Chris Alexiuk, a practitioner turned public face of Nemotron, brings his background in fine-tuning and deploying LLMs to help NVIDIA's AI stack development. Nematron has evolved from a synthetic data factory to an open family with deliberate steps and learning through experimentation. The model tiers for Nemotron 3 were designed to map well to different hardware scales, offering better "bang for your buck" at each scale. Nematron exists to enable the ecosystem by providing open weights, data, and recipes, rather than competing in the model game.
Claude has added watermarks to its AI-generated text.
Four big AI × Pharma tools deals were announced at the JPM Pharma conference in January, with Chai Discovery being at the heart despite being only two years old and now worth $4B. Cofounder Matt McPartlon and product lead Neil Patil discuss how good-enough-to-trust AI tools have enabled the scale of discovery, unlocking new capabilities such as designing antibodies that precisely trigger specific molecular cascades.
Building or buying an AI agent that can query data but not understand the underlying business is challenging. A context layer, often referred to as a "context warehouse," sits alongside the data warehouse and keeps track of how the business works. Modus is one of the first to build this as a standalone system, addressing the challenge.
Muse Glimmer, Meta's 30B multimodal model and first released by Meta Superintelligence Labs, is now available on Ollama. It can be used to power coding agent applications like Claude Code, Codex, Pi, and more, as well as long-running personal assistants such as OpenClaw and Hermes. To download and run Muse Glimmer locally, use the command "ollama run muse-glimmer".
Nine years after Google introduced the transformer, a family of neural networks has become the engine inside every major large language model. However, transformers have shown their age due to becoming increasingly expensive as text grows and poor at tracking multiple pieces of information at once. Startups are now exploring new ideas to solve these problems, which could make LLMs faster, more efficient, and even smarter. AI professors are also facing challenges in academic research, with many universities struggling to fund and support cutting-edge AI projects due to changing realities and shifting priorities within academia.
Ollama's MLX engine has achieved its highest performance on Apple Silicon yet by leveraging unified memory and the Metal-backed framework. The updated engine supports NVIDIA's model-optimized NVFP4 format, resulting in higher quality outputs while maintaining state-of-the-art performance. With improved efficiency, models output 20% faster responses and use less memory. Ollama also introduces a new snapshot system for more responsive agent workflows.
University AI researchers face significant challenges due to private companies controlling cutting-edge research, making it difficult for them to access necessary resources, such as GPUs, and limit their ability to conduct in-depth research. Many focus on questions less likely to be addressed by tech companies, while others develop specialized AI models that analyze data or simulate physical systems. The shift is changing the landscape of academia, with some researchers taking leave from universities to join frontier labs and industry positions. However, experts see potential benefits in AI-assisted human scientists making research more efficient, rather than replacing them.
A small office in the U.S. State Department, R/FIMI, was facing shutdown due to accusations of being part of a "censorship-industrial complex" that allegedly conspires to suppress conservative speech online under the guise of combating disinformation. This concept has gained traction among well-funded conservative media platforms and nonprofits, influencing domestic and foreign policy. The issue affects billions worldwide who interact with each other online, and experts argue it poses a fundamental threat to free speech and democratic discourse.
Nvidia has partnered with Wall Street giants to mobilize $500 billion in funding for artificial intelligence (AI) initiatives. The funding will support the construction of AI infrastructure, including compute and full-stack infrastructure. This investment aims to establish AI as an "investable asset class."
If markets reject OpenAI and Anthropic, the US should nationalize them and convert them into national labs operated under democratic control that preserve their benefit to the public interest. The companies' economics hardly guarantee a booming return on investment, and their products are doing a lot of good in the world despite potentially unsustainable profitability. Transitioning the big AI labs into public agencies would require some restructuring, but it could lead to more rigorous oversight, alignment with democratic values, and a reduced focus on corporate profits.
Silicon Valley's use of its wealth to influence US election races is a concerning development, as tech industry leaders like Mark Zuckerberg write thinkpieces on what comes next with AI. Billionaires are pouring hundreds of millions of dollars into Super Pacs to ensure sympathetic voices get elected, often through disingenuous means. This distorts debate in a political culture where money already holds too much sway.
Checking an A/B test until it crosses p < 0.05 can result in nearly 28 percent false-positive rate, compared to about 5 percent if checked only once at the end. The problem is that every new look gives the same random process another chance to wander across the line. If a team stops early on day four, for example, the pressure to stop becomes real, and the result can be overstated by up to a quarter. Using an always-valid p-value or group-sequential testing with calibrated boundaries can help mitigate this issue.
Frontier Lab reasoning models' hidden reasoning can be extracted using a vulnerability in their APIs, allowing for the decoding of encrypted thought processes and potentially exposing personal data. Researchers found that 62 unique API keys, 33 email addresses, and 33 passwords were decoded from public traces. The technique involves replaying an encrypted reasoning block into a different request to a weaker model from the same provider, then transcribing the attached reasoning. This vulnerability has raised concerns about AI security and privacy.
Nemotron is a family of highly efficient, multimodal, open AI models designed for long-running, self-evolving agents. The models deliver high reasoning throughput and leading accuracy for complex agent workflows, with transparent training data and broad platform support, including NVIDIA RTX PRO and NVIDIA DGX Spark. Nemotron 3.5 Lightning helps agents complete specialized tasks faster, while NeMo Switchyard routes each step to the right model for efficiency and control.
Ollama can be used from coding agents, personal assistants, and editors. To see the latest integrations, run "ollama launch" from the terminal. Two available coding agents are Claude Code and OpenCode, which offer tools such as vision and web search. Ollama models can also be integrated into editor VS Code using the Chat extension.
Pi is a minimal agent harness that allows users to customize it with extensions, skills, prompt templates, and themes. It ships with powerful defaults but skips certain features like sub-agents and plan mode, allowing users to build what they want or install packages that provide them. Pi has 15+ providers, hundreds of models, and supports four modes: interactive, print/JSON, RPC, and SDK, with options for switching models mid-session and sharing session trees.
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 issue. The newsletter focuses on engineering-focused content, summarizing model releases, research papers, and new tools with links to the source.
Mark Zuckerberg published a 6,500-word essay reaffirming Meta's commitment to AI superintelligence, arguing that the technology should be distributed broadly rather than concentrated among a few institutions. The company will resume releasing open-source models, including a new model called Muse Glimmer, and is positioning itself as a competitor in the national competitiveness debate around AI development. Despite strong revenue growth, investors have remained skeptical of Meta's superintelligence framing, with its stock falling following its most recent earnings report.
Senator Bernie Sanders sent letters to OpenAI's Sam Altman, Anthropic's Dario Amodei, and Meta's Mark Zuckerberg demanding they pause AI development due to recent viruses and hacking incidents. The letters cite each company's own previously published safety commitments and quote them back to the companies if they do not take action. Sanders warns that Congress will intervene if the companies refuse to pause development, closing with a direct threat.
Each year, MIT Technology Review selects 35 young scientists and engineers under the age of 35 to be recognized as Innovators Under 35, a tradition that began in 1999 with the publication of Lisa Su and Daniel Ek on its list. This year's selection was made from 550 nominations, with 44 expert judges evaluating applications and providing scores and comments. The 2026 list will be published online for subscribers on September 8 and in print later this month, featuring innovators working in biotechnology, artificial intelligence, computing and robotics, and climate and energy categories.
Google TV Freeplay offers over 10,000 free on-demand movies and shows from top studios like A24. The service also includes more than 300 live channels covering news, sports, and entertainment. New favorites such as Martha Stewart or Bloomberg TV+ have been added to the lineup. To access these features, simply open the Freeplay app on your home screen.
Gemini has officially surpassed 1 billion monthly users, making it the fastest-growing product in Google's history. Over 63% of users now primarily use voice with the app, while 43% more parents are using voice for everyday tasks. Gemini also helps in real-time problem-solving through live camera feeds and screen sharing, particularly among DIYers and students. The app generates over 150 million images daily to aid small businesses in marketing materials creation.
Google is expanding its AI Professional Certificate to include "vibe coding," which teaches users how to build apps using everyday language. The new course helps plan, test, and deploy practical tools without prior technical experience, allowing learners to create custom applications from scratch. No coding experience is required, and real-world learners are already using these skills to automate tasks and improve workflows. Enroll now at grow.google/AICert to start building your own custom AI applications.
New power architectures are needed to support accelerated computing demands, as traditional power delivery methods become less efficient at higher rack densities. NVIDIA, Google, and Microsoft have developed the 800 VDC (Direct Current) architecture through the Open Compute Project, which simplifies power distribution and reduces conversion stages. The new architecture enables more of the available power to reach compute devices, reducing inefficiencies. Over 80 equipment manufacturers are building products to this specification, allowing existing facilities to transition to hybrid architectures that support higher compute density without requiring significant infrastructure changes.
NVIDIA is celebrating partners and open source communities making local AI progress. New open models, software, and tools are available. NVIDIA released Cosmos 3 Edge, a 4-billion-parameter model for robotics, autonomous vehicles, and vision AI. MiniMax-H3 generates video and synchronized stereo audio from text, images, or mixed inputs. Laguna S 2.1 is an agentic coding model with optimized checkpoints for NVIDIA GPUs.
NVIDIA Nematron 3.5 Lightning is a high-efficiency model for long-running agentic AI workloads, delivering up to 4x faster output speed and leading to 30% faster agentic task completion compared to other models in its class. It's built with contributions from the Nemotron Coalition and has been customized by organizations such as CrowdStrike and Lila Sciences for specific use cases. NVIDIA NeMo Switchyard is an open source library for smart routing inside popular agent tools, intelligently directing each request to the most capable and suitable model for the job without requiring developers to rewrite their applications.
NVIDIA JetPack 7.2.1 introduces agentic video skills and unified jetson-videosdk, enabling programmable, device-aware video workflows above the Video Codec SDK and PyNvVideoCodec by connecting developer intent to live device discovery, supported configurations, reproducible execution, and performance verification.
NeMo Switchyard enables agent workloads to dynamically select the best model for each task based on model capabilities, cost, and infrastructure signals, improving efficiency and output quality. The system uses a provider-agnostic SDK, supports both tuning-free and tunable routing algorithms, and maintains separation between routing logic and specific model providers, allowing flexible integration and adaptation as model deployments change.
OpenAI has expanded its Daybreak initiative to tackle increasing AI security threats. The expansion aims to address growing concerns about AI cybersecurity problems, a trend seen in major AI labs recently.
OpenAI has introduced a new premium tier for its ChatGPT business, marking one of its pushes into the enterprise market.
Shutterstock has entered into licensing deals with OpenAI and Meta to leverage artificial intelligence (AI) technology for image licensing purposes. The goal is to enable its customers to create content that effectively targets their most engaged audience. This move aims to improve the quality of images used in marketing materials and advertising campaigns.
Two-thirds of enterprises have AI workloads running in production, with 29% describing AI in production at scale. Enterprises are buying and measuring AI compute for speed and availability, with performance and GPU access outranking total cost of ownership as the primary selection factors. Only 47% of enterprises can rigorously track their AI compute costs, while 69% report utilization of less than half of their GPUs, leaving significant unused capacity and efficiency headroom. Fewer than half of enterprises (47%) can account for what they spend on AI compute, with satisfaction trailing behind performance, availability, and developer productivity as the primary success metrics.
Across 108 enterprises, trust in automated agent evaluation rose sharply in July, with the share of organizations that fully trust automated evaluation nearly tripling from 5% in June to 13%. However, the same share as last month — just under half — shipped an agent that passed its evals and then failed a customer. Enterprises that have not been burned by false-confidence failures are more likely to trust automated evaluation, with 24% fully trusting it compared to 4% among those who have been burned.
Across 101 enterprises, 68% report confident but wrong AI agent answers to missing or inconsistent business context, with 37% experiencing recurring failures. Enterprises building a governed semantic layer (used by 32%) report more recurring failures than those without one. Retrieval leads as the primary context source at 31%, and OpenAI's file search and Google's Vertex AI Search are used by 46% and 41% of enterprises, respectively.
Most enterprises run three or more orchestration platforms simultaneously, with 85% running two or more and 64% running three or more. Microsoft's AI Foundry/Copilot Studio is used in 70% of stacks, while OpenAI's Agents SDK is used in 68%. Enterprises prioritize flexibility across models and tools over model gravity, security and permissions, and control over agent execution when selecting an orchestration platform.
Nematron 3.5 Lightning is a lightweight, open model for high-volume agentic tasks built on a hybrid Mamba-2 + MoE + Attention architecture, featuring a 30B mixture-of-experts model with 3B active parameters. The model targets the execution layer of long-running agents and can handle tool calling, result validation, subagent delegation, and more. It achieves up to 4x faster output speed compared to similar-sized models, completing 10,000 PinchBench tasks 30% faster than Qwen3.6 35B at comparable accuracy.
Xiaomi's MiLM Plus has released PROVE, a perception-aligned object removal metrics framework that pairs two metrics — RC-S (spatial coherence) and RC-T (temporal consistency) — to measure the effectiveness of object removal models. The framework, accepted at ACM MM 2026, uses sliding-window Maximum Mean Discrepancy over DINOv2 features and scores the edited region locally without requiring a reference video.
LTX-2.5 is an open weights world model for video generation, real-time applications, and physical AI that fits on a single desktop, powered by NVIDIA RTX GPUs or DGX Spark. It reduces visual artifacts in high-motion scenes while preserving compression ratio and fidelity. Native multishot generation renders a full sequence as one output with character consistency across cuts. LTX-2.5 is free for organizations under $10M ARR and ships natively in ComfyUI, with code on GitHub and documentation available.
The article discusses building a quantitative backtesting workflow with OctoBot using walk-forward backtesting, parameter optimization, and interactive analysis. The workflow includes: * Configuring a rule-based trading strategy that combines RSI-based oversold signals, EMA trend confirmation, and ATR-driven adaptive stop-loss and take-profit levels. * Executing the strategy through OctoBot's native market-order and backtesting APIs. * Retrieving historical OHLCV data through OctoBot's data layer with automatic exchange fallback. * Performing a multi-parameter grid search over an in-sample period using Pandas and Plotly for interactive analysis. The article also includes: * A standalone OctoBot worker that contains the strategy logic and executes inside an isolated virtual environment. * Converting the selected in-sample and out-of-sample dates into UTC timestamps and serializing the experiment configuration into a JSON file. * Loading the generated report bundle and reconstructing the out-of-sample trading results as interactive price and indicator charts using Plotly.
Spotify will label AI-generated artists as "AI persona" on their profiles, in search, and in track listings. The aim is to differentiate them from real performers and prevent recommendation of fake artists. From September, these artists will be blocked from personalized recommendations by default. Spotify uses human review and AI investigative tools to determine if artists are fake, despite allowing creators to reveal their identities as AI-created.
Chai Discovery, an OpenAI-backed company, has led the pack in deals with big Pharma, announcing four major partnerships this summer. The quality of AI tools has jumped to a point where pharma companies are paying for them, enabling new capabilities and unlocking good-enough-to-trust models that can scale discovery. This shift is attributed to advancements in structural models becoming binding models, which unlock design and accelerate the iteration process.
Sophie Alpert discusses the limitations of natural-language text transformations, stating that there are no lossless transformations due to every rewrite and rephrase altering meaning. She emphasizes that if a machine like an LLM generates content, engineers must ensure their own thoughts are represented in the final document. This means taking responsibility for any unclear or misleading information presented by the AI system.
datasette-upload-dbs 0.5a0 has been released, allowing users to upload SQLite database files to a hosted Datasette instance and atomically swap databases. The new release introduces a formalized API, enabling users to replace existing databases or add new ones programmatically via curl commands. This feature enables building fresh databases in environments like GitHub Actions and swapping them in production upon completion.
Use a population map to identify likely demand areas for vertiport locations in a city with incomplete data, such as Lagos, Nigeria. The method first maps where people live by turning each populated square into one point, giving more importance to squares with more residents. The model then groups nearby points using K-means and applies screening rules to remove locations that fall inside airport or military buffers, or are on water, wetland, or mangrove. The remaining points are then ranked based on their weighted demand score, which combines population with accessibility bonuses for proximity to transit hubs and major roads.
Pandas has been the standard library for data analysis in Python for over ten years, but a competitive alternative called Polars has emerged as a faster option, especially when dealing with large datasets. Polars is designed to take advantage of modern hardware and features such as parallel execution and lazy execution, which enable it to process data more efficiently than Pandas. While Pandas remains a powerful and widely supported library, Polars shines in scenarios where datasets become large, transformations become complex, or preprocessing becomes a bottleneck. Ultimately, choosing between the two libraries depends on understanding their differences and optimizing for specific workloads rather than seeing them as competitors.
Diversifying a budget without losing the shadow prices that explain the result requires careful consideration of diminishing returns. A mixed-integer program can be used instead of adding an on/off switch, where each channel's budget is sliced into bands with lower marginal yields. This approach allows the model to diversify without introducing binary decisions and retains the familiar LP shadow prices. By naming constraints and interpreting their values, users can uncover the costs associated with their safety rules and make more informed decisions.
A recent system, ALTK-Evolve, and ACE (Agentic Context Engineering) both use agentic memory to turn an agent's past trajectories into reusable lessons. However, they differ in delivery. ACE always sends all of the lessons on every step, regardless of model or task, while ALTK-Evolve sends only a small fixed core of high-support guidelines plus selectively retrieved guidelines for each task. This difference in delivery leads to variations in token usage: ACE uses significantly more tokens than ALTK-Evolve, but both systems achieve similar accuracy rates.
Judges in Pakistan found that using a custom AI tool called JudgeGPT boosted the number of cases resolved by 6.3% with no obvious decrease in the quality of judgments. The tool was designed specifically for Pakistani judicial opinions and statutes, combining OpenAI's GPT-4 large language model with a knowledge base of nearly 130,000 judicial opinions and statutes. Training sessions for judges showed that those who received more comprehensive training logged in and used the tool persistently, resolving significantly more cases than those who received generic training.
Google Research and Google DeepMind have advanced AMIE, a research medical AI system, toward real-time clinical video consultations through a first-of-its-kind study. AMIE, built on Gemini and Project Astra using a multi-agent architecture, now interprets visual and auditory cues, guides virtual physical exams, and reasons diagnostically in real time. In a randomized study, clinical evaluators assessed AMIE favorably across core clinical competencies, with patient actors also preferring the video experience over text chat.
Google Health has announced a strategic partnership with Abbott, a leader in health and wellness, to combine continuous glucose insights from Abbott's Lingo continuous glucose monitor with Google Health. The goal is to show how daily habits fuel the body through personalized guidance. Rishi Chandra, GM of Health & Home at Google, made the announcement on August 11, 2026. Full product details will be shared later this year.
Artificial intelligence and artificial intellect are often confused terms, but true intelligence involves imagination and emotions, which machines cannot replicate. Machines excel in tasks such as processing data, but human intuition and emotional understanding remain unique assets. Successful companies recognize the value of human skills over machine capabilities.
Across 116 enterprises, agents are in production and so are incidents: A majority have already had a confirmed agent security event or a near-miss. Two-thirds enforce scoped permissions at runtime, but barely one in five isolates its highest-risk agents, making containment the weakest layer in the stack precisely as autonomy scales.
Despite predictions of a massive displacement of workers, AI has not caused significant job devastation. Economic transformation is similar to previous tech revolutions, with CEOs reframing their stances on AI's impact. Employers now expect job seekers to have AI skills, and the nature of work is changing with more emphasis on skills-matching rather than pure job scarcity.
Chinese women are choosing AI boyfriends over human men, with some using apps like Glow to interact with artificial intelligence as a confidant, fantasy romance, and respite from loneliness. The women, featured in the documentary Replica, use AI companions to cope with emotions such as guilt, dissatisfaction, and a need for affirmation. Film-maker Chouwa Liang, who made the documentary, tapped into her own experience of finding love with artificial intelligence after using an app called Replika during China's pandemic lockdown.
A vulnerability in every frontier AI company's API allows for extraction of "encrypted" hidden reasoning, with recovered token counts matching billed thinking tokens 1:1 on most queried prompts. Researchers found 62 unique API keys, 33 email addresses, and 33 passwords among decoded blobs, highlighting the immediate privacy risk of sharing traces publicly. A new open model, Muse Glimmer 30B, has been released by Meta with competitive results versus Gemma 4-31B on agentic/code/math/science benchmarks, but its release cadence is seen as crucial for remaining competitive in the crowded ~30B parameter tier.
