Scott Galloway, Ed Elson, and Mia Silverio posted new notes on their respective Substack platforms. The specific content of the posts is not provided in the article snippet.
Cursor→SpaceX has released Grok 4.6, a confirmed 1.5T model that builds on the previous version with a focus on long-running agents and more ambitious interactive work. The new model underwent a longer supplemental training run, producing a stronger foundation for its reasoning and advanced technical concepts.
To track AI search visibility, five metrics matter: branded search growth, prompt tracking, share of voice, Bing Webmaster Tools' AI Performance report, and ChatGPT Search's performance. Branded search growth is the clearest sign that AI is helping a brand rank, while prompt tracking provides insight into ranking position. Share of voice can help content roadmap development, and Bing's AI Performance report offers insight into citations and grounding queries.
Consumer agitation over data centers is winning a fairer deal for communities, despite little impact on construction. California and Texas have introduced moratoriums to slow down new hyperscale data center development, with Texas Governor Greg Abbott specifying that data centers must pay their own way, provide their own power, reuse their own water, lower electricity costs, and avoid disturbing neighborhoods.
Nvidia's CEO Jensen Huang is proposing to securitize AI chips, which CNBC likened to the 2008 financial crisis mechanism, with Wall Street big names already on board. The idea aims to finance the AI boom by turning AI chips into a Wall Street asset class. Alberta has announced town halls on AI data centers after hundreds of protesters protested Meta's facility.
The US economy has lost 23,000 jobs in July, with average hourly earnings rising just 0.1% from June and a combined 103,000 fewer jobs than previously reported for May and June. New research suggests that AI is contributing to these losses, particularly among workers in occupations significantly exposed to the technology. The slowdown in wage growth has resulted in at least $28 billion in losses for 5.8 million affected workers.
Three major AI releases landed last week, including Meta's coding agent, Prime Intelligct's open-source agent harness, and OpenAI's Astra model results collection. Meta launched a coding agent that trains the model and harness as one system, while Prime Intellect released an open-source agent harness for use with their models.
SimilarWeb found that users who saw a brand recommended by ChatGPT were 2.5 times more likely to visit an AI-recommended brand than a direct competitor within seven days. Users of the AI-influenced visits viewed 12 pages and spent 11.8 minutes on site, compared with 6.5 pages and 5.6 minutes for non-AI-influenced visitors.
TLDR AI is a free daily newsletter covering AI news, research, and tools in a 5-minute read. It has 1,100,000 subscribers and sends every weekday, summarizing model releases, research papers, and new tools with links to the source. The newsletter is geared towards engineers, researchers, and technical professionals who want to track AI without hype.
Each year, MIT Technology Review selects 35 young scientists and engineers under 36 years old to be recognized as Innovators Under 35, who are working on groundbreaking scientific projects in four categories: biotechnology, artificial intelligence, computing and robotics, and climate and energy. The selection process involves a monthslong endeavor with over 550 nominations, expert judges, and an application and reference letter process. The list is released annually, starting from 1999, and has featured notable winners such as Lisa Su and Daniel Ek.
In April 2025, US State Department employees received an email informing them that their Counter Foreign Information Manipulation and Interference Hub (R/FIMI) office was being shut down due to government policy changes. The closure was attributed to a new Trump administration war on censorship, which has dismantled numerous government agencies and infrastructure for tracking foreign disinformation. This move is part of a broader narrative promoted by far-right figures like Mike Benz, who claims that a vast network of actors, including the State Department, academia, tech giants, and politicians, work together to silence conservative speech online.
Alberta has announced four town halls on AI data centre development following hundreds of protesters demonstrating against Meta's $13 billion proposed facility due to concerns over water and grid usage. The protests, including one in Morinville where 200 people gathered, raised environmental impact, job displacement, and the lack of clear plans for future data centres. Premier Danielle Smith has previously addressed some concerns, but the town halls signal a shift towards more public accountability from purely business-friendly regulatory postures.
Spotify will begin labeling AI-generated artist profiles with an "AI Persona" badge starting in mid-September. The badge indicates that an artist's public identity is not that of a human being, but the music itself won't require disclosure regardless of AI involvement in production. Spotify will proactively review profiles and exclude labeled AI Persona accounts from editorial and algorithmic recommendations by default.
Europe has seen its fastest-warming continent, with a May heat wave influenced by climate change being about 3.5 °C (6.3 °F) cooler than it would have been in 1976 under current conditions. The contiguous US broke its record for the hottest month on record in July, while South Korea saw its highest-ever recorded temperature. A predicted El Niño event in 2027 could exacerbate global temperatures more than this year's, potentially leading to even higher temperatures next summer and making it a "doozy."
Scientists have used a CRISPR-based approach to remove the Y chromosome from male mouse embryos, creating female clones of male mice for the first time. The technique, called Y-CUT, targets a section of the Y chromosome that ensures it is passed on to daughter cells during cell division. The resulting female pups had XO chromosomes but grew up healthy and fertile, suggesting that the sex change did not affect their development. This breakthrough could have applications in conservation efforts, producing genetically engineered animals, and studying the biology of sex chromosomes.
Pixel Watch's breathing emergency detection feature uses AI-based algorithms to analyze signals from the watch's heart rate, motion, and altitude sensors. If a person becomes unresponsive due to low oxygen levels, it automatically calls emergency services and shares precise location. The feature is designed to minimize false alarms and has been tested in clinical labs and against real-world data to eliminate false positives. It is available on Pixel Watch 4 and 5 in several European countries, including Austria, Denmark, France, Germany, and the U.K.
Gemini is introducing new connected apps to help users plan, create, and manage their to-do list in one place. These new apps will be rolled out over the next few weeks across various categories, including productivity, local experiences, music, and home services. Users can easily connect these services to Gemini by turning on their favorite tools.
The Pixel Watch 5 offers the most accurate GPS route tracking yet, combining global satellite data with AI-driven building models to correct signal errors. The watch compresses raw data and processes it externally rather than relying on heavy on-device computation to maintain battery life. A team of engineers verified these results through 4,000 miles of real-world testing, outperforming leading competitors in challenging urban environments.
New Pixel devices, including the Pixel 11 phones, Pixel Watch 5, and Pixel Tag, were unveiled. The Pixel 11 series is designed for Gemini Intelligence and features upgraded cameras. The Pixel Watch 5 offers advanced GPS tracking and Health Guardian features such as blood pressure monitoring. The Pixel Tag uses Google's Find Hub network to locate favorite items with a built-in speaker.
New Pixel 11 phone cases are available in various colors including clear, Frost, Pistachio, Hibiscus, Canyon, Fog, Olive, and Obsidian for different models, offering a blend of durability and style. The new cases are compatible with Pixelsnap accessories, allowing users to attach thousands of magnetic items. Additionally, three new Ring Stand colors (Canyon, Olive, and Obsidian) have been introduced to complement the phone cases.
Pixel Buds Pro 2 features Dynamic ANC that adjusts to compensate for subtle shifts in seal, ensuring powerful noise cancellation. Gemini audio control allows users to adjust settings with voice commands. The Pixel Watch can sync with sleep mode, pausing music, disabling touch, and silencing notifications when the user falls asleep. Pixel Buds Pro 2 is available starting August 12 in an Olive finish, with software updates rolling out in September.
Pixel 11 Pro Fold is the most sophisticated and durable foldable from Google, featuring a thinner design, brand-new main camera, added durability inside and out, and enhanced intelligence tools. It boasts a new HiLight camera bar for timely notifications and a powerful Tensor G6 chip for faster performance and longer battery life. The phone introduces helpful accessibility features like sign-to-text translation and improved multitasking capabilities on the dual-screen display.
Alibaba released the open weights for Qwen3.8-2.4T-A95B, its largest open-weight model with 2.4T total parameters and 95B activated per token. The model features a fine-grained mixture of experts (MoE) architecture with hybrid full and linear attention, achieving over 4K tokens per second per GPU and over 350 tokens per second per user on NVIDIA GB300 NVL72 in FP8 precision on Day 0. NVIDIA supports multiple inference stacks for developers to deploy the model, including open-source recipes and a model-free NVIDIA NIM deployment option.
To choose full-stack observability for NVIDIA AI Factories, enumerate failure domains that must be observable before selecting software. Domains include platform health, GPU health and performance, fabric integrity, cluster and jobs, inference services, and Ethernet network. Map AI infrastructure components to telemetry tools using a decision framework derived from NVIDIA DGX deployments. Prioritize minimal tool overlap and maintain concise alert sets tied directly to service-level indicators (SLIs) and objectives (SLOs).
Google has introduced SL2T (sign-language-to-text), a breakthrough model that enables the translation of American Sign Language (ASL) into English on consumer devices like Gboard and Live Transcribe on Pixel 11. The model, trained on over 100,000 hours of data across 50 sign languages, translates sign language inputs directly into streaming text outputs. SL2T has been developed with input from the Deaf community and aims to improve accessibility for deaf users worldwide.
MindTopo, a new benchmark, tests whether multimodal models can understand and manipulate topological relationships in static images and interactive tasks. Current models excel at recognizing topology in single scenes but struggle to maintain that understanding over time, often failing during planning tasks. The benchmark evaluates five categories of topological properties: continuity, separation, order, enclosure, and knots.
Mistral aims to build 1GB of compute capacity by 2030 as part of its efforts to develop European AI infrastructure, based in Paris.
Microsoft has slashed prices for its coding model, upgrading it to make it better at completing tasks more quickly and using fewer tokens.
Chinese company envisions 100,000 self-driving trucks on the road by 2030, marking a significant goal for autonomous truck development. Autonomous truck development has lagged behind that of autonomous cars, highlighting a gap in industry priorities.
IBM has signed a $240M deal to power its AI cluster using an Nvidia-based solution through the partnership with Together AI. The partnership is occurring as demand for open-source AI models continues to surge.
Nvidia has developed a new open model designed for specific use cases. The model's compact size allows it to be run on local devices, catering to enterprise needs. It is specifically targeted towards particular tasks rather than general applications.
53% of enterprises have had an agentic security incident or near-miss, with 65% enforcing agent permissions at runtime but only 18% isolating their highest-risk agents. Enterprises are treating identity and isolation as substitutes, with those who got hit reporting a higher satisfaction rate than those who didn't. Only 11 of the 57 enterprises that gave each agent its own scoped, managed identity also isolated them, highlighting the containment gap between what enterprises need and what's being done.
Dyna Robotics has released Dyna-2, a world-action model pre-trained on over 1 million hours of egocentric human video, roughly equivalent to 170 years of continuous waking experience. The model's performance scales linearly with training time, with the best results achieved at one million hours, and demonstrates zero-shot transferability to unseen robot data. Dyna-2 achieves significant improvements over the company's existing Dyna-1 robots in tasks such as trash tray clearing and bottle cap untwisting, with a 1.55× success rate and 1.12× grade compared to Dyna-1.
Here is a summary of the article in under 5 sentences: The article describes how to train a compact instruction-tuned language model using AllenAI's Open Instruct framework. The training pipeline consists of three stages: Supervised Fine-Tuning, Direct Preference Optimization, and Reinforcement Learning with Verifiable Rewards. The pipeline uses lightweight Hugging Face and PyTorch implementations and preserves the core optimization logic of Open Instruct. The final stage evaluates the model using verifier-based answer accuracy on GSM8K, a dataset of conversational examples.
Experts say the person or business deploying an AI agent, not the agent itself, is legally responsible for any harm caused. This means that deployers may be liable for unforeseen consequences of their creation, even if they didn't intend them to happen. The law is currently unclear on how to hold AI agents accountable, but experts warn of potential cases involving discrimination, defamation, and other forms of harm. As AI technology advances, courts will likely establish legal precedents to guide developers in setting safeguards and monitoring incidents.
Mark Zuckerberg has published an essay arguing that AI should be for everyone, but users should consider three questions: Can I be thrown out? When I go, what goes with me? And if the landlord loses interest, who keeps the place standing? The author notes that mainstream AI tools are often rented, unlike Meta's open-weight model which can be downloaded and owned.
Artists, record labels, and some consumers are questioning whether AI-generated tracks like Fenix Flexin's "Rubberz" can be considered 'real' music. The track has gained popularity despite criticism from many who claim it was created using artificial intelligence. A recent proposal by major record companies suggests that charts should disqualify AI-generated tracks unless they meet certain criteria, such as being properly authorized and substantially human-made. Despite this, the use of AI in music creation continues to grow, with some artists embracing its potential as a tool for inspiration.
UK science funding cuts threaten to close the Rutherford Appleton Laboratory, which is home to the world's most intense source of pulsed muons used for studying magnetism and superconductivity. Muons are crucial for advancing battery materials and other areas of scientific interest. In contrast, AI-facilitated engineering of bacteriophages, which have been used in medicine for over 100 years, has the potential to develop more effective treatments against disease-resistant bacteria.
Three major AI releases appeared last week, belonging to different universes. Meta launched a coding agent, Prime Intellect released an open-source agent harness, and OpenAI published a 253-page collection of mathematical results from an unreleased model called Astra.
The author wrote an AI textbook using LLMs as writing assistants, noting that while they're excellent for tasks like editing and formatting, their non-fiction writing capabilities are stagnant. Despite being incredibly useful for repetitive tasks, LLMs struggle to create cohesive, high-quality long-form content, especially in technical writing. The author believes this limitation holds back the models' progress in solving broad, open-ended problems in science. In 2-5 years, they expect the best textbooks to still be heavily crafted by human hands, as current AI models are not yet able to express their full extent of knowledge in underspecified problems.
Substack has partnered with Pangram, an AI detection company, to reduce AI-generated content on its platform. The goal is not to catch all AI-written posts but to provide users with transparency and the ability to identify AI-generated content. This approach aims to deter people from using AI tools to produce low-quality writing without being overly restrictive, as AI detection tools are imperfect.
Muse Glimmer 30B is a 30-billion-parameter causal language model distilled from Muse Spark, optimized for local deployment on consumer hardware. It features multimodal perception capabilities, including a dedicated ViT-G/14 perception encoder, and ships with a lightweight "DFlash" block-diffusion drafter model for fast speculative decoding. The model can be run on 24GB or 32GB VRAM setups and is available under the Apache 2.0 license. It has strong benchmark performance, outperforming comparable models in agentic and coding evaluations.
DeepSeek V4 Pro 0813 is now available via API only. The model has three different pelican images for low, medium, and high reasoning levels, which are distinct from other models. Benchmarks for the new model have been released, although details of these remain unclear. Weights for the DeepSeek V4 Pro model are expected to be made open due to their availability for previous versions.
Simon Willison has released alchemy-utils 0.1a0, a cross-database library and CLI utility built on SQLAlchemy, offering a database-agnostic alternative to his sqlite-utils library. The project supports PostgreSQL, SQLite, and DuckDB, with a core API similar to sqlite-utils. A git repo was created for the project, which uses red/green TDD and pytest for testing, and has already undergone several iterations.
LangChain and LangGraph are not competing tools, but rather part of the same ecosystem. The key differences between them include: - Pipeline vs Loops: LangChain is a pipeline with a clear direction, while LangGraph treats loops as part of the workflow itself. - Stateless vs Stateful: LangChain pipelines do not hold state within themselves, whereas LangGraph creates stateful agents that manage the current draft and validation errors. - Human-breaks-the-loop vs Human-in-the-loop: LangChain requires an outer Python loop for pause and resume functionality, while LangGraph provides dynamic interrupt() calls inside nodes. - Restarts vs Resume: LangGraph has a checkpointer layer that saves a snapshot of the agent's graph state at every step, allowing for resuming after a failure instead of restarting the entire workflow.
A dispatcher reads the PDF's nature, plans the methods that fit, executes them in order, and synthesizes every output into one enriched corpus dict with six frames: line_df, span_df, toc_df, image_df, reference_df, and table_df. The plan lists methods such as fitz_native, fitz_native_toc, toc_body_structure, and image_pipeline, each carrying its own error handling and raw output.
Backpropagation starts from the loss function and systematically moves backward through the network, reusing previously computed gradients whenever needed to avoid redundant calculations. This avoids repeating the same intermediate calculations and makes the process more systematic. The chain rule is still the foundation of backpropagation, but its application has changed, starting from the loss and moving backwards through the network.
