Generating realistic, context-aware two-person motion conditioned on diverse modalities remains a fundamental challenge for graphics, animation and embodied AI systems. Real-world applications such as VR/AR companions, social robotics and game agents require models capable of producing coordinated interpersonal behavior while flexibly switching between interactive and reactive generation. We introduce DualFlow, the first unified and efficient framework for multi-modal two-person motion generation. DualFlow conditions 3D motion generation on diverse inputs, including text, music, and prior motion sequences. Leveraging rectified flow, it achieves deterministic straight-line sampling paths between noise and data, reducing inference time and mitigating error accumulation common in diffusion-based models. To enhance semantic grounding, DualFlow employs a novel Retrieval-Augmented Generation (RAG) module for two-person motion that retrieves motion exemplars using music features and LLM-based text decompositions of spatial relations, body movements, and rhythmic patterns. We use contrastive rectified flow objective to further sharpen alignment with conditioning signals and add synchronization loss to improve inter-person temporal coordination. Extensive evaluations across interactive, reactive, and multi-modal benchmarks demonstrate that DualFlow consistently improves motion quality, responsiveness, and semantic fidelity. DualFlow achieves state-of-the-art performance in two-person multi-modal motion generation, producing coherent, expressive, and rhythmically synchronized motion. Code will be released upon acceptance.

Reinforcement learning (RL) has unlocked complex reasoning abilities in large language models (LLMs). However, most RL algorithms suffer from performance saturation, preventing further gains as RL training scales. This problem can be characterized by the collapse of entropy, a key diagnostic for exploration in RL. Existing attempts have tried to prevent entropy collapse through regularization or clipping, but their resulting entropy curves often exhibit instability in the long term, which hinders performance gains. In this paper, we introduce Entrocraft, a simple rejection-sampling approach that realizes any user-customized entropy schedule by biasing the advantage distributions. Entrocraft requires no objective regularization and is advantage-estimator-agnostic. Theoretically, we relate per-step entropy change to the advantage distribution under minimal assumptions, which explains the behavior of existing RL and entropy-preserving methods. Entrocraft also enables a systematic study of entropy schedules, where we find that linear annealing, which starts high and decays to a slightly lower target, performs best. Empirically, Entrocraft addresses performance saturation, significantly improving generalization, output diversity, and long-term training. It enables a 4B model to outperform an 8B baseline, sustains improvement for up to 4x longer before plateauing, and raises pass@K by 50

ABSTRACT High-dimensional population balance equations (PBEs) provide a natural framework for modeling heterogeneous cell populations, but their direct numerical solution becomes computationally prohibitive when the internal state space contains many molecular variables. We propose a hybrid mechanistic-machine learning framework for reducing and simulating PBEs defined over high-dimensional intracellular coordinates. The cell population is described by a number density n (x, t ), where x ∈ ℝ N represents gene and protein states associated with macrophage activation. A dynamics-preserving autoencoder maps this state space to a low-dimensional latent coordinate z ∈ ℝ d , with d ≪ N , while retaining key qualitative features of the underlying gene regulatory network, including attractor structure and multistability. Mechanistic information from the original regulatory dynamics is used to construct interpretable drift and diffusion terms for the reduced latent-space PBE. The reduced PBE is solved using a stochastic Lagrangian particle representation, in which particles evolve according to stochastic differential equations (SDEs) corresponding to the latent drift and diffusion fields. The resulting latent-space solution is subsequently decoded and propagated back into the original state space to recover physically interpretable cellular dynamics. We demonstrate the framework on macrophage polarization under cytokine-dependent regulation, including gene knockout perturbations. Overall, the proposed framework provides a computationally tractable and mechanistically interpretable route for integrating single-cell genomic data with population balance models of cell-state dynamics.

Motivation: Predicting cellular responses to drug perturbations requires capturing complex dependencies between transcriptomic and morphological changes. Existing approaches model these modalities in isolation, missing critical molecular-phenotypic relationships that occur simultaneously during drug treatment. No current method jointly predicts gene expression profiles and cellular morphology from chemical perturbations. Results: We introduce PertFlow, a unified computational framework that simultaneously predicts treatment gene expression (bulk RNA-seq) and generates cellular morphology (Cell Painting images) from control cellular states, conditioned on drug metadata. Evaluated on paired RNA-seq and imaging data from 3 cell lines and 40 compounds (17,242 samples), PertFlow achieves Pearson correlation of 0.780+-0.264 for transcriptomic prediction and FID of 24.06 for morphological generation. Board-certified pathologists rated generated images with median similarity scores of 7.11-7.89/10. The model successfully recovers known drug mechanisms including microtubule disruption, DNA damage, and MAPK pathway inhibition, with gene enrichment analysis confirming activation of expected biological pathways (EMT, p53, apoptosis). Availability: Code and pretrained models will be available at https://github.com/wangmengbo/PertFlow. ### Competing Interest Statement The authors have declared no competing interest.

"Best-of-N" selection is a popular inference-time scaling method for code generation using Large Language Models (LLMs). However, to reliably identify correct solutions, existing methods often depend on expensive or stochastic external verifiers. In this paper, we propose Symbolic Equivalence Partitioning, a selection framework that uses symbolic execution to group candidate programs by semantic behavior and select a representative from the dominant functional partition. To improve grouping and selection, we encode domain-specific constraints as Satisfiability Modulo Theories (SMT) assumptions during symbolic execution to reduce path explosion and prevent invalid input searches outside the problem domain. At N=10, our method improves average accuracy over Pass@1 from 0.728 to 0.803 on HumanEval+ and from 0.516 to 0.604 on LiveCodeBench, without requiring any additional LLM inference beyond the initial N candidate generations.

Whether citations objectively and reliably reflect the quality of articles and researchers is questionable. Even so, citation counts are widely used to estimate the productivity of researchers and institutions, which creates a ‘grubby’ motivation to be well-cited. We examine this motivation using a generative model of citation that is agent-based. In this model, new nodes are added to an existing citation network. These new nodes act as autonomous agents that cite other nodes based on a composite bias for preferential attachment, recency, fitness (epistemic quality), and community structure. We use the model to ask whether strategic citation behaviors can support an interest in being well-cited. Results from this model suggest that while fitness is influential, the number of references and community effects are also influential in attracting citations. These results raise questions about similar effects in the real world.

Reinforcement learning has become central to improving large reasoning models, but its success still relies heavily on verifiable rewards or labeled supervision. This limits its applicability to open ended domains where correctness is ambiguous and cannot be verified. Moreover, reasoning trajectories remain largely unconstrained, and optimization towards final answer can favor early exploitation over generalization. In this work, we ask whether general reasoning ability can be improved by teaching models how to think (the structure of reasoning) rather than what to produce (the outcome of reasoning) and extend traditional RLVR to open ended settings. We introduce structure aware reinforcement learning (SARL), a label free framework that constructs a per response Reasoning Map from intermediate thinking steps and rewards its small world topology, inspired by complex networks and the functional organization of the human brain. SARL encourages reasoning trajectories that are both locally coherent and globally efficient, shifting supervision from destination to path. Our experiments on Qwen3-4B show SARL surpasses ground truth based RL and prior label free RL baselines, achieving the best average gain of 9.1

Online learning is a foundational paradigm underlying applications from recommendation systems to the continual learning of modern AI models. Yet much of its theory centers on either fully adversarial or purely stochastic settings. However, real-world environments typically fall between these extremes, making classical models inadequate for describing practical behavior. This monograph develops a unified perspective for analyzing online learning under more nuanced and realistic environments. The authors approach the problem through the lens of universality from information theory and extend tools such as the Shtarkov sum, covering numbers and packing arguments to the online setting, revealing deeper structural connections between these two fields. Building on this viewpoint, they characterize minimax regret for logarithmic and Lipschitz losses, analyze expected regret under i.i.d. and more general stochastic processes and study hybrid adversarial–stochastic scenarios. The authors further develop constructive algorithms that achieve near-optimal regret guarantees, yielding a coherent and fine-grained information-theoretic framework of online universal learning.

