Structural connectomes are detailed graphs that map how different brain regions are physically connected, offering critical insight into aging, cognition, and neurodegenerative diseases. However, these connectomes are high-dimensional and densely interconnected, which makes them difficult to interpret and analyze at scale. While low-dimensional spaces like PCA and autoencoders are often used to capture major sources of variation, their latent spaces are generally continuous and cannot fully reflect the mixed nature of variability in connectomes, which include both continuous (e.g., connectivity strength) and discrete factors (e.g., imaging site). Motivated by this, we propose a variational autoencoder (VAE) with a hybrid latent space that jointly models the discrete and continuous components. We analyze a large dataset of 5,761 connectomes from six Alzheimer's disease studies with ten acquisition protocols. Each connectome represents a single scan from a unique subject (3579 females, 2182 males), aged 22 to 102, with 4338 cognitively normal, 809 with mild cognitive impairment (MCI), and 614 with Alzheimer's disease (AD). Each connectome contains 121 brain regions defined by the BrainCOLOR atlas. We train our hybrid VAE in an unsupervised way and characterize what each latent component captures. We find that the discrete space is particularly effective at capturing subtle site-related differences, achieving an Adjusted Rand Index (ARI) of 0.65 with site labels, significantly outperforming PCA and a standard VAE followed by clustering (p < 0.05). These results demonstrate that the hybrid latent space can disentangle distinct sources of variability in connectomes in an unsupervised manner, offering potential for large-scale connectome analysis.

Ligand-based drug discovery (LBDD) relies on making use of known binders to a protein target to find structurally diverse molecules similarly likely to bind. This process typically involves a brute force search of the known binder (query) against a molecular library using some metric of molecular similarity. One popular approach overlays the pharmacophore-shape profile of the known binder to 3D conformations enumerated for each of the library molecules, computes overlaps, and picks a set of diverse library molecules with high overlaps. While this virtual screening workflow has had considerable success in hit diversification, scaffold hopping, and patent busting, it scales poorly with library sizes and restricts candidate generation to existing library compounds. Leveraging recent advances in voxel-based generative modelling, we propose a pharmacophore-based generative model and workflows that address the scaling and fecundity issues of conventional pharmacophore-based virtual screening. We introduce \emph{VoxCap}, a voxel captioning method for generating SMILES strings from voxelised molecular representations. We propose two workflows as practical use cases as well as benchmarks for pharmacophore-based generation: \emph{de-novo} design, in which we aim to generate new molecules with high pharmacophore-shape similarities to query molecules, and fast search, which aims to combine generative design with a cheap 2D substructure similarity search for efficient hit identification. Our results show that VoxCap significantly outperforms previous methods in generating diverse \textit{de-novo} hits. When combined with our fast search workflow, VoxCap reduces computational time by orders of magnitude while returning hits for all query molecules, enabling the search of large libraries that are intractable to search by brute force.

Glioblastoma (GBM) remains the most aggressive tumor, urgently requiring novel therapeutic strategies. Here, we present a dry-to-wet framework combining generative modeling and experimental validation to optimize peptides targeting ATP5A, a potential peptide-binding protein for GBM. Our framework introduces the first lead-conditioned generative model, which focuses exploration on geometrically relevant regions around lead peptides and mitigates the combinatorial complexity of de novo methods. Specifically, we propose POTFlow, a \underline{P}rior and \underline{O}ptimal \underline{T}ransport-based \underline{Flow}-matching model for peptide optimization. POTFlow employs secondary structure information (e.g., helix, sheet, loop) as geometric constraints, which are further refined by optimal transport to produce shorter flow paths. With this design, our method achieves state-of-the-art performance compared with five popular approaches. When applied to GBM, our method generates peptides that selectively inhibit cell viability and significantly prolong survival in a patient-derived xenograft (PDX) model. As the first lead peptide-conditioned flow matching model, POTFlow holds strong potential as a generalizable framework for therapeutic peptide design.

Complex spatial connectivity patterns, such as interictal suppression and ictal propagation, complicate accurate drug-resistant epilepsy (DRE) seizure detection using stereotactic electroencephalography (SEEG) and traditional machine learning methods. Two critical challenges remain:(1)a low signal-to-noise ratio in functional connectivity estimates, making it difficult to learn seizure-related interactions; and (2)expert labels for spatial pathological connectivity patterns are difficult to obtain, meanwhile lacking the patterns' representation to improve seizure detection. To address these issues, we propose a novel node-graph dual contrastive learning framework, Seizure-NGCLNet, to learn SEEG interictal suppression and ictal propagation patterns for detecting DRE seizures with high precision. First, an adaptive graph augmentation strategy guided by centrality metrics is developed to generate seizure-related brain networks. Second, a dual-contrastive learning approach is integrated, combining global graph-level contrast with local node-graph contrast, to encode both spatial structural and semantic epileptogenic features. Third, the pretrained embeddings are fine-tuned via a top-k localized graph attention network to perform the final classification. Extensive experiments on a large-scale public SEEG dataset from 33 DRE patients demonstrate that Seizure-NGCLNet achieves state-of-the-art performance, with an average accuracy of 95.93%, sensitivity of 96.25%, and specificity of 94.12%. Visualizations confirm that the learned embeddings clearly separate ictal from interictal states, reflecting suppression and propagation patterns that correspond to the clinical mechanisms. These results highlight Seizure-NGCLNet's ability to learn interpretable spatial pathological patterns, enhancing both seizure detection and seizure onset zone localization.

Accurately recognizing human context from smartphone sensor data remains a significant challenge, especially in sedentary settings where activities such as studying, attending lectures, relaxing, and eating exhibit highly similar inertial patterns. Furthermore, social context plays a critical role in understanding user behavior, yet is often overlooked in mobile sensing research. To address these gaps, we introduce LogMe, a mobile sensing application that passively collects smartphone sensor data (accelerometer, gyroscope, magnetometer, and rotation vector) and prompts users for hourly self-reports capturing both sedentary activity and social context. Using this dual-label dataset, we propose DySTAN (Dynamic Cross-Stitch with Task Attention Network), a multi-task learning framework that jointly classifies both context dimensions from shared sensor inputs. It integrates task-specific layers with cross-task attention to model subtle distinctions effectively. DySTAN improves sedentary activity macro F1 scores by 21.8% over a single-task CNN-BiLSTM-GRU (CBG) model and by 8.2% over the strongest multi-task baseline, Sluice Network (SN). These results demonstrate the importance of modeling multiple, co-occurring context dimensions to improve the accuracy and robustness of mobile context recognition.

The rapid advancement of large language models (LLMs) has accelerated their integration into clinical decision support, particularly in prescription review. To enable systematic and fine-grained evaluation, we developed RxBench, a comprehensive benchmark that covers common prescription review categories and consolidates 14 frequent types of prescription errors drawn from authoritative pharmacy references. RxBench consists of 1,150 single-choice, 230 multiple-choice, and 879 short-answer items, all reviewed by experienced clinical pharmacists. We benchmarked 18 state-of-the-art LLMs and identified clear stratification of performance across tasks. Notably, Gemini-2.5-pro-preview-05-06, Grok-4-0709, and DeepSeek-R1-0528 consistently formed the first tier, outperforming other models in both accuracy and robustness. Comparisons with licensed pharmacists indicated that leading LLMs can match or exceed human performance in certain tasks. Furthermore, building on insights from our benchmark evaluation, we performed targeted fine-tuning on a mid-tier model, resulting in a specialized model that rivals leading general-purpose LLMs in performance on short-answer question tasks. The main contribution of RxBench lies in establishing a standardized, error-type-oriented framework that not only reveals the capabilities and limitations of frontier LLMs in prescription review but also provides a foundational resource for building more reliable and specialized clinical tools.

Reinforcement Learning (RL) following Supervised Fine-Tuning (SFT) has become the standard paradigm for post-training Large Language Models (LLMs). However, the mechanism by which RL contributes to reasoning capabilities-- whether it incentivizes the synthesis of new skills or merely amplifies existing behaviors--remains a subject of intense debate. In this work, we investigate this question through the lens of Complementary Reasoning, a complex task that requires integrating internal parametric knowledge with external contextual information. Using a controlled synthetic dataset of human biographies, we strictly decouple this ability into two atomic skills: Parametric Reasoning (relying on internal knowledge encoded in model parameters) and Contextual Reasoning (depending on novel information provided in the context window). To rigorously assess capability boundaries, we evaluate generalization across three distinct levels of difficulty: I.I.D., Composition, and Zero-shot settings. We find that while SFT is sufficient for in-distribution performance, it struggles with out-of-distribution generalization, particularly in Zero-shot settings where relational combinations are novel. Crucially, we identify the SFT Generalization Paradox: Models supervised solely on the composite task achieve near-perfect in-distribution accuracy (90%) but collapse on out-of-distribution generalization (18%), indicating their reliance on rote memorization of path shortcuts. In contrast, we find that RL acts as a reasoning synthesizer rather than a probability amplifier. However, we uncover a strict atomic prerequisite: RL can only synthesize these complex strategies if the base model has first mastered the independent atomic skills (Parametric and Contextual) via SFT. These findings challenge the view of RL as a mere amplifier, suggesting that given sufficient atomic foundations, RL can actively synthesize complex reasoning strategies from learned primitives without explicit supervision on such complex strategies. This indicates that decoupled atomic training followed by RL offers a scalable path to generalization for complex reasoning tasks. Code and data will be at https://github.com/sitaocheng/from The rapid evolution of Large Language Models (LLMs) has been fundamentally driven by advanced post-training strategies, specifically an initial Supervised Fine-Tuning (SFT) stage followed by a Reinforcement Learning (RL) stage (Achiam et al., 2023; Team et al., 2024; Guo et al., 2025). While SFT is effective at establishing behavioral norms and imparting foundational knowledge, it fundamentally relies on maximum likelihood estimation, which tends to favor the memorization of the training distribution.

The training of deep vision models is fundamentally a signal recovery problem amidst high-dimensional stochastic noise. Current optimization paradigms impose a static compromise on information channel capacity. For instance, magnitude-based methods, such as AdamW, operate on the assumption that gradient norms are high-fidelity curvature signals. While this allows for precision in smooth regimes, it leads to catastrophic noise amplification when applied to rugged, non-convex landscapes. Conversely, sign-based methods (e.g., Lion) perform a radical 1-bit quantization of the gradient, which aims to provide robust regularization at the cost of discarding fine-grained descent information. We propose that optimal convergence requires neither static prior, but rather a dynamic modulation of the update bitrate. We introduce ThermoLion, a vision-centric framework that utilizes local Signal-to-Noise Ratio (SNR) gating to autonomously transition parameters between a "low-bit" exploration phase and a "high-precision" exploitation phase. Furthermore, we introduce a Momentum Alignment mechanism that detects constructive interference between historical drift and instantaneous gradients to accelerate convergence during stable trajectories. Empirical benchmarks across 12 diverse vision datasets (including CIFAR, SVHN, and GTSRB) demonstrate that ThermoLion surpasses state-of-the-art optimizers, such as AdamW and Lion, in convergence speed and terminal accuracy.

Large language models (LLMs) frequently generate hallucinations -- plausible but factually incorrect outputs -- undermining their reliability. While prior work has examined hallucinations from macroscopic perspectives such as training data and objectives, the underlying neuron-level mechanisms remain largely unexplored. In this paper, we conduct a systematic investigation into hallucination-associated neurons (H-Neurons) in LLMs from three perspectives: identification, behavioral impact, and origins. Regarding their identification, we demonstrate that a remarkably sparse subset of neurons (less than $0.1\%$ of total neurons) can reliably predict hallucination occurrences, with strong generalization across diverse scenarios. In terms of behavioral impact, controlled interventions reveal that these neurons are causally linked to over-compliance behaviors. Concerning their origins, we trace these neurons back to the pre-trained base models and find that these neurons remain predictive for hallucination detection, indicating they emerge during pre-training. Our findings bridge macroscopic behavioral patterns with microscopic neural mechanisms, offering insights for developing more reliable LLMs.

We initiate the study of active learning algorithms for classifying strategic agents. Active learning is a well-established framework in machine learning in which the learner selectively queries labels, often achieving substantially higher accuracy and efficiency than classical supervised methods-especially in settings where labeling is costly or time-consuming, such as hiring, admissions, and loan decisions. Strategic classification, however, addresses scenarios where agents modify their features to obtain more favorable outcomes, resulting in observed data that is not truthful. Such manipulation introduces challenges beyond those in learning from clean data. Our goal is to design active and noise-tolerant algorithms that remain effective in strategic environments-algorithms that classify strategic agents accurately while issuing as few label requests as possible. The central difficulty is to simultaneously account for strategic manipulation and preserve the efficiency gains of active learning. Our main result is an algorithm for actively learning linear separators in the strategic setting that preserves the exponential improvement in label complexity over passive learning previously obtained only in the non-strategic case. Specifically, for data drawn uniformly from the unit sphere, we show that a modified version of the Active Perceptron algorithm [DKM05,YZ17] achieves excess error $ε$ using only $\tilde{O}(d \ln \frac{1}ε)$ label queries and incurs at most $\tilde{O}(d \ln \frac{1}ε)$ additional mistakes relative to the optimal classifier, even in the nonrealizable case, when a $\tildeΩ(ε)$ fraction of inputs have inconsistent labels with the optimal classifier. The algorithm is computationally efficient and, under these distributional assumptions, requires substantially fewer label queries than prior work on strategic Perceptron [ABBN21].