbook dataset, across a range of benchmark methods in Wild-Time. Moreover, we show that RIDER outperforms standard weighting strategies in two other real-world tasks: predicting stock market volatility and forecasting ride durations in NYC taxi data.

An increasing number of studies have focused on stochastic first-order methods (SFOMs) under heavy-tailed gradient noises, which have been observed in the training of practical deep learning models. In this paper, we focus on two types of gradient noises: one is sub-Weibull noise, and the other is noise under the assumption that it has a bounded $p$-th central moment ($p$-BCM) with $p\in (1, 2]$. The latter is more challenging due to the occurrence of infinite variance when $p\in (1, 2)$. Under these two gradient noise assumptions, the in-expectation and high-probability convergence of SFOMs have been extensively studied in the contexts of convex optimization and standard smooth optimization. However, for weakly convex objectives-a class that includes all Lipschitz-continuous convex objectives and smooth objectives-our understanding of the in-expectation and high-probability convergence of SFOMs under these two types of noises remains incomplete. We investigate the high-probability convergence of the vanilla stochastic subgradient descent (SsGD) method under sub-Weibull noises, as well as the high-probability and in-expectation convergence of clipped SsGD under the $p$-BCM noises. Both analyses are conducted in the context of weakly convex optimization. For weakly convex objectives that may be non-convex and non-smooth, our results demonstrate that the theoretical dependence of vanilla SsGD on the failure probability and number of iterations under sub-Weibull noises does not degrade compared to the case of smooth objectives. Under $p$-BCM noises, our findings indicate that the non-smoothness and non-convexity of weakly convex objectives do not impact the theoretical dependence of clipped SGD on the failure probability relative to the smooth case; however, the sample complexity we derived is worse than a well-known lower bound for smooth optimization.

Robots are increasingly integrated across industries, particularly in healthcare. However, many valuable applications for quadrupedal robots remain overlooked. This research explores the effectiveness of three reinforcement learning algorithms in training a simulated quadruped robot for autonomous navigation and obstacle avoidance. The goal is to develop a robotic guide dog simulation capable of path following and obstacle avoidance, with long-term potential for real-world assistance to guide dogs and visually impaired individuals. It also seeks to expand research into medical 'pets', including robotic guide and alert dogs. A comparative analysis of thirteen related research papers shaped key evaluation criteria, including collision detection, pathfinding algorithms, sensor usage, robot type, and simulation platforms. The study focuses on sensor inputs, collision frequency, reward signals, and learning progression to determine which algorithm best supports robotic navigation in complex environments. Custom-made environments were used to ensure fair evaluation of all three algorithms under controlled conditions, allowing consistent data collection. Results show that Proximal Policy Optimization (PPO) outperformed Deep Q-Network (DQN) and Q-learning across all metrics, particularly in average and median steps to goal per episode. By analysing these results, this study contributes to robotic navigation, AI and medical robotics, offering insights into the feasibility of AI-driven quadruped mobility and its role in assistive robotics.

Bayesian Optimization (BO) algorithm is a standard tool for black-box optimization problems. The current state-of-the-art BO approach for permutation spaces relies on the Mallows kernel-an $\Omega(n^2)$ representation that explicitly enumerates every pairwise comparison. Inspired by the close relationship between the Mallows kernel and pairwise comparison, we propose a novel framework for generating kernel functions on permutation space based on sorting algorithms. Within this framework, the Mallows kernel can be viewed as a special instance derived from bubble sort. Further, we introduce the \textbf{Merge Kernel} constructed from merge sort, which replaces the quadratic complexity with $\Theta(n\log n)$ to achieve the lowest possible complexity. The resulting feature vector is significantly shorter, can be computed in linearithmic time, yet still efficiently captures meaningful permutation distances. To boost robustness and right-invariance without sacrificing compactness, we further incorporate three lightweight, task-agnostic descriptors: (1) a shift histogram, which aggregates absolute element displacements and supplies a global misplacement signal; (2) a split-pair line, which encodes selected long-range comparisons by aligning elements across the two halves of the whole permutation; and (3) sliding-window motifs, which summarize local order patterns that influence near-neighbor objectives. Our empirical evaluation demonstrates that the proposed kernel consistently outperforms the state-of-the-art Mallows kernel across various permutation optimization benchmarks. Results confirm that the Merge Kernel provides a more compact yet more effective solution for Bayesian optimization in permutation space.

Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.

Accurate short-term electricity price forecasting is crucial for strategically scheduling demand and generation bids in day-ahead markets. While data-driven techniques have shown considerable prowess in achieving high forecast accuracy in recent years, they rely heavily on the quality of input covariates. In this paper, we investigate whether asynchronously published prices as a result of differing gate closure times (GCTs) in some bidding zones can improve forecasting accuracy in other markets with later GCTs. Using a state-of-the-art ensemble of models, we show significant improvements of 22% and 9% in forecast accuracy in the Belgian (BE) and Swedish bidding zones (SE3) respectively, when including price data from interconnected markets with earlier GCT (Germany-Luxembourg, Austria, and Switzerland). This improvement holds for both general as well as extreme market conditions. Our analysis also yields further important insights: frequent model recalibration is necessary for maximum accuracy but comes at substantial additional computational costs, and using data from more markets does not always lead to better performance - a fact we delve deeper into with interpretability analysis of the forecast models. Overall, these findings provide valuable guidance for market participants and decision-makers aiming to optimize bidding strategies within increasingly interconnected and volatile European energy markets.

Computational screening has become a powerful complement to experimental efforts in the discovery of high-performance photovoltaic (PV) materials. Most workflows rely on density functional theory (DFT) to estimate electronic and optical properties relevant to solar energy conversion. Although more efficient than laboratory-based methods, DFT calculations still entail substantial computational and environmental costs. Machine learning (ML) models have recently gained attention as surrogates for DFT, offering drastic reductions in resource use with competitive predictive performance. In this study, we reproduce a canonical DFT-based workflow to estimate the maximum efficiency limit and progressively replace its components with ML surrogates. By quantifying the CO$_2$ emissions associated with each computational strategy, we evaluate the trade-offs between predictive efficacy and environmental cost. Our results reveal multiple hybrid ML/DFT strategies that optimize different points along the accuracy--emissions front. We find that direct prediction of scalar quantities, such as maximum efficiency, is significantly more tractable than using predicted absorption spectra as an intermediate step. Interestingly, ML models trained on DFT data can outperform DFT workflows using alternative exchange--correlation functionals in screening applications, highlighting the consistency and utility of data-driven approaches. We also assess strategies to improve ML-driven screening through expanded datasets and improved model architectures tailored to PV-relevant features. This work provides a quantitative framework for building low-emission, high-throughput discovery pipelines.

A central question in computer science and statistics is whether efficient algorithms can achieve the information-theoretic limits of statistical problems. Many computational-statistical tradeoffs have been shown under average-case assumptions, but since statistical problems are average-case in nature, it has been a challenge to base them on standard worst-case assumptions. In PAC learning where such tradeoffs were first studied, the question is whether computational efficiency can come at the cost of using more samples than information-theoretically necessary. We base such tradeoffs on $\mathsf{NP}$-hardness and obtain: $\circ$ Sharp computational-statistical tradeoffs assuming $\mathsf{NP}$ requires exponential time: For every polynomial $p(n)$, there is an $n$-variate class $C$ with VC dimension $1$ such that the sample complexity of time-efficiently learning $C$ is $\Theta(p(n))$. $\circ$ A characterization of $\mathsf{RP}$ vs. $\mathsf{NP}$ in terms of learning: $\mathsf{RP} = \mathsf{NP}$ iff every $\mathsf{NP}$-enumerable class is learnable with $O(\mathrm{VCdim}(C))$ samples in polynomial time. The forward implication has been known since (Pitt and Valiant, 1988); we prove the reverse implication. Notably, all our lower bounds hold against improper learners. These are the first $\mathsf{NP}$-hardness results for improperly learning a subclass of polynomial-size circuits, circumventing formal barriers of Applebaum, Barak, and Xiao (2008).

Recent years have witnessed a growing interest for time series foundation models, with a strong emphasis on the forecasting task. Yet, the crucial task of out-of-domain imputation of missing values remains largely underexplored. We propose a first step to fill this gap by leveraging implicit neural representations (INRs). INRs model time series as continuous functions and naturally handle various missing data scenarios and sampling rates. While they have shown strong performance within specific distributions, they struggle under distribution shifts. To address this, we introduce MoTM (Mixture of Timeflow Models), a step toward a foundation model for time series imputation. Building on the idea that a new time series is a mixture of previously seen patterns, MoTM combines a basis of INRs, each trained independently on a distinct family of time series, with a ridge regressor that adapts to the observed context at inference. We demonstrate robust in-domain and out-of-domain generalization across diverse imputation scenarios (e.g., block and pointwise missingness, variable sampling rates), paving the way for adaptable foundation imputation models.

The Sliced Gromov-Wasserstein (SGW) distance, aiming to relieve the computational cost of solving a non-convex quadratic program that is the Gromov-Wasserstein distance, utilizes projecting directions sampled uniformly from unit hyperspheres. This slicing mechanism incurs unnecessary computational costs due to uninformative directions, which also affects the representative power of the distance. However, finding a more appropriate distribution over the projecting directions (slicing distribution) is often an optimization problem in itself that comes with its own computational cost. In addition, with more intricate distributions, the sampling itself may be expensive. As a remedy, we propose an optimization-free slicing distribution that provides fast sampling for the Monte Carlo approximation. We do so by introducing the Relation-Aware Projecting Direction (RAPD), effectively capturing the pairwise association of each of two pairs of random vectors, each following their ambient law. This enables us to derive the Relation-Aware Slicing Distribution (RASD), a location-scale law corresponding to sampled RAPDs. Finally, we introduce the RASGW distance and its variants, e.g., IWRASGW (Importance Weighted RASGW), which overcome the shortcomings experienced by SGW. We theoretically analyze its properties and substantiate its empirical prowess using extensive experiments on various alignment tasks.

Monotone gradient functions play a central role in solving the Monge formulation of the optimal transport problem, which arises in modern applications ranging from fluid dynamics to robot swarm control. When the transport cost is the squared Euclidean distance, Brenier's theorem guarantees that the unique optimal map is the gradient of a convex function, namely a monotone gradient map, and it satisfies a Monge-Amp\`ere equation. In [arXiv:2301.10862] [arXiv:2404.07361], we proposed Monotone Gradient Networks (mGradNets), neural networks that directly parameterize the space of monotone gradient maps. In this work, we leverage mGradNets to directly learn the optimal transport mapping by minimizing a training loss function defined using the Monge-Amp\`ere equation. We empirically show that the structural bias of mGradNets facilitates the learning of optimal transport maps and employ our method for a robot swarm control problem.

The effect of representation has been demonstrated in reinforcement learning, from both theoretical and empirical successes. However, the existing representation learning mainly induced from model learning aspects, misaligning with our RL tasks. This work introduces Spectral Bellman Representation, a novel framework derived from the Inherent Bellman Error (IBE) condition, which aligns with the fundamental structure of Bellman updates across a space of possible value functions, therefore, directly towards value-based RL. Our key insight is the discovery of a fundamental spectral relationship: under the zero-IBE condition, the transformation of a distribution of value functions by the Bellman operator is intrinsically linked to the feature covariance structure. This spectral connection yields a new, theoretically-grounded objective for learning state-action features that inherently capture this Bellman-aligned covariance. Our method requires a simple modification to existing algorithms. We demonstrate that our learned representations enable structured exploration, by aligning feature covariance with Bellman dynamics, and improve overall performance, particularly in challenging hard-exploration and long-horizon credit assignment tasks. Our framework naturally extends to powerful multi-step Bellman operators, further broadening its impact. Spectral Bellman Representation offers a principled and effective path toward learning more powerful and structurally sound representations for value-based reinforcement learning.