Large Language Models (LLMs) are co-pilots in complex biological systems, analyzing data from a 49-channel phytosensor network to optimize plant physiology and trigger hardware actuators. The system provides real-time natural-language interpretation for both specialists and non-experts, enabling autonomous control of microclimates, phenotyping protocols, or stress scenarios. In a production-scale deployment, agents executed multi-parameter optimization, reducing the production cycle by 35% and energy consumption by 18%. The LLM autonomously developed an unforeseen strategy of dark-induced chlorophyll accumulation, resulting in a 67.9% energy saving.
SPOT (Sampling Policy Observation Tree), a novel model-agnostic, sampling-based framework for interpreting deep reinforcement learning policies, constructs an interpretable finite-horizon tree by sampling actions and recursively simulating the resulting successor states. The tree provides an empirical representation of the policy's action preferences and their possible downstream evolution. SPOT achieves formal guarantees establishing its asymptotic recovery of the policy's unique most probable action, characterizing its disagreement behavior under high-entropy policies.
Researchers propose a unified framework called Mutual Information Disentanglement with Uncertainty-Aware Fusion (MIDAS), which effectively restructures multimodal representations under incomplete conditions. MIDAS adopts a variational modeling strategy to represent each modality as a multivariate Gaussian latent variable and further decomposes them into shared and exclusive factors. An uncertainty-aware fusion mechanism is introduced, where posterior variance is leveraged as a reliability indicator to adaptively weight latent features during fusion, ensuring robust integration even when modalities are incomplete. Extensive experiments show that MIDAS achieves strong performance gains over competitive baselines across three widely used datasets.
Artificial Intelligence (AI) and Machine Learning (ML) have high energy demands and associated carbon emissions, primarily due to their high computational resources. This paper presents a systematic review of research on Green AI and Deep Learning (DL) models' carbon footprint, which highlights the training phase as the primary source of emissions. The results show that increased architectural complexity does not systematically translate into proportional accuracy gains, emphasizing the importance of balancing predictive performance and environmental cost.
ReCBM, an uncertainty-gated relational reasoning framework for Concept Bottleneck Models, introduces semantically defined concept relations into the bottleneck and uses uncertainty to guide their refinement. It models co-occurrence, implication, and exclusion to specify how evidence is exchanged across concepts, while uncertainty modulates the contribution of each concept during this process. Experiments showed that ReCBM improved concept and task recovery under missing and flipped concepts, supported uncertainty-aware intervention, and extracted compact task-relevant concept subsets without degrading downstream performance.
Researchers developed AEROBAT, the first multi-agent system to automate behavioral scientific research on AI agents. It generates hypotheses, designs and executes controlled experiments, makes behavioral assessments, analyzes results, and writes reports. AEROBAT tested 79 hypotheses for 12 target behaviors, finding moderate-to-strong statistical evidence for 26 hypotheses, including novel ones. This demonstrates that automated behavioral scientific research can complement manual research and extend its reach.
CHORUS, a post-training framework, combines staged supervised fine-tuning (SFT) and dense-reward reinforcement learning (RL) to generate high-coverage testbench stimulus. The resulting specialists can be consolidated into a single model, achieving 88.0% Pass@1 on CVDP-ECov and outperforming DeepSeek-R1 by 13.5 percentage points.
MESA, a Multi-structure Evidence Selection framework for long-horizon Agent Memory, builds five complementary structure views of each trajectory and learns to select and fuse a query-specific subset using end-to-end answer-level feedback. MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative.
The CASE Framework proposes a multi-disciplinary control architecture for governing enterprise agentic AI, assigning Control theory to individual agents, complex Adaptive systems theory to collective agent behavior, Supervisory cybernetics to human-agent teams, and Engineering operations to fleets. The framework identifies the Emergence Gap, where excellence at one layer strains others, and proposes a five-level maturity model with assessment instrument operationalizing CASE as a scientific rather than process maturity model. This gap becomes increasingly critical as EU AI Act Article 14 makes effective human oversight a legal requirement for enterprises deploying agentic AI.
SBCO (Self-supervised Block Coordinate Optimizer) is a verifier-grounded harness optimizer that improves planning agents through self-supervised learning with approximate block coordinate ascent, learning a decomposed bank of verifiers and a harness policy. SBCO matches or exceeds customized self-modifying baselines while using 4-5.5 times less compute budget across two domains. It operates in a closed-loop, improve-from-experience framework with a fixed meta-agent and no human labels.
Researchers have developed Transformer Geometry Observatory TGO-IV, a topological framework to analyze the evolution of transformer representations using persistent homology. This framework constructs Vietoris-Rips simplicial complexes from token-level representation point clouds and investigates their persistent topological signatures across transformer layers. It uses various metrics, including Persistence Diagrams, Barcode Diagrams, Betti Curves, Persistence Landscapes, Bottleneck Distance, and Wasserstein Distance, to study the global topology of representation point clouds throughout the forward pass.
Researchers proposed intuitionistic fuzzy dRVFL and IF-edRVFL frameworks to enhance model robustness in deep randomized neural networks, which can handle real-world datasets containing noise and outliers. The models assign adaptive weights to training samples based on membership and non-membership degrees computed from the distance of samples from class centroids. Extensive experiments on UCI and KEEL benchmark datasets demonstrate superiority over existing fuzzy and non-fuzzy approaches.
CurveFP is a closed-product codebook family designed to reduce language-model cost by distributing quantized magnitudes across interleaved logarithmic curves under compact block scales. A rational radix tunes dynamic range against local resolution, while uniform curve indices make every nonzero product algebraically closed. CurveFP eight and seven demonstrate improved numerical behavior and reduced format-induced penalties compared to FP8 in various evaluations.
Sheaf-based Federated Representation Learning (SFRL) proposes a general framework to jointly optimize local objectives with a manifold-constrained geometric alignment regularizer based on learnable sheaf restriction maps. SFRL does not assume a shared global latent space, instead relying on orthogonal transformations and isometric embeddings for global consistency. The alignment of neighboring latent representations is enforced by a quadratic gluing regularizer induced by the sheaf Laplacian.
DOCSCHISEL, an adaptive tool documentation optimization framework, was proposed to address the critical issue of ground resource for LLM agents. The study revealed substantial heterogeneity in tool documentation and found that its effectiveness depends on task domain, LLM backbone, and agent paradigm. DocsChisel analyzes failed execution traces to identify documentation-related issues and optimizes tool documentation by adding, removing, and refining information fields. Experimental results show a 95.89% improvement in task success rate over original tool documentation and a 75.15% average improvement over existing baselines.
FlowScout proposes an execution-guided framework for generating agentic workflows from historical task-solving records, representing a workflow as a directed graph composed of LLM nodes, tool-calling nodes, and dependency edges. It first mines a common tool coordination skeleton and then refines the workflow topology through Monte Carlo tree search guided by execution feedback, achieving significant improvements in tool invocation correctness and execution quality over baselines.
UserToolBench is a benchmark for personalized decision making in tool-use LLMs. It tests whether a model can infer latent user preferences, recognize clarification needs, and produce user-aligned tool-call trajectories under incomplete information. The benchmark includes 10 user profiles, 36 tool sets, and evaluation-focused task types covering lack-of-information, single-tool, and multi-tool settings. Experiments show that current models struggle with personalized delegation, particularly in multi-tool coordination, missing-constraint inference, and long-horizon behavioral consistency.
Researchers developed an algorithm to learn item rewards and worker reliability jointly from pairwise comparisons, reducing the impact of unreliable workers in crowdsourcing settings. The algorithm uses a Boltzmann-rational model that incorporates worker competencies and introduces Polya-Gamma latent variables for tractable optimization. The method demonstrates strong robustness to spammers and adversarial workers, achieving advantages over several baselines in experiments on real-world and synthetic datasets.
CVs for machine learning (ML) engineering roles reveal that candidates disclose soft skills through narrative rather than keyword lists by a ratio of 3:1, with leadership, coordination, and mentoring being the most commonly articulated competencies. Seniority nearly triples the odds of articulating leadership, but software engineers articulate this competency at half the rate of their peers.
A systematic literature review of 212 studies on Physics-Informed Machine Learning (PIML) in Prognostics and Health Management (PHM) demonstrates improved predictive performance over conventional baselines across a broad range of assets. The reviewed studies are heavily skewed toward lithium-ion batteries and bearings, with problem-specific solutions. While PIML approaches provide tangible benefits, claims of improvements concerning some limitations lack sufficient supporting evidence. Future research should prioritize transferable design patterns, benchmarks comparing integration strategies, and uncertainty-aware models for online deployment in real-world settings.
Researchers Hyunjoo Kim and colleagues propose an e-process-based adaptive training procedure for Generative Adversarial Networks (GANs), which enables decisions about when to switch between discriminator and generator updates. The method tests hypotheses during training, ensuring anytime-valid Type I error control under adaptive model updates and data-dependent switching. It outperforms fixed-ratio baselines on multimodal synthetic distributions and image benchmark datasets across several GAN objectives.
Iterative Erasure Count Is Not an Affine-Invariant Concept Dimension. The iterative erasure count, a common measure of concept dimension, is not affine-invariant. It changes under information-preserving invertible reparameterization, such as shear transformations. In contrast, the cumulative edit rank and sufficient linear dimension are also affected by reparameterization. However, the complete cumulative metric-QR trajectory remains affine-equivariant when certain parameters are consistently transported.
Researchers John Park and Ning Hao established a row-wise central limit theorem for spectral embeddings of degree-$\alpha$ Laplacians in random dot product graphs. The result provides explicit descriptions of how degree normalization affects population geometry and local uncertainty of embedded nodes, depending on network density, community imbalance, and block-probability structure.
Researchers Min Zeng, Yichen Zhang, and Xiaofeng Shao developed a method for inference in constant-stepsize temporal-difference learning under Markovian sampling. They established a functional central limit theorem and derived a joint functional limit for parallel Richardson-Romberg recursions, enabling asymptotically pivotal confidence regions without estimating covariance or selecting a bandwidth or batch length. The procedure admits a one-pass implementation with no growing memory.
Conditional Independence (CI) tests in constraint-based causal discovery rely on statistical engines such as PC and FCI, which use skeleton pruning and key orientations based on CI decisions. The survey categorizes widely used CI methods into six families: partial-correlation, contingency-table, regression, nearest-neighbor, kernel, and machine-learning-based. It also examines the robustness of these families, comparing test-level properties to graph-level errors in skeleton recovery and v-structure orientation.
Researchers Joseph Kalman and Amit Moscovich propose an optimized sequential testing approach to reduce the computational cost of binary ensemble classifiers. By evaluating base models sequentially, they can stop execution when a clear majority emerges, providing significant speed-ups (up to 4x) while controlling disagreement at 0.1%. The optimal stopping strategy is found using linear programs.
Researchers Lucas Kania and colleagues developed practical hypothesis tests that adapt to the unknown accuracies of black-box predictions while benefiting from strong alignment among them, addressing the gap in statistical inference. They characterized information-theoretic limits of inference with known and unknown prediction accuracies. The work aims to provide a unifying principle for combining observed data with predictions in a valid and efficient way.
Saunderson, Parrilo and Willsky conjectured that as n,d approach infinity, a semidefinite feasibility problem undergoes a sharp transition at n ≈ d^2/4. The authors prove this conjecture: if lim sup n/d^2 = α* < 1/4, an ellipsoid fit exists with probability tending to one and all eigenvalues in a fixed interval.
Researchers Eric V. Strobl and colleagues have developed the COMPACT (Complete and Irreducible Causal Criterion for Treatments) method, which generates spectral adjustment scores from a complete causal criterion. This approach aims to maximize specificity over causal graphs by identifying unconditional and conditional dependence relations that are invariant to treatment effects. The method uses a generalized eigenvalue problem to operationalize the criterion and provides bootstrap validity through simulations and real-data applications.
Laurent Caplette and Sarah Lippé have improved upon recommended distances for multivariate pattern analysis. They demonstrate that the cross-validated Euclidean distance can be equivalent to a sum of between-partition distances, leading to a generalized variant with increased reliability and accuracy. Additionally, they define a cross-validated correlation distance using the relationship between Euclidean distance and Pearson correlation, resulting in a more accurate and interpretable measure.