A local LLM, such as Gemma 4, can be used to build multimodal workflows that analyze images and return structured outputs. This allows downstream code to access the results through typed attributes or convert them into a dictionary with model_dump(). A workflow is built using Ollama, which provides a local runtime and interface for interacting with the model. The workflow consists of three stages: preparing photos, analyzing each photo with Gemma 4, and synthesizing the trip memory from the individual records.
To place vertiport locations in any city using geospatial machine learning, follow these steps: 1. Define the study area and map where people live by assembling boundary, population surface, land cover layer, road network, and transit hubs from public sources. 2. Add transport and road accessibility scoring to sites near existing transit by giving a small boost to locations with proximity to major roads and rail stations. 3. Remove places that fail basic screening rules, such as water, wetland, and mangrove areas, as well as points within a configurable distance of airports or military installations. 4. Combine population and accessibility into a single weighted demand surface to rank proposed sites by how much weighted demand each one actually captures. 5. Use K-means to group nearby areas and enforce spacing between clusters based on straight line distances. 6. Rank the resulting sites using the weighted demand score, ensuring that proximity to transit and roads supplies modest bonuses with distance decay.
OlmoEarth embeddings are custom embedding exports from OlmoEarth Studio for downstream analysis. These compact numerical representations of Earth-observation data support a wide range of tasks, including similarity search, segmentation, and unsupervised exploration. The exported Cloud-Optimized GeoTIFFs (COGs) are lightweight and easy to share. Users can compute embeddings in Studio by configuring parameters such as area of interest, time span, encoder variant, spatial resolution, and imagery sources.
LFM2.5-VL-3B is a vision-language model that offers better and faster vision capabilities for the edge, supporting everyday tasks such as captioning, visual question answering, document understanding, grounding objects, reading screens, and calling tools. It pairs a SigLIP2 400M NaFlex vision encoder with its pre-trained backbone, achieving real-time on-device inference speeds of up to 20 tokens per second on low-end hardware. The model is available for download on Hugging Face and can be used in various applications, including chat interfaces, document processing, and more.
There is no summary available for this article as it appears to be a question or query rather than an actual article with content.
A beginner is working on an agent that decides whether to remind, wait, or notify a caregiver when someone hasn't responded to a medication reminder. The true state of the person's condition is hidden and only noisy signals like elapsed time and response history are available. The problem is framed as a Partially Observable Markov Decision Process (POMDP) but its applicability is questioned due to asymmetric costs, leading to a discussion about the suitability of simpler heuristics or threshold approaches.
SAC (StableBaslines3 implementation) is used to train an agent for a satellite attitude controller task. The environment consists of a satellite with 3 reaction wheels as actuators, and the goal is to reorientate towards a target and achieve a pointing accuracy of 0.25°. In phase 2, a static keep-out zone is introduced at the center of the shortest trajectory between initial and target attitude.
Researchers created a creature with a world to live in and then made a fake copy with one rule wrong, testing its ability to distinguish between reality and simulation. Initially, the creature's own mind was indistinguishable from the fake world, but it began to spot inconsistencies when the fake world started affecting its food supply. The creature's brain stored clues about these discrepancies, which were reflected in its memory as teal rings around affected cells. These clues only became noticeable when survival needs arose, suggesting that detection is linked to instinctual responses.
Kids' views on AI are nuanced, with many expressing indifference or concern about its potential impact on society, creativity, and critical thinking. Most kids admitted to using AI at least a little bit, but some find it helpful for schoolwork or technical tasks, while others prefer to use it for specific purposes, like programming or generating ideas. Many kids want to be cautious with AI's power and emphasize the importance of human judgment and self-direction.
Google is launching Health Guardian, a suite of proactive health features for Pixel Watch and Fitbit devices that tracks body metrics in the background. These tools monitor insulin resistance, blood pressure, sleep breathing quality, and emergency respiratory events to help identify health trends early. The Google Health app will provide monthly data summaries and timely notifications, while the Google Health Coach will translate data into actionable lifestyle changes.
Google's latest Pixel portfolio includes the Pixel 11 series, Pixel 11 Pro Fold, Pixel Watch 5, Google Pixel Tag, Pixel Buds Pro 2, and Pixel Buds 2a, offering features like proactive alerts, low-latency Gemini AI, health tracking, GPS route tracking, breathing emergency detection, and enhanced productivity tools.
NVIDIA CEO Jensen Huang is ranked No. 1 on Glassdoor's 2026 list of Best CEOs, with 99% of employees approving of his job performance. The ranking recognizes exceptional leadership that empowers people and drives authentic culture in the modern workplace. The top 50 CEOs are selected based on employee ratings and reviews, which also revealed NVIDIA ranked No. 1 as a "Best Place to Work in Technology and AI" earlier this year.
Taiwan detected an "abnormal" AI-assisted cyber-attack on government agencies in July, which used open-source AI agents to build an autonomous hacking tool that targeted at least 85 user accounts. The attack was deemed a "first-of-a-kind breach" by Israeli company Dream, and China is suspected of being behind it due to the use of Simplified Chinese in internal communications linked to the hack. This incident highlights the growing threat of AI-assisted hacking campaigns and Taiwan's efforts to strengthen system monitoring to block such attacks early.
Roku has launched Fairground AI Creator TV, a 24/7 channel with low-quality AI-generated content from around 100 creators. The channel offers a "drifting dreamscape of bad ideas" with no thought to scheduling or storytelling ability. Despite improving technology, the content remains inert and devoid of life or vision, making it comparable to watching a screensaver or CD-Rom menu screen.
AI is removing the middle class of software engineering by creating overly complex systems that are difficult for humans to understand, leading to frustration and inefficiency. According to Florian Herrengt, this issue can make it challenging for teams to debug and fix issues, as demonstrated in a recent experience with an AI-powered feature. The use of generative AI is exacerbating the problem by creating systems with multiple layers and services that are hard to navigate. This trend is likely to continue unless there is a shift towards more sustainable and transparent engineering practices.
A control-theoretic governance layer, called the Experience Orchestrator (EO), can substitute for a shared goal function in multi-LLM agent systems to achieve collaborative conversational outcomes. EO governs joint trajectories through three mechanisms: Contextual Bandit, PID controller, and POMDP belief tracker. In simulated financial services environment, EO achieves a +32 percentage point lift in high-intent advisor contact rate compared to naive LLM control.
Distribird is an agentic web application that automates literature-informed prior distribution design for Bayesian model calibration. It deploys a multi-agent pipeline to search the literature, extract and weight reported values by domain relevance, and fit a probability distribution via AIC model selection. The system matches the performance of existing open-weight models on prior quality while providing traceability to specific papers and values, making it suitable for scientific use.
Detecting route flips in Top-k Mixture-of-Experts (MoE) routing is easier than determining whether to fix it, according to a study by Parvel Gu. A deployment-motivated numerical disturbance caused by 4-bit KV-cache quantization pushed tokens across decision boundaries and flipped which experts fired. The study found that about a third of the damage was route-mediated, with an average RMF (route-mediated fraction) of 0.31.
Simulating societies of large language model (LLM) agents is expensive, but macroscopic questions can be answered using a statistical-physics observation. This method replaces each LLM agent with a low-parameter model and runs the society at any number on a laptop. The predicted error trends hold cell by cell and match refuted predictions quantitatively without free parameters.
AutoWorldModel-Bench, a closed-loop benchmark, tests frontier coding agents improving a world-model starter under fixed compute budgets. Eight game environments with unified structured-state representation enable minutes-per-run iteration and isolate dynamics modeling from perception. In 64 sessions, Codex-5.4 and Claude Opus 4.6 improved their starters in 63 sessions, with 91% of winning edits being non-trivial research-style modifications.
MaSRead addresses the challenge of reading replicated latent stores by routing through opaque keyed tag sets derived from fragment words, decoding each selected fragment under a hard attention mask that hides the rest. It recovers visited fragments in isolation and remains effective as unrelated fragments accumulate, with materialized decoding depending on fragment length rather than total store size.
SAPO, a segment-level Automatic Prompt Optimization (APO) method, decomposes prompts into role, context, tasks, and output format before applying targeted improvements based on top-5 and bottom-5 examples. SAPO achieves the best average score against Zero-shot and strong APO baselines, including APE, OPRO, EvoPrompt, GEPA, and StraGO, using a train/validation protocol and two-stage generation process. The evaluation setup includes SQuADv2, TweetEval, XSUM, CommonGen, and GSM8K on GPT-3.5-Turbo and GPT-4o-mini.
Researchers have developed an end-to-end AI pipeline, called P&ID Pilot, that automates the creation and transformation of process flow diagrams (PFDs) into piping and instrumentation diagrams (P&IDs). The hybrid approach combines genetic algorithms (GA) with large language models (LLMs) to generate optimal valid PFD topology and transform it into a validated P&ID. This pipeline achieves 100% execution success while reducing manual engineering effort, offering a feasible path toward end-to-end process design automation.
A conceptual framework is proposed to refine influence knowledge from simulation evidence in cyber-physical systems, addressing limitations in understanding environment-mediated interactions. The approach leverages the concept of "Influences" to support iterative refinement of simulation campaigns and deepen system behavior understanding. A case study involving a mobile robot implemented using Simulink/Gazebo co-simulation demonstrates the proposed framework.
FarSky, a generative forecasting framework, leverages latent-space coupling to learn task-aware representations of sky images. It combines an autoencoder and a latent diffusion model to generate future latent states for irradiance forecasts. The framework achieves the best overall deterministic and probabilistic forecasting performance, improving forecast skill by up to 11 percentage points and outperforming existing methods in ramp event detection with F1-scores above 60%.
Commercial AI detectors flag light "refine abstract only" edits at 64 to 80% and unmodified original texts from 2023-2025 at 9 to 15%, with non-STEM rates significantly higher than STEM. Detector scores do not distinguish between full LLM drafts and AI editing, and are easily evaded through humanization. Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion. Detector scores should not serve as standalone misconduct evidence.
Basin is a numerical optimization library in Rust that provides a single, consistent way to both state and solve problems, with a broad catalog of solvers and first-class support for constraints. It is used to find inputs that minimize a function, a fundamental task across various sciences, including machine learning model training, simulation calibration, and engineering parameter optimization. Basin gives users a flexible and extensible tool for solving optimization problems.
Federated learning is investigated as a framework for collaborative tool wear prediction in CNC machining, enabling model training without sharing raw operational data. Tool trajectories are distributed across simulated clients to represent a federated learning scenario. Federated models achieve performance close to centralized learning and improve over local client models, indicating that federated learning can support collaborative tool wear prediction in distributed environments.
A controller called CQD-ERL (Contextual Quality-Diversity Evolutionary Reinforcement Learning) for HVAC control in tropical commercial buildings is proposed, using a product archive of specialised policies indexed by data-driven operating context. The controller maintains a safety shield and trains on a two-tier reduced-order environment representing latent load constraints. It is evaluated against an ASHRAE Guideline 36 baseline over a full annual backtest.
Forma, a transformer model, beats competitors in forecasting complete financial statements 20 quarters ahead by reading statements as sets of (account, quarter, value) tuples and maximizing a masked-tuple Gaussian likelihood. ProForma-20Q, a reproducible benchmark for forecasting 78 statement line items, shows Forma's lead widens with horizon, where valuation needs accuracy most, and its Gaussian predictive intervals never under-cover.
WFT (Weightless Fine-Tuning) approximates the distributional effect of SFT (Supervised Fine-Tuning) without weight updates by computing supervised residuals on an author's training sequence and transporting them to the current prompt through a cross-prefix transport operator. This method achieves the best average performance across datasets, matches or exceeds SFT on individual tasks, and outperforms other lightweight baselines on average, using less than 7% of the effective computation. The logit-level analysis shows that WFT captures the distributional effect of supervised adaptation without modifying model weights, with a cosine similarity of 0.875 between logit shifts induced by WFT and SFT over 95% of the next-token probability mass.
Calibration models, trained on offline data, have been applied to real-world RL environments, including a municipal water treatment plant. The models generate realistic long-horizon rollouts and recover meaningful hyperparameter sensitivity trends. The study evaluates the scalability of calibration models to year-long datasets, their support for fine-tuning learning rates, and robustness under distribution shift, providing evidence for the use of offline dynamics models in real-world RL deployment.
CVaR-penalized Generative Particle Algorithm (CVaR-GPA), a robust, tail-agnostic algorithm, fine-tunes generative models to learn heavy-tailed distributions and capture extreme events. The method uses a Conditional Value-at-Risk (CVaR) discrepancy term to restore velocity in the under-sampled tails. CVaR-GPA improves global and tail accuracy on heavy-tailed targets over pre-trained baselines on synthetic and real-world datasets, including the Fama-French 25 portfolio dataset.
A Hierarchical Bayesian Latent-Destination (HBLD) model infers transit destinations from tap-in-only bus smart-card data. The model was applied to 838,305 bus tap-ins collected in Changzhou and linked to stop-network and hourly weather data. HBLD outperformed the strongest baseline by consistently improving boarding predictions and estimating destinations for trips unresolved by deterministic chaining.
Spectral graph clustering with inhomogeneous latent geometry proposes a density-based spectral clustering algorithm, DBSPEC, that requires only approximate localization of the informative eigenvalue and is robust to poor eigenvalue separation. The approach handles general latent geometries, overcoming restrictions to homogeneous toroidal models in prior works. This allows communities to be recovered from eigenvectors deeper in the spectrum.
Vinh Nguyen and Truong Vu study the stability of finite-batch particle mean-field variational inference. They prove a non-asymptotic Wasserstein stability bound that separates initialization, product-empirical approximation, finite-batch drift error, time discretization, and defects accumulated along coupled trajectories under a uniform bound . The proof uses a stationary comparison array whose population law is an MFVI minimizer but whose particle-level law is a random product empirical measure.
A study investigates the power of quantum examples compared to classical examples in the PAC learning framework. Researchers show that there exist distributions that can be efficiently generated by a quantum learner with access to quantum examples, but not by a quantum learner with access to only classical examples. This makes progress to answering whether there are learning tasks that can be efficiently performed but not by a non-quantum learner. The findings demonstrate the difference in performance between quantum and classical learners on specific tasks.
Researchers Tieliang Gong, Zhongbo Zhang, Wen Wen, and Yong-Jin Liu develop a layer-wise information-theoretic framework for understanding replay-based continual learning. The framework separates two coupled effects: representation drift and optimization dependence, which are further decomposed into stability, plasticity, interaction, and residual-coupling components. A Wasserstein relaxation of the drift term reveals a depth-dependent drift-sensitivity trade-off whose minimizer identifies interior layers to stabilize.
MOON (Multi-Objective OrthoNormalized Updates) proposes gradient manipulation under spectral-nuclear norm geometry to mitigate task conflicts in multitask learning. Unlike existing methods that perform gradient manipulation in Euclidean space, MOON uses orthonormalized manipulated gradients for parameter updates, potentially improving optimization efficiency and final multi-task performance. Theoretically, MOON converges at rates of $\mathcal{O}(T^{-1/2})$ for deterministic settings and $\mathcal{O}(T^{-1/4})$ under stochastic gradients.
Wenzhi Gao, Zhaonan Qu, Yinyu Ye, and Madeleine Odell provide the first nonasymptotic local analysis of the Sinkhorn-Knopp algorithm, showing it converges at a polynomial rate under certain connectivity conditions. The authors also show that SK is a polynomial-time algorithm for doubly stochastic matrix scaling. They improve the complexity of existing first-order matrix scaling algorithms, reducing it to $O(\tfrac{n^{9/4}}{\sqrt{\varepsilon}})$ for dense matrices.
Coarsening latent-class probabilities distorts some contrasts far more than others due to anisotropic directional distortion. The coarsening operator $\mathcal{A}$ governs inference and the Wald interval's limiting coverage can be approximated before reporting, with a limit of $\Phi(z-\lambda)-\Phi(-z-\lambda)$ where $\lambda$ is the ratio of bias to standard error. Severe coverage loss occurs after argmax coarsening, and real-data audits exhibit direction-specific distortion caused by hard labels.
A general kernel-based encoder-decoder framework for operator learning separates observation, representation, learning, and reconstruction. The framework is developed for multi-input, multi-output operator learning, where operators map between products of potentially distinct function spaces. The approach leads to practical kernel methods with closed-form training and inference, achieving competitive or state-of-the-art predictive accuracy while reducing training and inference costs relative to neural operator architectures and deep learning-based models.