Multimodal reasoning has become a cornerstone of modern AI research. Standardized exam questions offer a uniquely rigorous testbed for such reasoning, providing structured visual contexts and verifiable answers. While recent progress has largely focused on algorithmic advances such as reinforcement learning (e.g., GRPO, DPO), the data centric foundations of vision language reasoning remain less explored. We show that supervised fine-tuning (SFT) with high-quality data can rival proprietary approaches. To this end, we compile a 161.4 million token multimodal dataset combining textbook question-solution pairs, curriculum aligned diagrams, and contextual materials, and fine-tune Qwen-2.5VL-32B using an optimized reasoning syntax (QMSA). The resulting model achieves 78.6% accuracy, only 1.0% below Gemini 2.0 Flash, on our newly released benchmark YKSUniform, which standardizes 1,854 multimodal exam questions across 309 curriculum topics. Our results reveal that data composition and representational syntax play a decisive role in multimodal reasoning. This work establishes a data centric framework for advancing open weight vision language models, demonstrating that carefully curated and curriculum-grounded multimodal data can elevate supervised fine-tuning to near state-of-the-art performance.

Understanding the nuances in everyday language is pivotal for advancements in computational linguistics & emotions research. Traditional lexicon-based tools such as LIWC and Pattern have long served as foundational instruments in this domain. LIWC is the most extensively validated word count based text analysis tool in the social sciences and Pattern is an open source Python library offering functionalities for NLP. However, everyday language is inherently spontaneous, richly expressive, & deeply context dependent. To explore the capabilities of LLMs in capturing the valences of daily narratives in Flemish, we first conducted a study involving approximately 25,000 textual responses from 102 Dutch-speaking participants. Each participant provided narratives prompted by the question, "What is happening right now and how do you feel about it?", accompanied by self-assessed valence ratings on a continuous scale from -50 to +50. We then assessed the performance of three Dutch-specific LLMs in predicting these valence scores, and compared their outputs to those generated by LIWC and Pattern. Our findings indicate that, despite advancements in LLM architectures, these Dutch tuned models currently fall short in accurately capturing the emotional valence present in spontaneous, real-world narratives. This study underscores the imperative for developing culturally and linguistically tailored models/tools that can adeptly handle the complexities of natural language use. Enhancing automated valence analysis is not only pivotal for advancing computational methodologies but also holds significant promise for psychological research with ecologically valid insights into human daily experiences. We advocate for increased efforts in creating comprehensive datasets & finetuning LLMs for low-resource languages like Flemish, aiming to bridge the gap between computational linguistics & emotion research.

Subgrid machine-learning (ML) parameterizations have the potential to introduce a new generation of climate models that incorporate the effects of higher-resolution physics without incurring the prohibitive computational cost associated with more explicit physics-based simulations. However, important issues, ranging from online instability to inconsistent online performance, have limited their operational use for long-term climate projections. To more rapidly drive progress in solving these issues, domain scientists and machine learning researchers opened up the offline aspect of this problem to the broader machine learning and data science community with the release of ClimSim, a NeurIPS Datasets and Benchmarks publication, and an associated Kaggle competition. This paper reports on the downstream results of the Kaggle competition by coupling emulators inspired by the winning teams' architectures to an interactive climate model (including full cloud microphysics, a regime historically prone to online instability) and systematically evaluating their online performance. Our results demonstrate that online stability in the low-resolution, real-geography setting is reproducible across multiple diverse architectures, which we consider a key milestone. All tested architectures exhibit strikingly similar offline and online biases, though their responses to architecture-agnostic design choices (e.g., expanding the list of input variables) can differ significantly. Multiple Kaggle-inspired architectures achieve state-of-the-art (SOTA) results on certain metrics such as zonal mean bias patterns and global RMSE, indicating that crowdsourcing the essence of the offline problem is one path to improving online performance in hybrid physics-AI climate simulation.

