Current AI builds a free global web for AI systems using $400 million in funding from the French government and other organizations.
A Power Query approach assigns seats to organizational units (OUs) based on a floor-by-floor rule, where unassigned seats are assigned to the OU that owns the most seats in that room or on the entire floor. To do this, an M-function is created to find the latest file, read its data, and count the number of seats per OU in a room. The function then sorts the results by seat count in descending order and assigns the first row's OU to unassigned seats. This process must be performed twice: once for the same room and once for the entire floor. However, this approach performs poorly on network-based sources due to performance issues. A DAX solution is proposed as an alternative, which creates two calculated columns using SUMMARIZECOLUMNS and TOPN functions to determine the OU with the most assigned seats in a room or on the entire floor. The final result shows that the DAX solution is less complex than the Power Query approach but has limitations, such as requiring intermediary results to be stored in the data model.
AI systems are more likely to form biases when hiring than humans due to their tendency to generalize from limited data. Researchers found that large language models (LLMs) can stereotype job applicants by demographic group, even more so than human participants in similar studies. The models' bias was reduced when they were given a bonus for diverse hiring and when they received personal information relevant to the individual's ability to adapt to a new role. However, the finding suggests that AI systems may learn novel biases from their experience and are not always capable of avoiding existing social biases.
Weather data sabotage is a growing risk due to the increasing importance of weather forecasts in various industries, such as agriculture, utilities, and prediction markets. Manipulation of single weather stations, like the Paris Charles de Gaulle Airport incident, can be caught by human monitoring or statistical methods. However, coordinated manipulation of multiple stations at once poses a greater challenge, introducing new risks for data quality and accuracy. The shift towards artificial intelligence in weather prediction amplifies these risks, requiring strengthened oversight and accountability structures to protect observational data and models.
Bristol Myers Squibb is deploying its second NVIDIA DGX SuperPOD, a life science industry's most advanced AI factory on NVIDIA Vera Rubin. The new system uses eight rack-scale systems, each comprising NVIDIA Vera CPUs and Rubin GPUs, delivering up to 10x the performance per megawatt of the previous infrastructure. This will enable researchers at BMS to access a unified AI platform for running predictions, training models, and powering agentic workflows across the full drug discovery pipeline.
As AI spending continues to rise, companies are facing challenges with opaque pricing models that do not clearly disclose costs. These unclear costs can lead to backward-looking billing, where expenses are only tracked after a project has been completed. As a result, organizations are re-evaluating their AI model strategies in response to these financial complexities.
A community developer has published a 1 billion-parameter model called MiniCPM5-1B-Claude-Opus-Fable5-Thinking, built on OpenBMB's MiniCPM5-1B base model and fine-tuned on Fable 5 data to improve coding and instruction following. The model runs fully on local hardware without API keys or cloud calls, with a footprint of around 657MB in its smallest quantization. It uses supervised fine-tuning rather than weight-level distillation, capturing response format and style but not frontier reasoning capability.
Qwen3.6-27B is a dense 27B model released in April 2026 under Apache 2.0, suitable for agentic coding, repository-level reasoning, and frontend workflows, requiring about 16GB of memory at Q4_K_M. The Qwen3.6-35B-A3B is a Mixture-of-Experts model with 35B total parameters, decoding faster than a dense 35B model due to fractional weight firing per step, but still tracking total parameters for memory consumption. Gemma 4 is a multimodal and multilingual model released under Apache 2.0, suitable for vision input and language coverage, requiring about 14GB of memory at Q4_K_M.
Feyn AI has released SQRL, a text-to-SQL model family that inspects the database before writing a query. The flagship model, SQRL-35B-A3B, reaches 70.6% execution accuracy on the BIRD Dev benchmark, outperforming other models such as Claude Opus 4.6 at 68.77%. SQRL combines single-shot and pipeline strategies, paying for inspection only when necessary.
Alibaba's Qwen team has previewed Qwen3.8-Max-Preview, a 2.4 trillion-parameter multimodal model, days after Moonshot AI released Kimi K3. The model is described as "second only to Fable 5" among systems benchmarked, but benchmarks and active-parameter counts are not verified. The model can process text, images, video, and documents and will be made open-weight in the future. A preview build is available through Alibaba's Token Plan subscription at 10% of standard pricing.
Scientists warn that AI-altered images on birdwatching forums are putting research at risk, as fake and enhanced images threaten the credibility of popular citizen science platforms like iNaturalist and Macaulay Library. Experts say many users unknowingly use AI to edit photos, which can introduce significant changes and create false sightings. Hundreds of fake images have already been discovered on popular databases for recording species, contaminating records collected by the public. Citizen science organizations are working to understand the scale of the issue, urging users to be vigilant when sharing information online.
OpenAI plans to create a language model with the capability of GPT-3 that can run locally on consumer hardware and release it soon, potentially discouraging others from releasing similarly-powerful models. The goal is to make it harder for new efforts to get funded. This strategy was discussed in an email to OpenAI's board on October 1, 2022.
Here is a summary of the article in under 5 sentences: To train a neural network, it needs to learn by figuring out which parameters to increase and decrease to reduce the loss. The learning process involves calculating partial derivatives of the loss function with respect to each parameter using the chain rule. This derivative tells us how sensitive the loss is to each parameter and helps determine whether that parameter should be increased or decreased to reduce the loss. In a neural network, the output layer has multiple parameters that need to be optimized simultaneously, making manual derivation of gradients challenging.
Loop engineering, specifically in question parsing, involves designing small loops that improve pipeline quality by filling missing fields before retrieval. A small loop consists of one field being filled from a fixed schema. The flow has six steps, with the loop only firing from step 3 to step 4. This approach allows for bounded latency and is grounded in real state, unlike agentic RAG where the loop spins for tens of seconds.
An AI system develops a strategy that appears strange or bad to humans, who then study it and adapt their behavior, causing the AI to adapt to a changing human population. This ongoing loop, known as synthetic counteradaptation, is different from one-off knowledge transfers because it doesn't close, but rather continues as both sides evolve. The phenomenon has implications for multi-agent RL, opponent pool design, curriculum development, and understanding of human baselines in AI systems. It highlights the non-stationarity of humans in response to AI strategies, making traditional evaluation setups inadequate.
Researchers at Ant Group developed "Ring-Zero," an algorithm that scales zero RL (Reinforcement Learning) to trillion parameters, enabling emergent reasoning in complex models. Ring-Zero achieves this through a novel method of parameter optimization and regularization techniques. This advancement aims to bridge the scalability gap between existing methods and large-scale models.
A researcher created a Soft Actor-Critic (SAC) reinforcement learning agent that trades US stocks using features derived from OHLCV, VWAP, and transactions data. The model performed reasonably well, with +12% returns across three independently seeded runs on a held-out test window, but did not beat the buy-and-hold strategy. It achieved under half the return of doing nothing for an identical drawdown.
Staying connected with others, cultivating uniquely human skills such as curiosity, humility, and emotional intelligence, can help mitigate the fear of becoming obsolete in a job market where AI is increasingly prevalent. This approach is recommended by experts who suggest that workers view AI developments as an opportunity for growth rather than a threat. By prioritizing human skills and networking with others, individuals like Paul can prepare actively for a future where AI assumes more technical tasks.
GraphDx, a cost-aware knowledge-enhanced multi-agent framework, improves diagnostic success rates from 50--68% to 79--93% and reduces test costs by 20--54%, providing a robust, economical, and interpretable solution for automated clinical diagnosis using Large Language Models (LLMs). GraphDx leverages an automated pipeline to construct Medical Diagnosis Knowledge Graphs with quantized typicality, action-centric topology, and dual-objective attributes. Three collaborative agents handle language understanding, generation, evidence scoring, and cost-aware planning on the MDKG.
Causal-Audit proposes an explicit and auditable graph-based reasoning framework for context-free intervention-based question answering. It formulates causal inference as structured reasoning over a causal graph through four modular stages, including target-aware causal graph construction and path-level causal evidence aggregation. The method outperforms existing LLM-based methods while providing interpretable and auditable causal reasoning traces. Extensive experiments demonstrate its consistency on three benchmarks.
Cura 1T is a healthcare-specialized large language model (LLM) trained through a human-gated self-evolution loop. This approach involves training an agent to plan target capabilities, refine data mixture from observed failures, and improve the model through targeted synthetic and curated examples. Cura 1T ranks at or near the top among frontier baselines in healthcare evaluation suites, while remaining competitive on out-of-domain reasoning and agentic benchmarks.
AnovaX is a local, multi-agent voice assistant that runs entirely on the user's computer and treats the desktop as its action surface. It uses a single Python process to wire together various components, including an LLM planner, typed executors, and adaptive recovery mechanisms. The assistant can coordinate concurrent actions, recover from single-step failures, and be controlled entirely from a phone in another room without exposing any raw audio or touching the keyboard.
Reviewer precision alone does not guarantee critique uptake in multi-agent math reasoning. A protocol may spot errors well yet still fail to solve more problems if it does not act on those critiques. Forcing explicit acknowledgment lowers final accuracy, while embedding reviewer guidance directly in the solver's working context partially improves follow-through without closing the gap.
DrawingVQA, a benchmark for multimodal large language models (LLMs), evaluates LLMs on real-world construction drawings. The benchmark includes 33 drawings and 92 expertly curated question-answer pairs across three reasoning depths: perceptual understanding, contextual interpretation, and domain-expert reasoning. Evaluations reveal a substantial gap between model performance and that of experts in the field, particularly at higher reasoning depths.
Sergey Rodionov investigates the necessity of executable world models, simplification, and verification in coding agents to solve ARC-AGI-3. Four nested Codex-based agents were developed to evaluate these components: a textual baseline, an executable model without replay verification, an executable model with scheduled simplification, and a fixed-interface verification treatment. The most robust result is that every agent variant improves with a stronger model and greater reasoning effort, but requiring a persistent executable deliverable is not universally beneficial.
A novel framework, MAR-12, is introduced to detect and explain harmful humor in memes by leveraging Vision Language Models (VLMs). It interprets memes through twelve structured perspectives derived from humor and hate theories, applies a role-aware soft-gated attention mechanism, and synthesizes explanations using perspective-specific reasoning and learned attention weights. MAR-12 achieves up to 80.3% accuracy for humor detection and 75.9% accuracy for hate detection on the PrideMM and Memotion datasets, outperforming state-of-the-art approaches.
Researchers have developed a three-stage post-hoc transformation to convert a trained deep reinforcement learning policy into an executable Prolog program that reproduces its behavior and can be executed by a logic engine. The transformation extracts a frozen proximal policy optimization teacher, induces an ordered rule list from the teacher's decisions, and emits the result as a Prolog program. Empirical results show that the converted Prolog program attains exact optimal return in every seed on a two-room key-and-door task, while recovering partially or exceeding the stochastic neural teacher's performance on three continuous-control tasks.
Critical analysis of trustworthy AI tools and mark frameworks reveals significant asymmetries in ethical focus, lifecycle coverage, stakeholder targeting, and tool typology. Most tools concentrate on post-development stages with limited guidance for early design or data collection phases. The study argues that bridging implementation gaps requires expanding ethical objectives, embedding ethics across the AI lifecycle, and fostering broader multi-stakeholder participation.
Dyadic and circular convolution can be computed in O(N log N) time using the Hadamard transform and the FFT-computed discrete Fourier transform, respectively. The substitution for the DFT introduces an algebraic error that asymptotically doubles the output energy, except for filters in the universal zero-error subspace, which incur no error.
Quantum program generation must prioritize validity over probabilistic scaling, as the scaling hypothesis assumes that increasing model parameters yields emergent reasoning capabilities, but this approach ignores the strict adherence to mathematical constraints in quantum circuits. Training on unverified programs leads to models learning syntax without capturing physical semantics, and post-hoc filtering is mathematically intractable due to an exponential decay of valid circuit designs with qubits. A pivot towards verifier-centric agents that integrate hierarchical constraints and symbolic proxies can offer a viable path for modular quantum program generation.
A transportable threshold-based framework provides fully interpretable, rule-based clinical classification using the Bernoulli Naïve Bayes (BNB) model. The method applies supervised χ2-guided statistical binarization to continuous variables, identifying thresholds that maximize association with clinical outcomes within training data. The approach was evaluated on three benchmark datasets, achieving area-under-the-curve (AUC) scores of 0.800 for Pima Indians Diabetes, 0.984 for Wisconsin Breast Cancer, and 0.919 for Heart Failure Prediction.
This paper proposes a method, MoRe, to improve the convergence rate of stochastic multi-objective regularity-aware MGDA in nonconvex settings. The traditional vanilla stochastic MGDA method has a suboptimal convergence rate due to noise in gradients from mini-batch sampling. The proposed method exploits the Lipschitz continuity of the CA direction when the subproblem is regular and switches to linear scalarization update otherwise.
Large Language Models (LLMs) have emerged as powerful tools for processing the heterogeneous information environments of modern financial markets. A systematic evaluation of five prominent LLMs, including GPT-4 Turbo and FinGPT, was conducted to assess their capacity for technical market analysis. The models outperformed a passive S&P 500 benchmark in backtesting, with GPT-4 Turbo achieving the highest annualized return and Sharpe ratio among general-purpose models. However, persistent failure modes were identified across all evaluated models, including numerical hallucination and context-window limitations.
qZACH-ViT, a quantization-aware extension of the zero-token ZACH-ViT backbone, achieves significant improvements in efficiency and interpretability by introducing Recursive Attribution-Stabilized Optimization (RASO). Deployed mixed-precision INT8 qZACH-ViT with Adam improves the FP32 ZACH-ViT baseline mean on all seven datasets. RASO reduces sufficiency error and improves input-noise stability over Adam, but doesn't dominate every predictive or explainable artificial intelligence metric.
A standard linear model and a single-qubit mixed-state model for binary classification both learn hyperellipsoids to classify data, rather than hyperplanes. The geometric inductive bias of each model leads to different feature importance inductive biases. This characterization offers an accessible route into quantum machine learning for readers unfamiliar with quantum concepts.
Stochastic Reset Pathfinding (SRP) is an episodic learning problem on a known directed graph with unknown stationary edge success probabilities. The optimal policy for SRP is open-loop, placing it within the combinatorial cascading bandit framework. A Log-Dijkstra meta-algorithm with UCB and Thompson Sampling instantiations achieves path-level regret bounds that are complementary to edge-level bounds, supporting experiments on various domains.
This study analyzed Medical Expenditure Panel Survey (MEPS) data from 2019 and 2021 to examine healthcare financial vulnerability before and after the COVID-19 pandemic. High financial burden was defined as out-of-pocket healthcare expenditures exceeding 10% of family income, and subgroup analyses indicated persistent disparities across population groups. The study found that models trained on pre-pandemic data exhibited only modest reductions in predictive performance when evaluated on post-pandemic data, suggesting that the principal predictors of healthcare financial vulnerability remained relatively stable over time.
Researchers have proposed LLM4EHR, a clinical foundation model trained on ICU Electronic Health Records (EHR) data to align clinical time series with medical event sequences via large language models. The model combines domain-adapted large language models with a transformer time series encoder and learns robust EHR time series representations conditioned on EHR event embeddings. A regularised contrastive objective is used to pre-train the model, resulting in improved performance on downstream clinical tasks and transferable clinical time series embeddings that can be deployed to new cohorts via k-shot adaptation.
Researchers use automated classifiers to label unstructured data, but existing methods treat audit labels as correct. A new method, Partially Adjudicated Design-Based Supervised Learning (PA-DSL), uses adjudicated cases to correct noisy human labels and then debias analyses using full sets of automated labels. PA-DSL maintains nominal coverage and reduces RMSE by 10-17% relative to using only adjudicated labels in synthetic and Wikipedia Detox experiments.
Retraining seeks stable signals, where repeated risk minimization converges geometrically to the direction of a small model-independent component, or "stable signal," in the prediction target. This stable signal principle governs retraining even when the model's influence on the target is arbitrarily large relative to the stable signal. Regularization emerges naturally as a force to control performativity, revealing a new facet of an old concept.
A game-theoretic framework is presented to analyze the influence of hyperparameters on machine learning objectives, employing Shapley Effects for global sensitivity analysis and Pareto front sets to identify effective configurations. The proposed framework provides interpretable insights into objective-aware hyperparameter interactions, enabling practitioners to optimize models, reduce search space, and perform early-stage evaluation. This approach is demonstrated using three neural network architectures across different problem domains under multi-objective settings.
Deep and Probabilistic Models for Gene Regulatory Network Inference casts GRN inference as a probabilistic graphical model optimized by variational inference, enabling principled model selection and uncertainty-aware edge estimates. GLM-Prior addresses the prior bottleneck by fine-tuning pretrained Nucleotide Transformer to predict TF-target gene interactions directly from nucleotide sequence, generalizing across yeast, mouse, and human settings. The two frameworks motivate a dual-stage view of GRN reconstruction in which sequence-derived priors provide a transferable starting scaffold and probabilistic inference refines regulatory estimates with quantified uncertainty under incomplete evaluation resources.
Cluster-Aware Matching via Laplacian Optimal Transport proposes a novel approach to matching point clouds with an intrinsic cluster structure. The key idea is to regularize the optimal transport problem with quadratic Laplacian terms, which encourages respect for the cluster structure of both point sets. This leads to more consistent and meaningful alignments between point clouds through Refined Simultaneous Clustering (RSC).
Proactive inpatient bed requests can reduce average boarding times for admitted patients by 30-70% and average length of stay for all ED patients by 6-15%. These predictions are made using information about current patients and bed availability, guiding early inpatient bed requests. Policies based on approximate dynamic programming, reinforcement learning, and a newsvendor-type approach were evaluated across a wide range of settings. The newsvendor heuristic offers the most attractive tradeoff between ED performance and inpatient bed idle time.
Researchers Hien Dang, Pratik Patil, and Alessandro Rinaldo introduced a new method for self-distillation (SD) in linear and logistic regression, where the student is trained on fresh unlabeled covariates pseudo-labeled by the teacher. They showed that this "prediction-only" distillation scheme results in risk strictly smaller than the teacher's risk under general anisotropic covariance.
Diffusion models can recover accurate mixture weights despite score function insensitivity. This is because generated samples can be informative about the mixture weights, even when the target score is insensitive to them. The diffusion score sensitivity index (DSSI) measures the variation in the DSM loss relative to changes in a parameter and governs the accuracy with which the parameter of the target distribution can be estimated from generated samples.
The original Binary Iterative Hard Thresholding (BIHT) algorithm converges to a sparse vector in the noiseless setting with $\widetilde O(s/\epsilon)$ measurements, matching the sample-optimal convergence rate achieved by its normalized variant. However, under sign corruptions, BIHT without per-iterate normalization still reaches optimal recovery at an early iterate but is unstable.
Conventional tabular evaluation methods fail to capture the failures of generative models, which can reproduce every marginal and foreign-key relationship while emitting timestamps that run backwards or repeat. A taxonomy-guided evaluation protocol assesses temporal fidelity by measuring timestamp validity, cross-sectional structure at aligned time points, within-entity dynamics, and time-varying relational structure across eight generative models on thirteen datasets spanning six domains. The protocol reveals substantial disagreements between conventional and temporal evaluations, highlighting the need to measure temporal fidelity on the time axis itself rather than inferred from pooled record distributions.