How do language models use contextual information to answer health questions? How are their responses impacted by conflicting contexts? We assess the ability of language models to reason over long, conflicting biomedical contexts using HealthContradict, an expert-verified dataset comprising 920 unique instances, each consisting of a health-related question, a factual answer supported by scientific evidence, and two documents presenting contradictory stances. We consider several prompt settings, including correct, incorrect or contradictory context, and measure their impact on model outputs. Compared to existing medical question-answering evaluation benchmarks, HealthContradict provides greater distinctions of language models' contextual reasoning capabilities. Our experiments show that the strength of fine-tuned biomedical language models lies not only in their parametric knowledge from pretraining, but also in their ability to exploit correct context while resisting incorrect context.

Multimodal large language models (MLLMs) are expected to jointly interpret vision, audio, and language, yet existing video benchmarks rarely assess fine-grained reasoning about human speech. Many tasks remain visually solvable or only coarsely evaluate speech, offering limited insight into whether models can align who speaks, what is said, and when it occurs. We introduce AV-SpeakerBench, a curated benchmark of 3,212 multiple-choice questions focused on speaker-centric audiovisual reasoning in real-world videos. It features: (1) a speaker-centered formulation that treats speakers-not scenes-as the core reasoning unit; (2) fusion-grounded question design embedding audiovisual dependencies into question semantics; and (3) expert-curated annotations ensuring temporal precision and cross-modal validity. Comprehensive evaluations show that the Gemini family consistently outperforms open-source systems, with Gemini 2.5 Pro achieving the best results. Among open models, Qwen3-Omni-30B approaches Gemini 2.0 Flash but remains far behind Gemini 2.5 Pro, primarily due to weaker audiovisual fusion rather than visual perception. We believe AV-SpeakerBench establishes a rigorous foundation for advancing fine-grained audiovisual reasoning in future multimodal systems.

Effective text generation and chat interfaces for low-resource languages (LRLs) remain a challenge for state-of-the-art large language models (LLMs) to support. This is mainly due to the difficulty of curating high-quality instruction datasets for LRLs, a limitation prevalent in the languages spoken across the African continent and other regions. Current approaches, such as automated translation and synthetic data generation, frequently yield outputs that lack fluency or even orthographic consistency. In this paper, we introduce InstructLR, a novel framework designed to generate high-quality instruction datasets for LRLs. Our approach integrates LLM-driven text generation with a dual-layer quality filtering mechanism: an automated filtering layer based on retrieval-augmented-generation (RAG)-based n-shot prompting, and a human-in-the-loop validation layer. Drawing inspiration from benchmarks such as MMLU in task definition, InstructLR has facilitated the creation of three multi-domain instruction benchmarks: ZarmaInstruct-50k, BambaraInstruct-50k, and FulfuldeInstruct-50k.

Sperm whales communicate in short sequences of clicks known as codas. We present WhAM (Whale Acoustics Model), the first transformer-based model capable of generating synthetic sperm whale codas from any audio prompt. WhAM is built by finetuning VampNet, a masked acoustic token model pretrained on musical audio, using 10k coda recordings collected over the past two decades. Through iterative masked token prediction, WhAM generates high-fidelity synthetic codas that preserve key acoustic features of the source recordings. We evaluate WhAM's synthetic codas using Fréchet Audio Distance and through perceptual studies with expert marine biologists. On downstream classification tasks including rhythm, social unit, and vowel classification, WhAM's learned representations achieve strong performance, despite being trained for generation rather than classification. Our code is available at https://github.com/Project-CETI/wham

We introduce Bin2Vec, a new framework that helps compare software programs in a clear and explainable way. Instead of focusing only on one type of information, Bin2Vec combines what a program looks like (its built-in functions, imports, and exports) with how it behaves when it runs (its instructions and memory usage). This gives a more complete picture when deciding whether two programs are similar or not. Bin2Vec represents these different types of information as views that can be inspected separately using easy-to-read charts, and then brings them together into an overall similarity score. Bin2Vec acts as a bridge between binary representations and machine learning techniques by generating feature representations that can be efficiently processed by machine-learning models. We tested Bin2Vec on multiple versions of two well-known Windows programs, PuTTY and 7-Zip. The primary results strongly confirmed that our method compute an optimal and visualization-friendly representation of the analyzed software. For example, PuTTY versions showed more complex behavior and memory activity, while 7-Zip versions focused more on performance-related patterns. Overall, Bin2Vec provides decisions that are both reliable and explainable to humans. Because it is modular and easy to extend, it can be applied to tasks like auditing, verifying software origins, or quickly screening large numbers of programs in cybersecurity and reverse-engineering work.

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.

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.

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].

Educators frequently rely on diagrams to explain complex concepts during lectures, yet creating clear and complete visual representations in real time while simultaneously speaking can be cognitively demanding. Incomplete or unclear diagrams may hinder student comprehension, as learners must mentally reconstruct missing information while following the verbal explanation. Inspired by advances in code completion tools, we introduce DrawDash, an AI-powered white-board assistant that proactively completes and refines educational diagrams through multimodal understanding. Draw-Dash adopts a T AB-completion interaction model: it listens to spoken explanations, detects intent, and dynamically suggests refinements that can be accepted with a single keystroke. We demonstrate DrawDash across four diverse teaching scenarios--spanning topics from computer science and web development to biology. This work represents an early exploration into reducing instructors' cognitive load and improving diagram-based pedagogy through real-time, speech-driven visual assistance, and concludes with a discussion of current limitations and directions for formal classroom evaluation.