Electric Vehicles (EVs) offer pathways to lower emissions and increased energy efficiency. However, a broad and equitable adoption of electric vehicles can only be realized with well-planned growth in public charging infrastructure. We propose EV-Planner, a software tool that formulates EV charging station placement as a constrained multi-objective optimization problem and computes approximate solutions to these problems. In contrast to previous efforts, EV-Planner considers multiple criteria, including minimizing average user-to-station distance, maximizing fairness in coverage, and balancing load across stations. We propose a clustering-based approximation to solve the constrained multi-objective problem, along with a Suggest-Accept (SA) criterion that iterates over two steps: (i) generation of suitable clusters; and (ii) acceptance of a subset of clusters based on the specified cost function and constraints. To tune the hyperparameters in SA clustering, we develop a neural network model and an associated training procedure. Based on real-world data from 10 US states, EV-Planner reduces average distances to charging stations by 52.3%, the number of users who are more than three miles from a charging station by up to 10.7 times, and overloaded EV stations by 71.7% over current baselines, offering a promising solution to the problem of planning EV infrastructure. The code is available at https://github.com/bharathanand0/EV-Planner.

Motivation Phenotypic drug discovery generates rich multi-modal biological data from transcriptomic and morphological measurements, yet translating complex cellular responses into molecular design remains a computational bottleneck. Existing generative methods operate on single modalities and condition on post-treatment measurements without leveraging paired control-treatment dynamics to capture perturbation effects. Results We present Pert2Mol, the first framework for multi-modal phenotype-to-structure generation that integrates transcriptomic and morphological features from paired control-treatment experiments. Pert2Mol employs bidirectional cross-attention between control and treatment states to capture perturbation dynamics, conditioning a rectified flow transformer that generates molecular structures along straight-line trajectories. We introduce Student-Teacher Self-Representation (SERE) learning to stabilize training in high-dimensional multi-modal spaces. On the GDP dataset, Pert2Mol achieves Fréchet ChemNet Distance of 4.996 compared to 7.343 for diffusion baselines and 59.114 for transcriptomics-only methods, while maintaining perfect molecular validity and appropriate physicochemical property distributions. The model demonstrates 84.7% scaffold diversity and 12.4 times faster generation than diffusion approaches with deterministic sampling suitable for hypothesis-driven validation. Availability Code and pretrained models will be available at . ### Competing Interest Statement The authors have declared no competing interest.

Real-world large language model deployments (e.g., conversational AI systems, code generation assistants) naturally generate abundant implicit user dissatisfaction (DSAT) signals, as users iterate toward better answers through refinements, corrections, and expressed preferences, while explicit satisfaction (SAT) feedback is scarce. Existing preference learning approaches are poorly aligned with this data profile, as they rely on costly human annotations or assume plentiful positive responses. In this paper, we introduce \textbf{DRIFT} (\textbf{D}issatisfaction-\textbf{R}efined \textbf{I}terative pre\textbf{F}erence \textbf{T}raining), which anchors training on real-world DSAT signals and samples positives dynamically from the evolving policy. Empirically, DRIFT models trained on real-world \textit{WildFeedback} datasets and synthetic \textit{UltraFeedback} datasets achieve up to +6.23\% (7B) / +7.61\% (14B) on WildBench Task Score and up to +8.95\% (7B) / +12.29\% (14B) on AlpacaEval2 win rate over base models, outperforming strong baseline methods such as iterative DPO and SPIN. At larger scales, the improvements are particularly pronounced: 14B models trained with DRIFT surpass GPT-4o-mini on WildBench. Further analysis shows that DRIFT also preserves exploratory capacity, yielding more diverse high-reward solutions rather than collapsing to narrow subsets. Theoretically, we demonstrate that this design preserves preference margins and avoids the gradient degeneration. These results show that DRIFT is an effective and scalable recipe for real-world post-training that leverages the most abundant and informative signal.

Aligning large language models (LLMs) with human values is an increasingly critical step in post-training. Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback (RLHF). Synthetic preference data with its low cost and high quality enable effective alignment through single- or multi-model generated preference data. Our study reveals a striking, safety-specific phenomenon associated with DPO alignment: Although multi-model generated data enhances performance on general tasks (ARC, Hellaswag, MMLU, TruthfulQA, Winogrande) by providing diverse responses, it also tends to facilitate reward hacking during training. This can lead to a high attack success rate (ASR) when models encounter jailbreaking prompts. The issue is particularly pronounced when employing stronger models like GPT-4o or larger models in the same family to generate chosen responses paired with target model self-generated rejected responses, resulting in dramatically poorer safety outcomes. Furthermore, with respect to safety, using solely self-generated responses (single-model generation) for both chosen and rejected pairs significantly outperforms configurations that incorporate responses from stronger models, whether used directly as chosen data or as part of a multi-model response pool. We demonstrate that multi-model preference data exhibits high linear separability between chosen and rejected responses, which allows models to exploit superficial cues rather than internalizing robust safety constraints. Our experiments, conducted on models from the Llama, Mistral, and Qwen families, consistently validate these findings.

Generative models have shown impressive capabilities in synthesizing high-quality outputs across various domains. However, a persistent challenge is the occurrence of "hallucinations", where the model produces outputs that are plausible but invalid. While empirical strategies have been explored to mitigate this issue, a rigorous theoretical understanding remains elusive. In this paper, we develop a theoretical framework to analyze the learnability of non-hallucinating generative models from a learning-theoretic perspective. Our results reveal that non-hallucinating learning is statistically impossible when relying solely on the training dataset, even for a hypothesis class of size two and when the entire training set is truthful. To overcome these limitations, we show that incorporating inductive biases aligned with the actual facts into the learning process is essential. We provide a systematic approach to achieve this by restricting the facts set to a concept class of finite VC-dimension and demonstrate its effectiveness under various learning paradigms. Although our findings are primarily conceptual, they represent a first step towards a principled approach to addressing hallucinations in learning generative models.

Seizures affect millions worldwide, significantly impacting quality of life and increasing the risk of premature death. Although seizure-like activity in neuronal networks is commonly characterized by synchronized firing, the mechanisms underlying network reorganization and dynamics remain incompletely understood. A better understanding of how network firing changes during seizure-like activity can improve diagnosis and treatment approaches for seizures. Here, we combine an in vitro model of primary cortical networks cultured on microelectrode arrays with a multilayered machine learning (ML) pipeline to investigate bicuculline-induced seizure-like activity. Our multistep analysis approach, which consists of using a long short-term memory (LSTM) autoencoder for dimensionality reduction, followed by uniform manifold approximation and projection (UMAP) and hierarchical clustering, revealed the emergence of distinct neuronal subpopulations with characteristic activity profiles after seizure-like activity induction, even within globally synchronized network activity states. Furthermore, deep Granger causality, an advanced analysis technique for identifying predictive relationships in data, applied to our nonlinear time series, revealed disproportionate neuronal responses to seizure-like activity-driven network firing changes. We also trained ML classifiers to distinguish native and seizure-like activity states with high accuracy using different firing features. Spike rate was the most significant feature for achieving high classification accuracy. Ultimately, these findings demonstrate the power of our analytical framework for characterizing seizure-like activity. Our low-cost, two-dimensional model of seizure-on-a-chip, combined with novel performance metrics, could serve as a valuable tool for screening potential new antiepileptic drugs and for gaining a deeper understanding of how seizure-like activity alters the functional organization of neuronal networks.NEW & NOTEWORTHY This study introduces a novel approach for analyzing seizure-like activity in neuronal networks using a multilayered machine learning pipeline. We show that even during drug-induced network disinhibition leading to global network synchronization, neurons organize into distinct, spatially distributed subnetworks with unique firing profiles. Our innovative seizure-on-a-chip model enables the comprehensive investigation of network reorganization and dynamics during seizure-like activity, which could potentially be useful for understanding pathophysiology, developing diagnostic tools, and screening new antiepileptic drugs.

Eigenvalue problems serve as fundamental substrates for applications in large-scale scientific simulations and machine learning, often requiring computation on massively parallel platforms. As these platforms scale to hundreds of thousands of cores, hardware failures become a significant challenge to reliability and efficiency. In this paper, we propose and analyze a novel fault-tolerant eigenvalue solver based on erasure-coded computations – a technique that enhances resilience by augmenting the system with redundant data a priori. This transformation reformulates the original eigenvalue problem as a generalized eigenvalue problem, enabling fault-oblivious computation while preserving numerical stability and convergence properties. We formulate the augmentation scheme, establish the necessary conditions for the encoded blocks, and prove the relationship between the original and transformed problems. We implement an erasure-coded TraceMin eigensolver and demonstrate its effectiveness in extracting eigenvalues in the presence of faults. Our experimental results show that the proposed solver incurs minimal computational overhead, maintains robust convergence, and scales efficiently with the number of faults, making it a practical solution for resilient eigenvalue computations in large-scale systems.

We study agnostic online learning from continuous-time data streams, a setting that naturally arises in applications such as environmental monitoring, personalized recommendation, and high-frequency trading. Unlike classical discrete-time models, learners in this setting must interact with a continually evolving data stream while making queries and updating models only at sparse, strategically selected times. We develop a general theoretical framework for learning from both *oblivious* and *adaptive* data streams, which may be noisy and non-stationary. For oblivious streams, we present a black-box reduction to classical online learning that yields a regret bound of $T \cdot R(S)/S$ for any class with discrete-time regret $R(S)$, where $T$ is the time horizon and $S$ is the *query budget*. For adaptive streams, which can evolve in response to learner actions, we design a dynamic query strategy in conjunction with a novel importance weighting scheme that enables unbiased loss estimation. In particular, for hypothesis class $\mathcal{H}$ with a finite Littlestone dimension, we establish a tight regret bound of $\tilde{\Theta}(T \cdot \sqrt{\mathsf{Ldim}(\mathcal{H})/S})$ that holds in both settings. Our results provide the first *quantitative* characterization of agnostic learning in continuous-time online environments with limited interaction.

In recent years, Graph Convolutional Networks (GCNs) and their variants have been widely utilized in learning tasks that involve graphs. These tasks include recommendation systems, node classification, among many others. In node classification problem, the input is a graph in which the edges represent the association between pairs of nodes, multi-dimensional feature vectors are associated with the nodes, and some of the nodes in the graph have known labels. The objective is to predict the labels of the nodes that are not labeled, using the nodes features, in conjunction with graph topology. While GCNs have been successfully applied to this problem, the caveats that they inherit from traditional deep learning models pose significant challenges to broad utilization of GCNs in node classification. One such caveat is that training a GCN requires a large number of labeled training instances, which is often not the case in realistic settings. To remedy this requirement, state-of-the-art methods leverage network diffusion-based approaches to propagate labels across the network before training GCNs. However, these approaches ignore the tendency of the network diffusion methods in biasing proximity with centrality, resulting in the propagation of labels to the nodes that are well-connected in the graph. To address this problem, here we present an alternate approach to extrapolating node labels in GCNs in the following three steps: (i) clustering of the network to identify communities, (ii) use of network diffusion algorithms to quantify the proximity of each node to the communities, thereby obtaining a low-dimensional topological profile for each node, (iii) comparing these topological profiles to identify nodes that are most similar to the labeled nodes.

Noninvasive neural recording methods like electroencephalography (EEG) offer high temporal resolution for capturing neural activity. However, interpreting EEG data is challenging scalp-recorded signals (sensor space) reflect complex, integrated activity from multiple cortical regions (source space), complicating the reconstruction of underlying neural dynamics. Traditional approaches like minimum norm estimation require extensive subject-specific data, including MRI scans, precise electrode placement, and detailed anatomical atlases. To address these limitations, we propose a two-part framework: (1) an unsupervised biLSTM autoencoder that reveals clustering patterns in EEG electrode activations and their temporal dynamics during auditory stimulus processing; and (2) a deep learning architecture to predict temporally evoked neural features in sensor space EEG from source representations using a dual-path network with independent stimulus processing and dilated convolutional layers.The clustering identifies evolving spatiotemporal co-activation patterns between stimulus onset and gaps, revealing functional reorganization. The reconstruction network reduces input dimensionality and integrates features via convolutional blocks with residual connections, trained using a hybrid loss that combines feature-based and spectral terms. Our results demonstrate accurate reconstruction of stimulus-related neural correlates and reveal topographical patterns consistent with the clustering findings. The model generalizes well across subjects. By analyzing both functional organization in sensor signals and source-to-sensor mappings, our framework enhances understanding of EEG transformations. This has significant implications for brain-computer interfaces, neuroimaging, and EEG processing where accurate reconstruction and interpretation are essential.