Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, 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 explores federated learning (FL) for continuous client onboarding without catastrophic forgetting, using hypernetworks and data-free replay.
ABot-World-0, an action-conditioned video world model, enables real-time, long-horizon interactive world rollout on a single desktop GPU.
Research introduces E2E-CDiff, a conditional diffusion model generating realistic, controllable visual traffic scenarios for autonomous driving evaluation.
Research explores using algebraic signatures derived from observed vanishing binomials in probability tensors to identify probabilistic model structure.
Research details BETA, an indirect evasion attack targeting Graph Neural Network-based anomaly detection in sensor networks by perturbing non-target sensors.
Research introduces a method for exact recovery of Gaussian graphical models from a single trajectory of Gaussian Glauber dynamics.
Research explores using LLMs for agentic calibration of grey-box simulation models, addressing high-dimensional parameter spaces and expensive evaluations.
Research benchmarks deep learning models for layout detection and information extraction specifically on Architecture, Engineering, and Construction (AEC) drawings.
Research details the exact error conditions required for tractable, unbiased sampling with inexact score oracles, ruling out weaker error assumptions.
Research introduces SkewAdam, an optimizer that allocates state differently for MoE models to reduce memory usage during training.
Research paper proposes a two-stage online learning framework for detecting service-affecting failures in mobile core networks using traffic data.
Researchers introduced E-SpecFormer, an Edge-Efficient Transformer for end-to-end RF spectrum monitoring, using a Softmax- and LayerNorm-free attention mechanism called LiTAN.
New research addresses quantifying the risk of rare failures in language models, proposing improved methods for low probability estimation to avoid bias.
A new benchmark, HALLMARK, diagnoses three failure modes in LLM citation verifiers, addressing high rates of hallucinated references in academic papers.
New research details cost accounting for exhaustive site-by-site interventions on neural network computational graphs, focusing on recomputation costs.
RAPT is a new model-predictive out-of-distribution detection method for preventing silent failures and hardware damage in Sim-to-Real deployments.
Research compares Monte Carlo Dropout and Deep Ensemble for uncertainty quantification in AI-driven crash simulation surrogates, using an open-source bumper beam.
New research proposes Chi-Square Wavelet Graph Neural Networks for anomaly detection in complex, heterogeneous networks, addressing GNN limitations.
Research details Quantum Reservoir Computing (QRC), using quantum system dynamics for temporal input transformation and classical readout training.
MAAT framework reconstructs latent dynamical states from partial, noisy, and heterogeneous observations using a knowledge-informed kernel approach.
Research finds aggregating judgments from multiple LLMs can outperform individual models but risks 'contamination' from poor performers.
Research introduces a scalable spiking embedding predictive architecture for dynamic graphs, improving fraud detection and recommender systems.
Neural Kolmogorov Equations (NKEs) are proposed for learning stochastic dynamics with parallelizable training and support for general noise types.
Research explores network-agnostic feature initialization for Graph Neural Networks to improve spatial transferability in traffic assignment models.
Research introduces a decision-aware weak-to-strong (W2S) learning framework to improve predictive models using both labeled and unlabeled data.
New research argues that 'circuit extraction' methods for identifying AI model mechanisms are under-determined, with findings depending on reporting and comparison choices.
Researchers propose SFGA, a statistics-first gating architecture for evaluating the quality of supervised fine-tuning (SFT) data based on diversity, utility, and redundancy.
New research introduces hierarchical physics-embedded adaptive Fourier neural operator for spatiotemporal systems with partial physical knowledge.
Research presents a generative state-space model learning directly from sparse, noisy observations to improve ocean modeling beyond complete datasets.
Research explores Parallel Noising in Neural Markov Logic Networks (NMLNs) to improve generative modeling performance on larger relational structures.
New research proposes an optimization stack for Reinforcement Learning with Verifiable Rewards (RLVR), focusing on reusing base model weight spectra.
Research explores decentralized multi-agent reinforcement learning (MARL) for improving resilience in critical infrastructures.
New research introduces Posterior Prefix Tuning (PPT), a method to steer transformer behavior by optimizing the latent posterior without backpropagation.
Research introduces Hybrid Latent-Structural Fusion (HLSF), combining tensor decomposition and normalizing flows for cyber anomaly detection.
Research introduces GNNAS-TSP, a Graph Neural Network (GNN)-based framework for automated algorithm selection (AS) in the Traveling Salesman Problem.
Researchers propose physical self-supervised learning for IMU-based sensing, reducing reliance on labeled data and improving robustness to heterogeneity.
New research introduces AHEAD, a multi-class label aggregation method using interpretable cross-annotator modeling to infer true labels from noisy crowdsourced data.
A research paper systematically compares two geospatial foundation models, TerraMind and THOR, developed under the European Space Agency's $\Phi$-lab.
Research finds common benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) overstate correctness by using fixed-shape, small-sample checks.
ConceptCF introduces a method for generating counterfactual explanations for time series models based on human-interpretable concepts.
Research proposes a censoring-aware in-context learning method for generalized supplier lead time estimation, addressing right-censored data.
Research introduces domain-conditional position offsets to reduce cold-start inaccuracy in autoregressive language models by adding a learned vector.
Research introduces Decafs, a flow-based generative model using Lie groups to disentangle latent spaces, improving interpretability and controlled generation.
Interactive Training 2 introduces an open-source control plane for auditable steering of live model training, abstracting trainer-specific code.
Research proposes a federated fine-tuning method that reduces communication bottleneck by averaging small trainable latents instead of full model weights.
Research introduces a unified framework for Riemannian deep learning with reusable modules and manifold-specific architectures.
QScheduler proposes adaptive gradient sampling for Zeroth-Order (ZO) on-device training on INT8 NPUs, reducing memory needs.
Research explores Wireless Physical Neural Networks (WPNNs) that embed neural computation in analog hardware using MIMO relays and power amplifiers.
Research proposes a multi-view classification method using global anchor consensus to improve robustness under noisy data supervision.
FALCON-Discover proposes a model-agnostic framework to find concentrated regions of false confidence in ML predictions, improving local calibration.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion