Search signals, briefings, benchmarks and glossary terms.
Search signals, briefings, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research analyzes the convergence speed of Low-rank Adaptation (LoRA) for fine-tuning large models, finding exponential oracle calls for $\epsilon$-stationary points.
A new research study investigates scaling laws for classical machine learning models on tabular data across 18 datasets and 6 model families.
Research explores Cross-Domain Off-Policy Evaluation and Learning (OPE/L) for contextual bandits to address few-shot data and new actions in real systems.
Researchers introduced VRDQ, a variance-reduced distributed Q-learning algorithm for multi-agent reinforcement learning over static and dynamic networks.
Research proposes detecting talking-face deepfakes by analyzing physiological signals like remote photoplethysmography (rPPG), which are absent in synthetic video.
Researchers propose DCS, a unified framework for detecting cross-modal copyright infringement in foundation model outputs by analyzing conditional sensitivity.
New research proposes Sharpness-Guided Equilibrium Sampling to address poor generalization in long-tailed learning by combining re-sampling with geometry-aware techniques.
Researchers introduced CEL, a new library and benchmark for evaluating counterfactual explanations in explainable AI, focusing on properties like actionability.
A research paper proposes Deep Convolutional Large-Margin $\ell_p$-SVDD for visual anomaly detection, combining deep features with explicit margin-aware boundaries.
Research proposes a "Reasoning Denoiser" to improve hallucination detection in large reasoning models by identifying and removing irrelevant or repetitive steps in reasoning traces.
Research paper proposes RED-PIM, a Processing-In-Memory (PIM) architecture to reduce data movement during transformer attention operations, improving efficiency.
Researchers propose Evaluation-as-a-Service (EaaS), a cloud-native microservices architecture for scalable AI monitoring with conformal guarantees.
Research introduces a supervisory runtime stability framework for neural network training to detect and recover from severe destabilizing updates.
IFCLoRA is a new parameter-efficient fine-tuning method for LLMs that optimizes rank allocation across Transformer modules without extra memory or computation.
Research characterizes self-training (ST) in linear classifiers using pseudo-labels on Gaussian mixture data to understand generalization improvement.
Research proposes a robust predict-then-optimize method addressing prediction shifts from noisy covariate features in decision-making.
Research identifies common metrics for synthetic tabular data generation are blind to inter-column dependencies critical for fraud and risk models.
Research identifies optimization collapse and topology as causes preventing Logic Gate Networks (LGNs) from reliably benefiting from increased depth.
Research explores multi-horizon latent consistency in video predictors, analyzing how the weighting of multi-step agreement affects prediction error.
Research introduces Neural Atom Prevalence (NAP), a Bayesian framework for structured node-level model selection in feedforward neural networks.
CARNet proposes a novel linear-complexity model for multivariate time series forecasting that addresses cross-variate dependencies and periodic patterns.
Research introduces a parameter-free adaptive sparse attention method using data compression, outperforming fixed patterns and dense attention.
Research explores quantum federated learning to enable distributed quantum neural network training without sharing sensitive local data for intelligent services.
Research proposes a method to evaluate the causal impact of ML-assisted decision-making using counterfactual correctness without full RCTs.
Research introduces 'adjustment speed' as a safety constraint for reinforcement learning in nonstationary environments, addressing delayed adaptation risks.
Research demonstrates Quasi-Monte Carlo (QMC) initialization improves training convergence in meta-reinforcement learning, outperforming orthogonal defaults.
Research proposes MA-DAR, a method for continual temporal knowledge graph reasoning that integrates new facts while preserving old knowledge.
Research identifies numerical fragility in Transformer models due to low-precision execution, proposing a layer-wise risk estimator and controller.
New research addresses scaling Graph Neural Networks (GNNs) on heterophilic graphs, where prior coarsening-based training methods struggled.
Research overviews Bayesian and frequentist simulation-based inference with machine learning for inverse problems and parameter estimation.
Research finds the simple quadratic model can effectively predict optimization dynamics in a 150M parameter LLM, challenging assumptions of neural network complexity.
RIS-Kernel introduces a model-agnostic architecture, RIS, reducing LLM self-attention complexity to O(N log N) for long-context inference.
Research benchmarks federated learning strategies for in-hospital mortality prediction using heterogeneous and imbalanced clinical data.
Research explores safety in In-Context Reinforcement Learning (ICRL), addressing unexamined test-time behavior for real-world deployments.
Research proves ReLU networks with two hidden layers can exactly represent the maximum of up to 10 real numbers using rational linear algebra.
Research explores physically constrained federated additive models (FAMs) for auditable and privacy-preserving SLA-risk prediction in O-RAN networks.
Research identifies explicit iteration complexity for exact data-driven inverse optimization of Integer Linear Programs using gradient-based methods.
Hopformer, a two-stage Transformer framework, is introduced for multi-variate time series forecasting by separating common trends from specific information.
Research explores agentic AI for automated, evidence-grounded root cause analysis of industrial anomalies, addressing explainability and data scarcity.
Researchers propose SEM-DNN, a heteroscedastic neural simultaneous-equation estimator, to learn bidirectional causal interactions from observational data.
Research proposes a convex optimization framework to generate theoretical correlation matrices with graph-based sparsity patterns, improving matrix completion.
Research explores making large text-to-image diffusion models more interpretable and manipulable for creative uses, focusing on interactive explainability.
Research proposes TRACE-ROUTER, a new routing mechanism for agentic AI applications that optimizes LLM selection based on long-horizon, task-level outcomes.
HiKV proposes a novel algorithm-hardware co-design to compress the KV cache in LLM decoding, tackling memory bottlenecks for long-context models.
Research proposes General Value Functions (GVF) for remaining useful life (RUL) and failure-mode prediction in predictive maintenance.
Research explores a Spatially-Enhanced Temporal Fusion Transformer for interpretable multi-output prediction in parametric dynamical systems with time-varying inputs.
Research proposes contract-based incentives for federated learning (FL) to prioritize high-quality data contributions during critical early training periods.
Research introduces SurvDiff, a diffusion model designed for generating synthetic survival data, addressing challenges of incomplete event information.
Research proposes Generalized Gaussian Temporal Difference Error (GGD-TDE) for uncertainty-aware reinforcement learning, addressing non-Gaussian TD residuals.
Researchers introduced LiMuon, an optimizer designed for large model training, claiming reduced sample complexity and memory usage compared to prior Muon variants.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion