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.
TEmBed-T introduces a new multi-dimensional benchmark for evaluating table-level embeddings, crucial for applications like table retrieval and data lake discovery.
Agent-UCT introduces an algorithm leveraging Upper Confidence Bounds Applied to Trees for cost-aware optimization of agentic workflows like RAG.
Research highlights the lack of systematic, temporally valid, and adversarial-robust evaluations for AI-based Windows malware detectors.
Research establishes minimax lower bounds for kernel discrepancy estimation (MMD, HSIC, KSD), relevant for two-sample and goodness-of-fit testing.
Research characterizes trade-offs between LLM jailbreak defenses, performance, over-refusal, and inference costs, categorizing defenses by operational strategy.
Research proposes MTSF-ANO, a hybrid quantum-classical model using adaptive non-local observables for multivariate time series forecasting.
Research paper introduces proxymate, a method to diagnose and adjust for systematic bias in proxy-based estimates used in place of primary outcomes.
Researchers propose ESRVS, a semi-supervised method for retinal vessel segmentation requiring only one annotated image and unlabeled data.
Research paper proposes a model for aggregating imbalanced crowd-sourced labels, focusing on class-dependent annotator accuracy, especially for rare classes.
Research paper argues for evaluating AI-driven autonomous research (AR) systems based on solution-search efficiency, not just final outcome quality.
Research explores efficient expert design for diffusion transformers, aiming to make AIGC foundation models scale more practically.
Research on co-learning addresses missing data modalities in multi-modal classification, focusing on robust fusion despite operational constraints.
Research explores how on-policy distillation behaves under classifier-free guidance, identifying issues with existing methods in adapting diffusion models.
Research explores learning an unknown distribution from multiple heterogeneous data providers, querying conditional samples from restricted sets.
Research examines fairness interventions in AI classification, comparing Demographic Parity and Equalized Odds criteria for explainability.
Research explores novel loss function designs for Generative Flow Networks (GFlowNets) to enhance training stability and sampling from unnormalized distributions.
Researchers propose CausAdv, a causal reasoning framework for detecting adversarial examples in Convolutional Neural Networks.
Research on Fisher Information based Stochastic Gradient Ascent for online learning of Dirichlet Process Mixture, improving scalable Bayesian nonparametrics.
A new arXiv survey reviews Graph Transformers (GTs), detailing architectures, theories, and applications, addressing GNN limitations like over-smoothing.
Researchers propose EvoCL, a gradient-free evolutionary algorithm for continual learning in neural networks to prevent catastrophic forgetting without storing past data.
Research explores CTC-based knowledge distillation for Automatic Speech Recognition (ASR) models, focusing on blank token handling to improve efficiency.
Research proposes synthetic benchmarking to systematically characterize sequence modeling architectures like RNNs, Transformers, and state-space models.
Research evaluates accuracy of diffusion models for inverse problems, such as inpainting and super-resolution, especially with Gaussian data.
Loong introduces a method for synthesizing long chain-of-thoughts using verifiers to improve LLM reasoning, particularly in domains like mathematics and programming, extending RLVR.
Research proposes CHARM, a graph-based method using attention flows and activations to detect hallucinations in Large Language Models.
Research introduces "Seesaw," a principled framework for scheduling batch size and learning rate to accelerate large language model pretraining.
Research proposes a LoRA-based approach for Domain Incremental Learning to prevent catastrophic forgetting by consolidating shared knowledge across tasks.
Research explores energy trade-offs between on-device processing and cloud streaming for multi-modal deep learning on wearable cardiovascular patches.
Research explores infinite-precision autoregressive modeling for vector graphics and layouts, addressing token discretization limits in continuous domains.
New research proposes an extended score matching framework for causal discovery to learn DAG structures from purely observational data, improving upon existing methods.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion