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 finds quantum reservoir computing performance parity with carefully tuned classical methods, challenging prior claims of quantum advantage.
GATAS introduces a black-box testing method for ASR systems, generating phoneme-level adversarial inputs to induce failures while maintaining naturalness.
Research explores Task-Conditioned Synthetic Data Generation (SDG) to enhance machine learning performance, addressing data scarcity in agricultural tasks.
Researchers propose ECG-LDC, a low-dimensional computing framework for energy-efficient ECG arrhythmia classification on resource-constrained hardware.
Research explores 'Requential Coding,' a method using self-generated training data for model compression, suggesting efficiency gains.
New research proposes Active Offline-to-Online Reinforcement Learning (O2O-RL) for training and improving policies using historical data and limited online interaction.
DiffEEG, a 9.6M-parameter self-supervised diffusion model, is proposed for EEG seizure detection, addressing data scarcity and class imbalance.
Research introduces a new method for anytime-valid and feature-aware auditing of black-box conditional quantile forecasters, addressing data drift challenges.
Research finds common post-hoc calibration methods for probabilistic classifiers, Temperature Scaling and Isotonic Regression, lack robustness across varied operating conditions.
Researchers propose CDFM, a 'Causal Discovery Foundation Model' approach for recovering causal structures from diverse observational data.
Researchers propose PRISM Edit, a new LLM editing method allowing temporal facts to be updated while retaining historical accuracy without full retraining.
Research introduces PolicySynth, a framework for generating synthetic data that ensures decision alignment for campaign decision support, beyond just distributional similarity.
Research proposes ToolAtlas, a system for LLM agents where tool knowledge and memory are maintained by the tool provider, enabling reuse across agents.
Research explores a new training priority weighting framework for physics-informed neural networks (PINNs) to improve stability and accuracy.
Research evaluates TabPFN as a zero-gradient classification head for few-shot multimodal classification, improving calibration and accuracy.
New research proposes a spectral method to identify heterogeneous treatment effects under unobserved confounding, extending prior Synthetic Potential Outcomes work.
Research explores reliability of quantized LLMs on perturbed inputs, addressing a gap in standard performance metrics for robust deployment.
Research proposes Diachronic Sample Integration to improve generative models' ability to accurately estimate rare, adverse tail-risk scenarios.
Research paper introduces Hierarchical Bayesian Quadrature, an advanced numerical integration method using Gaussian process surrogates with quantified uncertainty.
Research details a "Policy-Driven CT-Agent" model for optimizing CT imaging phase selection to reduce radiation exposure and improve diagnosis.
Research introduces a 'structural-frontier split' for molecular property model evaluation, revealing traditional scaffold splits hide generalization failures.
Research demonstrates explainable machine learning (XML) pipelines can produce misleadingly robust risk hierarchies for composite outcomes.
New research proposes M+Adam, an optimizer addressing accuracy degradation in low-precision AI model training without high-precision copies.
MemDecay is a new KV cache eviction policy that improves LLM agent efficiency by prioritizing semantically important context over less relevant data.
Research uses LLMs to guide symbolic hypothesis generation for discovering partial differential equations from noisy data, improving robustness.
EvidentialRAG (ERAG) is a new RAG framework using evidential deep learning to quantify and mitigate information conflict in multi-source retrieval.
Research presents sharper convergence results for single-loop bilevel optimization methods, improving efficiency for hyperparameter and meta-learning.
Research introduces Delay State Space Models (DSSMs) to overcome fixed-size state limitations of traditional SSMs for long-context retention.
Research extends Neural Tangent Kernel (NTK) analysis to Physics-Informed Neural Networks (PINNs) for PDEs, focusing on NTK positivity.
Research introduces PhysMRV, a method to improve video-language models' (VLMs) physical plausibility reasoning, addressing limitations in object interaction and causal dynamics understanding.
FlashTrie is a GPU-accelerated constrained beam search method for generative retrieval, addressing CPU bottlenecks in trie-based decoding.
New research proposes 'conservation laws' for diffusion models to exactly characterize data-model cross-entropy via local information-theoretic derivatives.
Research finds that data imbalance can improve model generalization and robustness to spurious correlations in sufficiently capable models.
New research proposes an exponential-linear weight reparameterization technique to improve neural network optimization by better handling weight magnitudes.
Research proposes a foundation model for multimodal event sequences, aiming to unify predictive tasks in financial services with heterogeneous data.
Research presents Nonparametric Bayesian Inverse Reinforcement Learning with data-parallel Gibbs sampling to infer reward functions from diverse expert demonstrations.
Researchers propose Learned Predictive Ambiguity Sets (LPAS) for decision-focused distributionally robust optimization, enhancing protection against forecast errors.
Research proposes a formal safety argument for an AI Predictor (Scientist AI) designed for honesty and disinterest to avoid implicit agency in highly capable systems.
EvoClawBench is a new benchmark for evaluating whether AI agents can learn reusable skills from their own operational runs to improve future task executions.
Researchers introduced Embodied-R1.5, a unified Embodied Foundation Model (EFM) aiming for general physical intelligence by integrating reasoning.
Research identifies and mitigates early training collapse in deep neural network click-through rate (CTR) prediction models using industrial datasets.
Research explores minimax regret in contextual bilateral trade with infinite variance valuations, extending self-bounding properties to real-valued data.
MawForge proposes memory-bounded expert materialization for local Mixture-of-Experts (MoE) inference, enabling large models on memory-constrained devices.
Research compares Turbo-Quant and SpectralQuant KV-cache compression methods, identifying eigenbasis limitations on heavy-tailed data and successes in structured data.
Research explores optimizing decision trees and set cover problems under precedence constraints, improving structured decision-making algorithms.
Research presents methods for reliable mislabel detection in medical imaging datasets, addressing challenges of limited expert annotators.
Researchers introduced NMIRacle, a two-stage generative AI framework for molecular structure elucidation from IR and NMR spectroscopic data.
Researchers propose a graph-based Bayesian optimization framework using GNNs to automate the design and refinement of variational quantum circuits.
Research proposes EHR-MPC, a framework using generative patient digital twins for adaptive, inference-time controlled sepsis treatment, decoupling learning from optimization.
Research explores statistical efficiency and inference for quantile distributional reinforcement learning, focusing on characterizing return distributions.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion