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.
PsiLogic introduces an optimizer that enhances Adam with a dynamic Active Cancellation Term to improve stability during chaotic training phases for deep learning models.
A research paper proposes a predict-then-correct framework using few-shot continuous contextual bandits for adaptive retail demand forecasting.
Research indicates constant-stepsize SGD's behavior fundamentally changes for convex objectives with flat minima, deviating from Gaussian limits.
Research explores training-free interpretability methods for neural networks, questioning if expensive training-based methods yield superior insights.
Research explores compact convolutional neural networks for detecting drones via their radio-frequency video transmissions, focusing on lightweight models.
K-IPO, a new oversampling method, addresses class imbalance in tabular data while preserving feature importance rankings for model explanations.
Research on multi-task deep learning for predictive maintenance highlights evaluation robustness challenges, specifically data leakage in time-series splits.
Research explores how team-averaged rewards in Multi-Agent Reinforcement Learning (MARL) affect learned representations, proposing new metrics.
Research explores hierarchical domain generalization, showing how train/test domain partitions, not just hypothesis class complexity, obstruct extrapolation.
Researchers introduced Building2Building (B2B), a large-scale benchmark for generalizable real-world reinforcement learning.
Research indicates model uncertainty metrics in model-based reinforcement learning (RL) are not reliable risk signals and can increase collision rates.
Research proposes a new group fairness criterion for differentially private machine learning, considering privacy cost itself as a form of harm.
Research applies self-supervised learning (JEPA, VICReg) and federated learning to predict sepsis from electronic health records, addressing data challenges.
Research demonstrates pruning Mixture-of-Experts (MoE) LLMs by up to half their experts for coding tasks without performance degradation.
New research proposes a general framework for designing robust loss functions using univariate base functions to mitigate noisy labels in deep neural networks.
MultiLoReFT, a new fine-tuning method, aims to improve multimodal learning by disentangling shared and modality-specific information, addressing data scarcity.
New research proposes a concordance loss function for deep survival models, aligning training objectives with the C-index evaluation metric.
New research proposes a method for sparse-support uncertainty quantification, addressing issues of overconfident conclusions when dictionaries are learned from latent mixtures.
Research explores local task adaptation in trained transformers, noting individual perturbations are predictable but pairwise interactions are fragile.
New research introduces Public AI Memory (PAIM) and Public Transformers (PTr), benchmarks for evaluating LLM memory systems on scientific papers.
New arXiv research explores intrinsic motivation in unsupervised reinforcement learning using model-free epistemic free-energy estimators.
Research proposes a physics-informed learning framework for robust health monitoring of lithium-ion batteries, integrating aging origins.
Research models real-time truckload bid acceptance as a dynamic program, optimizing fleet state, hours-of-service, and appointment windows.
Research introduces TVGL-CFM, a model for generating and forecasting time-varying trajectories of dynamic networks using conditional flow matching.
Research explores safe adaptive control, enabling online learning with guaranteed safety across varied initial models while matching a safe oracle's performance.
Research introduces CADENCE, a distillation method addressing cold-start collapse, state-agnostic scheduling, and sparse rewards in on-policy knowledge transfer.
Researchers introduced SurvCF(t), a framework to generate counterfactual explanations for deep survival models used in predictive maintenance.
TurboVec, an open-source vector index, addresses data leakage and recall degradation in multi-tenant RAG systems via codebook-oblivious quantization.
Periodic Bootstrap Thompson Sampling (PBTS) is a new algorithm for bandit problems, addressing bias from non-stationary reward distributions.
Research finds a specific regularization method (\sigreg) can reduce hidden-state anisotropy in autoregressive LLMs, which may improve KV-cache quantization.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion