In early May, Sam Altman traveled to Washington to tell a story about America. Appearing before a Senate committee, Altman described how he came of age as the internet took off, how he stayed up late in his family's attic and learned to code on products that were invented in the United States--a personal computer, its silicon chips and accompanying software. That early experience with the "spirit of American innovation," Altman told the senators, put him on a path to found OpenAI, launch ChatGPT, and set off the AI boom. "I think America is just an incredible and special thing," he said, "and it will not only be the place where the AI revolution happens but all the revolutions after." Altman's written testimony, which was submitted to the Senate, added an important asterisk that he did not speak aloud that day.

We present Hydra as an architectural proposal for hybrid long-context language models that combine conditional computation, long-context memory mechanisms, and sparse mixture-of-experts within an approximately 1.6B parameter design envelope. Hydra integrates a Mamba-style Structured State Space Model (SSM) backbone with intermittent sparse global attention, chunk-level MoE feed-forward routing, and dual (workspace plus factual PKM) memories. We formalize the component interfaces, give transparent parameter and complexity accounting, and outline a staged curriculum intended to stably activate the parts. We accompany the specification with illustrative toy-scale prototype measurements (tens of millions of parameters on synthetic data) whose sole purpose is to demonstrate implementation feasibility and qualitative scaling behaviors (for example, long-context throughput crossover and controllable expert routing), not to claim competitive full-scale performance. We explicitly delineate assumptions and open risks (training complexity, memory utilization, specialization dynamics) and position Hydra as a blueprint to stimulate empirical follow-up rather than a finished system. By combining SSM efficiency, selective sparse attention, MoE capacity, and learnable memory, Hydra sketches a path toward modular, input-adaptive long-context language models; validating end-task gains at target scale remains future work.

Transformers have emerged as a powerful class of sequence models with remarkable expressive capabilities. Originally popularized in the context of natural language processing, they leverage self-attention mechanisms to in-context learn new tasks without any parameter updates (Vaswani, 2017; Liu et al., 2021; Dosovitskiy, 2020; Yun et al., 2019; Dong et al., 2018). In other words, a large transformer model can be given a prompt consisting of example input-output pairs for an unseen task and subsequently produce correct outputs for new queries of that task, purely by processing the sequence of examples and queries (Lee et al., 2022; Laskin et al., 2022; Yang et al., 2023; Lin et al., 2024). This ability to dynamically adapt via context rather than gradient-based fine-tuning has spurred extensive interest in understanding the theoretical expressivity of transformers and how they might learn algorithms internally during training. Recent theoretical work has begun to analyze the various aspects of transformers.

The affective attitude of liking a recommended item reflects just one category in a wide spectrum of affective phenomena that also includes emotions such as entranced or intrigued, moods such as cheerful or buoyant, as well as more fine-grained affective states, such as "pleasantly surprised by the conclusion". In this paper, we introduce a novel recommendation task that can leverage a virtually unbounded range of affective states sought explicitly by the user in order to identify items that, upon consumption, are likely to induce those affective states. Correspondingly, we create a large dataset of user preferences containing expressions of fine-grained affective states that are mined from book reviews, and propose a Transformer-based architecture that leverages such affective expressions as input. We then use the resulting dataset of affective states preferences, together with the linked users and their histories of book readings, ratings, and reviews, to train and evaluate multiple recommendation models on the task of matching recommended items with affective preferences. Experiments show that the best results are obtained by models that can utilize textual descriptions of items and user affective preferences.

Full-text screening is the major bottleneck of systematic reviews (SRs), as decisive evidence is dispersed across long, heterogeneous documents and rarely admits static, binary rules. We present a scalable, auditable pipeline that reframes inclusion/exclusion as a fuzzy decision problem and benchmark it against statistical and crisp baselines in the context of the Population Health Modelling Consensus Reporting Network for noncommunicable diseases (POPCORN). Articles are parsed into overlapping chunks and embedded with a domain-adapted model; for each criterion (Population, Intervention, Outcome, Study Approach), we compute contrastive similarity (inclusion-exclusion cosine) and a vagueness margin, which a Mamdani fuzzy controller maps into graded inclusion degrees with dynamic thresholds in a multi-label setting. A large language model (LLM) judge adjudicates highlighted spans with tertiary labels, confidence scores, and criterion-referenced rationales; when evidence is insufficient, fuzzy membership is attenuated rather than excluded. In a pilot on an all-positive gold set (16 full texts; 3,208 chunks), the fuzzy system achieved recall of 81.3% (Population), 87.5% (Intervention), 87.5% (Outcome), and 75.0% (Study Approach), surpassing statistical (56.3-75.0%) and crisp baselines (43.8-81.3%). Strict "all-criteria" inclusion was reached for 50.0% of articles, compared to 25.0% and 12.5% under the baselines. Cross-model agreement on justifications was 98.3%, human-machine agreement 96.1%, and a pilot review showed 91% inter-rater agreement (kappa = 0.82), with screening time reduced from about 20 minutes to under 1 minute per article at significantly lower cost. These results show that fuzzy logic with contrastive highlighting and LLM adjudication yields high recall, stable rationale, and end-to-end traceability.

The non-invasive assessment of increasingly incidentally discovered renal masses is a critical challenge in urologic oncology, where diagnostic uncertainty frequently leads to the overtreatment of benign or indolent tumors. In this study, we developed and validated RenalCLIP using a dataset of 27,866 CT scans from 8,809 patients across nine Chinese medical centers and the public TCIA cohort, a visual-language foundation model for characterization, diagnosis and prognosis of renal mass. The model was developed via a two-stage pre-training strategy that first enhances the image and text encoders with domain-specific knowledge before aligning them through a contrastive learning objective, to create robust representations for superior generalization and diagnostic precision. RenalCLIP achieved better performance and superior generalizability across 10 core tasks spanning the full clinical workflow of kidney cancer, including anatomical assessment, diagnostic classification, and survival prediction, compared with other state-of-the-art general-purpose CT foundation models. Especially, for complicated task like recurrence-free survival prediction in the TCIA cohort, RenalCLIP achieved a C-index of 0.726, representing a substantial improvement of approximately 20% over the leading baselines. Furthermore, RenalCLIP's pre-training imparted remarkable data efficiency; in the diagnostic classification task, it only needs 20% training data to achieve the peak performance of all baseline models even after they were fully fine-tuned on 100% of the data. Additionally, it achieved superior performance in report generation, image-text retrieval and zero-shot diagnosis tasks. Our findings establish that RenalCLIP provides a robust tool with the potential to enhance diagnostic accuracy, refine prognostic stratification, and personalize the management of patients with kidney cancer.

Training large language models (LLMs) from scratch is increasingly impractical, making post-training methods such as supervised fine-tuning (SFT) and reinforcement-learning fine-tuning (RL-FT, e.g., PPO) central to modern practice. Using an out-of-distribution (OOD) variant of the 24-point card game and new spectrum-based diagnostics, we revisit how these two stages reshape model representation and OOD performance. Our key findings are- (1) RL-FT can restore much of the OOD performance loss from SFT (e.g., Llama-11B 8.97% to 15.38%, Qwen-7B 17.09% to 19.66%). But when SFT induces severe overfitting and a clear distribution shift, RL-FT cannot fully recover OOD performance. (2) Direction shifts of singular vectors matter more than singular value magnitudes. These shifts concentrate on directions linked to the largest and smallest singular values, leaving the bulk spectrum intact. (3) Low-rank and shallow recovery is effective: restoring singular vector directions for the top 20% of values or first 25% of layers recovers 70-80% of OOD performance. (4) Stronger SFT checkpoints enable better recovery by RL, while overfitted ones resist restoration. These results reconcile prior reports of RL superior OOD performance: RL primarily counteracts SFT-induced directional drift rather than finding new solutions. Our spectrum-aware analysis highlights inexpensive recovery knobs low-rank UV merging and shallow-layer resets that practitioners can use before costly RL fine-tuning.

For decades, Computer Vision has aimed at enabling machines to perceive the external world. Initial limitations led to the development of highly specialized niches. As success in each task accrued and research progressed, increasingly complex perception tasks emerged. This survey charts the convergence of these tasks and, in doing so, introduces Open World Detection (OWD), an umbrella term we propose to unify class-agnostic and generally applicable detection models in the vision domain. We start from the history of foundational vision subdomains and cover key concepts, methodologies and datasets making up today's state-of-the-art landscape. This traverses topics starting from early saliency detection, foreground/background separation, out of distribution detection and leading up to open world object detection, zero-shot detection and Vision Large Language Models (VLLMs). We explore the overlap between these subdomains, their increasing convergence, and their potential to unify into a singular domain in the future, perception.

Recent works improving LLM math reasoning with synthetic data have used unique setups, making comparison of data synthesis strategies impractical. This leaves many unanswered questions about the roles of different factors in the synthetic data pipeline, such as the impact of filtering low-quality problems. To address this gap, we introduce FLAMES, a Framework for LLM Assessment of Math rEasoning Data Synthesis, and perform a systematic study of 10 existing data synthesis strategies and multiple other factors impacting the performance of synthetic math reasoning data. Our FLAMES experiments provide several valuable insights about the optimal balance of difficulty and diversity of synthetic data. First, data agents designed to increase problem complexity lead to best improvements on most math metrics. Second, with a fixed data generation budget, keeping higher problem coverage is more important than keeping only problems with reliable solutions. Third, GSM8K- and MATH-based synthetic data can lead to improvements on competition-level benchmarks, showcasing easy-to-hard generalization. Leveraging insights from our FLAMES experiments, we design two novel data synthesis strategies for improving out-of-domain generalization and robustness. Further, we develop the FLAMES dataset, an effective blend of our novel and existing data synthesis strategies, outperforming public datasets on OlympiadBench (+15.7), CollegeMath (+4.5), GSMPlus (+6.5), and MATH (+3.1). Fine-tuning Qwen2.5-Math-7B on the FLAMES dataset achieves 81.4% on MATH, surpassing larger Llama3 405B, GPT-4o and Claude 3.5 Sonnet.

In the digital era, individuals are increasingly exposed to online harms such as toxicity, manipulation, and grooming, which often pose emotional and safety risks. Existing systems for detecting abusive content or issuing safety alerts operate in isolation and rarely combine digital safety with emotional well-being. In this paper, we present SafeSpace, a unified web application that integrates three modules: (1) toxicity detection in chats and screenshots using NLP models and Google's Perspective API, (2) a configurable safety ping system that issues emergency alerts with the user's live location (longitude and latitude) via SMTP-based emails when check-ins are missed or SOS alerts are manually triggered, and (3) a reflective questionnaire that evaluates relationship health and emotional resilience. The system employs Firebase for alert management and a modular architecture designed for usability, privacy, and scalability. The experimental evaluation shows 93% precision in toxicity detection, 100% reliability in safety alerts under emulator tests, and 92% alignment between automated and manual questionnaire scoring. SafeSpace, implemented as a web application, demonstrates the feasibility of integrating detection, protection, and reflection within a single platform, with future deployment envisioned as a mobile application for broader accessibility.