This paper introduces ToolOrchestra, a method for training small AI models (orchestrators) to efficiently coordinate other, often more powerful, AI models and tools. The Orchestrator, an 8B parameter model, learns through reinforcement learning to balance task outcome, efficiency, and user preferences, achieving higher accuracy at significantly lower cost on complex benchmarks like Humanity's Last Exam (HLE) compared to larger, monolithic models. The study's evaluations rely on computational benchmarks and synthetic data, which may not fully capture real-world complexities.

This preclinical mouse study found that dietary restriction (DR) significantly slows tumor growth by reprogramming CD8+ T cells in the tumor microenvironment. DR promotes the expansion of highly functional, cytotoxic T cells and limits exhausted T cells, primarily by increasing ketone body oxidation. Furthermore, DR synergizes with anti-PD1 immunotherapy, enhancing its anti-tumor effects in these mouse models.

This comprehensive review synthesizes existing research, predominantly from animal models, demonstrating that regular physical activity can mitigate all 14 hallmarks of biological aging. The findings suggest exercise is a promising non-pharmacological strategy to promote healthy aging, but emphasize the critical need for more human studies to confirm and extend these benefits.

This paper presents six case reports (three involving MSG and three involving lupin seed) suggesting a link between excessive consumption of certain food items and cramp-fasciculation syndrome (CFS) along with other neurological symptoms. While the small, observational nature of these cases precludes strong causal claims, the authors hypothesize that high intake of glutamate or quinolizidine alkaloids may contribute to these conditions. The findings highlight potential health concerns related to popular foods and call for further research into their neurological safety.

This paper introduces MAKER, a framework leveraging Massively Decomposed Agentic Processes (MDAPs) to enable large language models (LLMs) to reliably solve million-step tasks with zero errors. By breaking down complex problems into minimal subtasks, implementing subtask-level voting for error correction, and red-flagging unreliable outputs, MAKER successfully completed a 20-disk Towers of Hanoi puzzle (over 1 million steps). The research suggests that extreme decomposition combined with robust error correction offers a scalable paradigm for long-horizon AI tasks, rather than relying solely on continually improving base LLMs.

This paper introduces WorldTest and AutumnBench to evaluate AI world models, revealing that humans significantly outperform current frontier AI models (Claude, Gemini, OpenAI 03) in learning grid-world environment dynamics. Human success is attributed to more effective exploration strategies, such as frequent use of "resets" to test hypotheses and more flexible belief updating. The findings highlight substantial shortcomings in AI's current world-modeling capabilities, particularly in experimental design and adaptive learning.

The paper presents Scholar Inbox, a new open-access platform providing personalized paper recommendations and research tools for scientists. Utilizing a content-based model trained on a large dataset of explicit user ratings, the platform aims to combat information overload and improve research efficiency. A comprehensive evaluation, including a large user study, demonstrates high user satisfaction and retention, though the system's ability to explicitly model multiple diverse research interests is still an area for future improvement.

This paper introduces Kosmos, an AI system designed to automate scientific discovery by performing iterative cycles of data analysis, literature search, and hypothesis generation. While Kosmos demonstrates impressive scale in executing complex tasks and reading numerous papers, it struggles with accurate interpretation of results (only 57% accurate) and is prone to generating conceptually obscure metrics, which significantly limits its reliability for truly autonomous discovery.

This study reveals widespread age- and gender-related distortion across online media (images, videos, text) and large language models (LLMs). It found that women are consistently depicted as younger than men, especially in higher-status roles, and demonstrated how mainstream algorithms amplify these biases, influencing people's beliefs and even AI-driven hiring recommendations. The research combined large-scale data analysis with human experiments and LLM audits to show these biases are pervasive and disconnect from real-world demographics.

This paper introduces SPICE, a novel reinforcement learning framework where a single large language model (LLM) trains itself by generating challenging reasoning tasks from a vast document corpus and then solving them. By interacting with external, verifiable information, SPICE successfully overcomes common issues like hallucination and performance plateaus seen in ungrounded self-play, leading to significant improvements in both mathematical and general reasoning abilities across various LLMs.

This theoretical paper argues that a complete, purely algorithmic "Theory of Everything" (ToE) is impossible due to fundamental mathematical limitations like Gödel's incompleteness theorems. It proposes a "Meta-Theory of Everything" (MToE) that incorporates non-algorithmic understanding, suggesting some aspects of reality are inherently uncomputable. This framework further implies that the universe cannot be a computer simulation, as any simulation would be algorithmic and thus incomplete.

This paper proposes a practical three-pass method for efficiently reading research papers, designed to save time and improve comprehension for researchers. The method systematically moves from a general overview to a deep understanding, building knowledge with each pass, and can also be applied effectively to conduct literature surveys. It aims to address the common challenge faced by graduate students who struggle with efficient paper consumption.

This study found that two types of highly drug-resistant bacteria (ST78 and ST133) are widespread in Japanese healthcare settings, including many small hospitals, suggesting a silent and ongoing clonal dissemination. The research used whole-genome sequencing of isolates collected across 58 hospitals, identifying ST133 as a particularly common and underreported lineage in Japan, emphasizing the need for broader molecular surveillance.

This study used causal inference to investigate how well CMIP6 climate models represent the mechanisms of subpolar gyre variability and potential tipping points in the North Atlantic. While basic links between surface conditions and deep convection are generally captured, key feedback loops involving subsurface temperature, ocean density, and gyre strength often show inconsistent or even contradictory signs across models, or are completely absent. Crucially, models that do capture more of the proposed mechanisms are also the ones that predict abrupt shifts in the subpolar gyre.

This paper introduces WorldMirror, a novel AI model that can reconstruct 3D scenes from images and various "hints" like camera data or depth maps, generating multiple 3D representations simultaneously. It achieves state-of-the-art performance across diverse 3D reconstruction tasks by flexibly integrating these priors, although it shows suboptimal performance on dynamic scenes due to training data limitations. The model demonstrates strong generalization and efficiency, showcasing a promising direction for universal 3D scene understanding.

This paper introduces a transformer-based 3D embedding that efficiently approximates global illumination in computer-generated scenes, allowing for generalizable and view-independent rendering without traditional ray tracing. The model learns to encode scene geometry, materials, and lighting into latent codes, which can then be decoded to predict indirect lighting effects, with preliminary results showing potential for complex tasks like glossy reflections and path guiding.

This review highlights a critical, understudied area: the interactive effects of chemical pollution and biological invasions on animal populations. It synthesizes existing knowledge, showing how pollutants can influence various traits (e.g., reproduction, behavior, thermal tolerance) that are crucial at different stages of an invasion, from uptake and transport to establishment and spread. The paper emphasizes that despite the individual impacts of pollution and invasions being well-documented, their combined effects are largely unexplored and represent a significant gap in ecological research.

This report from the ANZDATA Registry found that deceased kidney donor numbers declined in Australia and plateaued in New Zealand during 2019-2020, likely influenced by the COVID-19 pandemic. Live kidney donation remained stable, with New Zealand maintaining a rate more than double that of Australia, accounting for 47% versus 20% of transplants, respectively.

This paper introduces Pico-Banana-400K, a large-scale dataset of approximately 400,000 text-guided image edits, which is primarily generated and quality-controlled by AI models rather than humans. The dataset leverages Nano-Banana for diverse edit generation from real images and Gemini-2.5-Pro for automated quality assessment, providing examples for single-turn, multi-turn, and preference-based editing scenarios. It aims to establish a robust foundation for training and benchmarking the next generation of text-guided image editing models, despite inherent biases from its AI-on-AI generation and judging process.

This study investigated how neighborhood disadvantage, social vulnerability, and environmental injustice (place-based social determinants of health, SDoH) relate to Alzheimer's disease biomarkers. It found that SDoH measures were significantly higher among Black participants compared to White participants. Crucially, these higher SDoH measures were associated with worse neuroimaging biomarkers (like lower cortical thickness and higher cerebral blood flow variability) exclusively among Black participants, not White participants.