Grammatical error correction for Indic languages faces limited supervision, diverse scripts, and rich morphology. We propose an augmentation-free setup that uses instruction-tuned large language models and conservative decoding. A 12B GEMMA 3 model is instruction-tuned in bnb 4-bit precision with parameter-efficient fine-tuning (PEFT) and Alpaca-style formatting. Decoding follows a deterministic, constraint-aware procedure with a lightweight normaliser that encourages minimal, meaning-preserving edits. We operationalise inference, subsequent to instruction fine-tuning (IFT), via a fixed, language-specific prompt directly synthesised from a deterministic error classifier's taxonomy, label distributions, and precedence ordering computed on the training corpus. Under the official untuned GLEU evaluation, the system scores 92.41 on Malayalam, sixth overall, and 81.44 on Hindi, third overall. These results indicate that classifier-informed prompt design, adapter-based instruction tuning, and deterministic decoding provide a reproducible and a computationally efficient alternative to augmentation-centred pipelines for Indic GEC. The approach also motivates future work on stronger morphosyntactic constraints and human-centred evaluation of conservative edits.

Fairness in clinical prediction models remains a persistent challenge, particularly in high-stakes applications such as spinal fusion surgery for scoliosis, where patient outcomes exhibit substantial heterogeneity. Many existing fairness approaches rely on coarse demographic adjustments or post-hoc corrections, which fail to capture the latent structure of clinical populations and may unintentionally reinforce bias. We propose FAIR-MTL, a fairness-aware multitask learning framework designed to provide equitable and fine-grained prediction of postoperative complication severity. Instead of relying on explicit sensitive attributes during model training, FAIR-MTL employs a data-driven subgroup inference mechanism. We extract a compact demographic embedding, and apply k-means clustering to uncover latent patient subgroups that may be differentially affected by traditional models. These inferred subgroup labels determine task routing within a shared multitask architecture. During training, subgroup imbalance is mitigated through inverse-frequency weighting, and regularization prevents overfitting to smaller groups. Applied to postoperative complication prediction with four severity levels, FAIR-MTL achieves an AUC of 0.86 and an accuracy of 75%, outperforming single-task baselines while substantially reducing bias. For gender, the demographic parity difference decreases to 0.055 and equalized odds to 0.094; for age, these values reduce to 0.056 and 0.148, respectively. Model interpretability is ensured through SHAP and Gini importance analyses, which consistently highlight clinically meaningful predictors such as hemoglobin, hematocrit, and patient weight. Our findings show that incorporating unsupervised subgroup discovery into a multitask framework enables more equitable, interpretable, and clinically actionable predictions for surgical risk stratification.

Modern recommender systems struggle to effectively utilize the rich, yet high-dimensional and noisy, multi-modal features generated by Large Language Models (LLMs). Treating these features as static inputs decouples them from the core recommendation task. We address this limitation with a novel framework built on a key insight: deeply fusing multi-modal and collaborative knowledge for representation denoising. Our unified architecture introduces two primary technical innovations. First, we integrate dimensionality reduction directly into the recommendation model, enabling end-to-end co-training that makes the reduction process aware of the final ranking objective. Second, we introduce a contrastive learning objective that explicitly incorporates the collaborative filtering signal into the latent space. This synergistic process refines raw LLM embeddings, filtering noise while amplifying task-relevant signals. Extensive experiments confirm our method's superior discriminative power, proving that this integrated fusion and denoising strategy is critical for achieving state-of-the-art performance. Our work provides a foundational paradigm for effectively harnessing LLMs in recommender systems.

Objectives: We present and evaluate a Mamba-based deep-learning model for diagnosis and event-level characterization of sleep disordered breathing based on signals from the ANNE One, a non-intrusive dual-module wireless wearable system measuring chest electrocardiography, triaxial accelerometry, chest and finger temperature, and finger phototplethysmography. Methods: We obtained concurrent PSG and wearable sensor recordings from 384 adults attending a tertiary care sleep laboratory. Respiratory events in the PSG were manually annotated in accordance with AASM guidelines. Wearable sensor and PSG recordings were automatically aligned based on the ECG signal, alignment confirmed by visual inspection, and PSG-derived respiratory event labels were used to train and evaluate a deep sequential neural network based on the Mamba architecture. Results: In 57 recordings in our test set (mean age 56, mean AHI 10.8, 43.86\% female) the model-predicted AHI was highly correlated with that derived form the PSG labels (R=0.95, p=8.3e-30, men absolute error 2.83). This performance did not vary with age or sex. At a threshold of AHI$>$5, the model had a sensitivity of 0.96, specificity of 0.87, and kappa of 0.82, and at a threshold of AHI$>$15, the model had a sensitivity of 0.86, specificity of 0.98, and kappa of 0.85. At the level of 30-sec epochs, the model had a sensitivity of 0.93 and specificity of 0.95, with a kappa of 0.68 regarding whether any given epoch contained a respiratory event. Conclusions: Applied to data from the ANNE One, a Mamba-based deep learning model can accurately predict AHI and identify SDB at clinically relevant thresholds, achieves good epoch- and event-level identification of individual respiratory events, and shows promise at physiological characterization of these events including event type (central vs. other) and event duration.

Echocardiogram datasets enable training deep learning models to automate interpretation of cardiac ultrasound, thereby expanding access to accurate readings of diagnostically-useful images. However, the gender, sex, race, and ethnicity of the patients in these datasets are underreported and subgroup-specific predictive performance is unevaluated. These reporting deficiencies raise concerns about subgroup validity that must be studied and addressed before model deployment. In this paper, we show that current open echocardiogram datasets are unable to assuage subgroup validity concerns. We improve sociodemographic reporting for two datasets: TMED-2 and MIMIC-IV-ECHO. Analysis of six open datasets reveals no consideration of gender-diverse patients and insufficient patient counts for many racial and ethnic groups. We further perform an exploratory subgroup analysis of two published aortic stenosis detection models on TMED-2. We find insufficient evidence for subgroup validity for sex, racial, and ethnic subgroups. Our findings highlight that more data for underrepresented subgroups, improved demographic reporting, and subgroup-focused analyses are needed to prove subgroup validity in future work.

Recent efforts on Diffusion Mixture-of-Experts (MoE) models have primarily focused on developing more sophisticated routing mechanisms. However, we observe that the underlying architectural configuration space remains markedly under-explored. Inspired by the MoE design paradigms established in large language models (LLMs), we identify a set of crucial architectural factors for building effective Diffusion MoE models--including DeepSeek-style expert modules, alternative intermediate widths, varying expert counts, and enhanced attention positional encodings. Our systematic study reveals that carefully tuning these configurations is essential for unlocking the full potential of Diffusion MoE models, often yielding gains that exceed those achieved by routing innovations alone. Through extensive experiments, we present novel architectures that can be efficiently applied to both latent and pixel-space diffusion frameworks, which provide a practical and efficient training recipe that enables Diffusion MoE models to surpass strong baselines while using equal or fewer activated parameters. All code and models are publicly available at: https://github.com/yhlleo/EfficientMoE.

Generative diversity varies significantly across discrete latent generative models such as AR, MIM, and Diffusion. We propose a diagnostic framework, grounded in Information Bottleneck (IB) theory, to analyze the underlying strategies resolving this behavior. The framework models generation as a conflict between a 'Compression Pressure' - a drive to minimize overall codebook entropy - and a 'Diversity Pressure' - a drive to maximize conditional entropy given an input. We further decompose this diversity into two primary sources: 'Path Diversity', representing the choice of high-level generative strategies, and 'Execution Diversity', the randomness in executing a chosen strategy. To make this decomposition operational, we introduce three zero-shot, inference-time interventions that directly perturb the latent generative process and reveal how models allocate and express diversity. Application of this probe-based framework to representative AR, MIM, and Diffusion systems reveals three distinct strategies: "Diversity-Prioritized" (MIM), "Compression-Prioritized" (AR), and "Decoupled" (Diffusion). Our analysis provides a principled explanation for their behavioral differences and informs a novel inference-time diversity enhancement technique.