This study compared greenhouse gas emissions from a conventional pasture-based beef system (CONV) and a partial-confinement system (ALT). The CONV system was found to be a net carbon sink due to perennial grassland sequestration, while the ALT system was a net source of emissions. This highlights that the design of beef production systems significantly influences their environmental footprint.

This large prospective cohort study of over one million Swedish male adolescents found that low muscular strength in adolescence is an independent risk factor for premature death from any cause, cardiovascular disease, and suicide. Weaker adolescents had significantly higher mortality rates, with the lowest tenth showing the highest risk, an effect size comparable to elevated BMI or blood pressure.

This paper introduces "sparse memory finetuning," a novel method for Large Language Models (LLMs) to learn new information without catastrophically forgetting previously acquired knowledge. By selectively updating only the most relevant memory slots using a TF-IDF-like ranking, the method significantly reduces interference between new and existing knowledge. Evaluated on two question answering tasks, sparse memory finetuning demonstrated substantially less forgetting (e.g., an 11% drop in F1 score vs. 89% for full finetuning) while effectively acquiring new knowledge.

This paper introduces "mementum," a new concept describing periods where social media coordination significantly drives stock prices and trading volumes. Using a regime-switching cointegration model on Twitter data, the authors successfully identify these "meme periods" in well-known meme stocks like GameStop and AMC, distinguishing them from other periods of high social media activity in non-meme stocks. The study is limited to Twitter data for its empirical application, though the methodology is presented as platform-agnostic.

This large cross-cultural study involving over 6,000 participants across four countries found that support for economic redistribution is primarily predicted by an individual's dispositional compassion, dispositional envy, and expected personal gain. Conversely, a "taste for fairness," whether defined as uniformity in laws or low variance in outcomes, did not reliably predict attitudes towards redistribution. The findings suggest that evolved psychological mechanisms related to interpersonal interactions are key drivers, rather than abstract notions of societal fairness.

This study found that observing large groups of people moving synchronously reduced physiological arousal (measured by skin conductance) in undergraduate participants, an effect not seen when observing small groups or asynchronous movement. The calming effect for large synchronous groups was partially mediated by viewer enjoyment and interest in the videos.

This mouse study suggests that male SARS-CoV-2 infection can lead to increased anxiety-like behaviors in first-generation offspring and subtle changes in grand-offspring, mediated by alterations in sperm small noncoding RNAs. Microinjecting these altered sperm RNAs into oocytes partially replicated the anxiety phenotype. However, these findings are from an animal model and their direct relevance to humans requires further research.

The 2019 Ultra Low Emission Zone (ULEZ1) significantly reduced nitrogen dioxide (NO2) and nitrogen oxides (NOx) in central London and had positive spillover effects, though its impact on fine particulate matter (PM2.5) was limited. However, the city-wide expansion in 2023 (ULEZ3) showed no detectable impact on air pollution, and overall NO2 and PM2.5 levels remain well above World Health Organization guidelines, indicating ULEZ alone is insufficient to achieve truly clean air.

This editorial asserts that continuous use of phones and smart tablets causes a phenomenon dubbed "brain rot," leading to weakened perception, indifference, and various psychological issues, especially in children. It offers advice like consuming natural herbs, avoiding medications, and limiting screen time to prevent this purported condition, without providing any empirical data or research to support these claims.

This study, funded by Novavax and not yet peer-reviewed, suggests that Novavax's COVID-19 vaccines (monovalent and bivalent) significantly increased anti-SARS-CoV-2 IgA levels in the saliva of previously vaccinated adolescents. The researchers also observed cross-reactivity against other variants, highlighting the potential of mucosal IgA for vaccine development, though clinical significance is yet to be formally established due to study limitations.

This study found statistically significant associations between mysterious short-lived star-like objects captured in the Palomar Observatory Sky Survey (POSS-I) from 1949-1957, global above-ground nuclear weapons testing, and reports of Unidentified Anomalous Phenomena (UAP). Transients were 45% more likely on dates around nuclear tests, and an 8.5% increase in transient count was observed for every additional UAP reported, though these are correlations, not proven causal links. The findings are intriguing but the small magnitude of these associations and noisy data preclude definitive conclusions about their nature or origin.

This computational study introduces Lambda-ABF-OPES, a novel enhanced sampling method that accurately models water dynamics to improve protein-ligand binding affinity predictions. The approach, which shows good agreement with experimental data and significantly increases efficiency, could accelerate drug discovery. However, two authors are co-founders of Qubit Pharmaceuticals, which could directly benefit from this technology.

This modeling study suggests that explosive volcanic eruptions on early Mars could have delivered substantial amounts of water ice (up to ~5 meters) to equatorial regions, where ice is not expected to be stable under normal conditions. These volcanically induced ice deposits offer a compelling explanation for the elevated hydrogen content observed at low Martian latitudes, providing a mechanism for ice delivery independent of planetary obliquity changes. The model employs several simplifications regarding atmospheric processes and ash-water interactions, which may influence the exact distribution and persistence of the ice.

This paper presents a new mechanochemical method to break down Polytetrafluoroethylene (PTFE), commonly known as Teflon, using sodium metal and ball milling at room temperature, converting it into sodium fluoride and elemental carbon with 98% efficiency. The recovered fluoride can then be directly used to synthesize valuable sulfonyl fluoride and acyl fluoride fine chemicals through in situ mechanochemical fluorination reactions, offering a more sustainable way to recycle fluorine from PTFE waste.

This paper argues that Britain's justice system, contrary to its ideal of equality, is fundamentally biased towards the wealthy and powerful. It details how financial barriers, judicial culture, and lax regulation enable a system where access to justice depends on financial means, eroding public trust.

This mathematical modeling study, presented as a hypothetical analysis from mid-2025, projects that sharp cuts in international health aid from the United States and other donors to TB and HIV programs could lead to millions of additional pediatric tuberculosis cases and deaths between 2025 and 2034. The most severe scenario predicts an additional 8.9 million pediatric TB cases and 1.5 million deaths, particularly in Sub-Saharan Africa and South-East Asia, underscoring the critical importance of sustained funding.

This study, primarily using lab-grown mini-brains (organoids) and cultured human astrocytes, reveals that SARS-CoV-2 directly infects brain astrocytes using a novel receptor, Neuropilin-1 (NRP1), instead of the usual ACE2. This infection triggers severe inflammation in astrocytes, leading to a hostile brain environment that causes uninfected neighboring neurons to malfunction and die.

This paper introduces the Tiny Recursive Model (TRM), a simplified AI approach that uses a single small neural network (7M parameters) to recursively refine answers, significantly outperforming larger models like the Hierarchical Reasoning Model (HRM) and even some LLMs on tasks like Sudoku, Maze, and ARC-AGI. While achieving better generalization and requiring fewer computational resources, the model's optimal architecture and benefits are task-dependent, and the exact theoretical reason for recursion's effectiveness is not fully understood.

This paper introduces Physics-IQ, a comprehensive real-world benchmark to evaluate if generative video models truly understand physical principles like gravity or fluid dynamics. The study found that across a range of current models (e.g., Sora, VideoPoet), physical understanding is severely limited and largely unrelated to visual realism, despite some models generating highly realistic-looking videos. The research concludes that visual realism does not imply physical understanding, highlighting a significant gap in current AI capabilities.

This paper is a protocol outlining a pilot randomized controlled trial (INTERFAST-MI) to investigate the feasibility and safety of intermittent fasting in patients who have recently experienced a severe heart attack. The study aims to gather data on protocol adherence, recruitment rates, and preliminary effects on heart function (left ventricular ejection fraction) to inform the design and sample size calculation for a larger, definitive clinical trial. Key limitations, such as a small sample size and lack of blinding, are acknowledged as inherent to its pilot study design.

This study found that Australians with Postural Orthostatic Tachycardia Syndrome (POTS) experience significantly worse health-related quality of life across multiple domains compared to age- and sex-matched healthy individuals. Key predictors for this reduced quality of life in POTS patients included severity of orthostatic intolerance, female sex, fatigue scores, and a comorbid diagnosis of myalgic encephalomyelitis/chronic fatigue syndrome. While the findings highlight a substantial burden, the normative population data used for comparison was collected almost a decade prior to the POTS data.

This paper proposes "props," a new conceptual system for machine learning to securely access vast amounts of private "deep web" data while preserving user privacy and ensuring data integrity. It aims to solve the problem of limited high-quality training data and improve trustworthiness in ML models, outlining how such a system could be built using existing privacy-preserving oracle technologies without providing a full implementation or empirical validation.

This paper introduces GENERAL-REASONER, a novel training approach that significantly enhances large language models' (LLMs) reasoning capabilities across diverse domains beyond just math and coding. The method leverages a large, verifiable dataset curated from web crawling and a generative model-based verifier to provide robust reward signals for reinforcement learning. The results demonstrate superior generalizable reasoning performance compared to existing open-source baselines, while maintaining effectiveness in mathematical tasks.

This observational study compared health outcomes in a convenience sample of vaccinated and unvaccinated children from three pediatric practices. It found that vaccination before one year of age was associated with increased odds of developmental delays, asthma, and ear infections, but could only calculate unadjusted observational associations, not causal relationships, and did not control for many confounding factors. The sample selection and lack of demographic data severely limit the generalizability and reliability of these correlations.

This paper introduces Semantic Similarity Rating (SSR), a new method allowing large language models (LLMs) to accurately simulate human purchase intent by having them generate free-text responses, which are then mapped to Likert scales based on semantic similarity. The method, tested on 57 personal care product surveys, achieved 90% human test-retest reliability and produced realistic response distributions, outperforming direct numerical rating requests. It also generated rich qualitative feedback, though the reference statements were manually optimized for this dataset and not all demographics were replicated consistently.

This paper presents a novel method for detecting and controlling single electrons trapped on liquid helium at temperatures above 1 Kelvin, a significant step towards more practical quantum computing environments. Using a superconducting resonator, researchers observed frequency shifts corresponding to the loading and unloading of individual electrons, with these experimental results aligning well with their classical coupling model. This advancement provides a foundation for developing large-scale quantum processors that can operate with higher cooling powers than traditional millikelvin systems.

This paper introduces Self-Adapting Language Models (SEAL), a framework enabling Large Language Models (LLMs) to generate their own finetuning data and update instructions using reinforcement learning. This self-adaptation significantly improves performance in knowledge incorporation and few-shot learning tasks, often outperforming synthetic data generated by powerful models like GPT-4.1 for finetuning. However, the study acknowledges that SEAL is still susceptible to catastrophic forgetting, where new updates can interfere with previously learned knowledge.

This theoretical paper proposes a unifying framework linking quantum entanglement and classical chaos in many-particle systems. It finds that entanglement measures (entropy, Fisher information, square commutator) grow logarithmically/quadratically in regular systems but linearly/exponentially in chaotic ones, confirming previous conjectures. Numerical simulations of kicked top and Dicke models support these analytical predictions, though numerical precision limitations affect classical simulations in chaotic regimes over long times.

The paper introduces Agentic Context Engineering (ACE), a framework that allows large language models (LLMs) to self-improve by evolving their operational contexts into structured "playbooks." This approach, which avoids monolithic rewriting and instead uses incremental updates, consistently outperforms strong baselines across agent and domain-specific benchmarks (e.g., +10.6% on agents, +8.6% on finance). Furthermore, ACE significantly reduces adaptation latency (86.9%) and token costs (83.6%) compared to existing adaptive methods.

This community report summarizes discussions from a Nordita program on Quantum Gravity, highlighting the current status, commonalities, and disagreements among various theoretical approaches. It covers fundamental questions like the nature of spacetime, unitarity, causality, and stability, as well as the prospects for observational detection of quantum gravity effects. The paper emphasizes the lack of consensus on many core issues and the ongoing challenges in bridging theoretical frameworks with experimental data.