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 introduces SALT-GNN, a graph neural network method to improve anti-money laundering detection in high-activity transaction neighborhoods.
New research explores distance-preserving embeddings in inhomogeneous random graphs, improving traditional worst-case distortion guarantees for node representations.
Research identifies residual stream distributional drift as the root cause of sharp degradation in LLM post-training quantization below 4-bit precision.
GAE (Graph-Augmented Evolution) is a new framework addressing limitations in LLM-guided evolutionary program search for scientific discovery.
AuditWeave, a new research paper, proposes a tamper-evident, auditor-navigable evidence layer for AI-assisted workflows in regulated domains.
TabLoRA, a neural ensemble method, addresses computational challenges of deep learning models on large-scale tabular data, outperforming GBDTs.
Research proposes a method for Implicit Neural Representations to encode relative error scales by predicting distributions, avoiding expensive uncertainty estimation.
A research survey from arXiv outlines a new taxonomy for integrating Graph Neural Networks (GNNs) with Knowledge Graphs (KGs).
Research identifies a silent learning freeze in low-precision AI training, predictable from high-precision data and mantissa length.
Research evaluates Scientific Machine Learning (SciML) models like NODEs and PINNs, comparing them to ARIMA and LSTM for macroeconomic forecasting under violated structural prior assumptions.
Research indicates coding agents waste most context window reading, focusing on what context is actually needed for code editing rather than maximal use.
Research introduces Depth-Entropy Guided Sampling, a training-free method to improve LLM reasoning by using internal transformer dynamics.
Research identifies a trade-off in MLLM safety mechanisms, proposing output-aware safety guardrails to mitigate over-refusal.
Research proposes Manifold Constrained Tabular Deep Neural Networks to better handle rule-partitioned data, addressing the mismatch with Euclidean representations.
Research proposes a new Direct Preference Optimization (DPO) method, Meta-Reweighted DPO, to improve LLM alignment performance with noisy preference data.
Research paper argues ground truth datasets are human constructions, not objective truths, urging ML community to articulate hidden choices.
Research paper proposes "Quota Marketplace" for dynamic pricing and efficient allocation of scarce ML training resources to optimize ROI.
New research proposes a framework for Markov chain choice models with panel data, improving parameter estimation, personalized prediction, and assortment optimization.
Research proposes Balanced-Softmax Classifier-Retraining (BS-cRT) as a two-stage method to improve long-tailed recognition by refining the classifier.
Vrbo research explores training-free LLM candidate generation for long-tail vacation rental listings, addressing data sparsity for personalized recommendations.
Research proposes SMETA-ZSL, a zero-shot learning method using LLMs to classify emerging cyber threats without labeled data, leveraging CTI reports.
Research proposes embedding fairness constraints into ARDL models for retail sales forecasting and dynamic pricing, balancing profitability with consumer welfare.
Research proposes a multimodal routing framework for clinical prediction, enhancing interpretability and auditability over deep fusion methods.
Research introduces LeRoPE, a method for learnable Rotary Positional Encoding (RoPE) frequencies, improving language model performance by optimizing rotation rates.
New research suggests MLPs in LLMs store facts at an information-theoretically optimal rate, developing a theoretical account for this empirical phenomenon.
New research proposes Riemannian Isometric Policy Optimization to address exploration collapse in LLM reinforcement learning, fixing a flaw in PPO-Clip.
Research paper explores the physical mechanisms governing collective dynamics in large language models through Cognitive Field Theory and time-scale density of states.
New research proposes "predictive divergence masks" to improve the stability and performance of reinforcement learning for LLMs, enhancing policy updates.
Research explores how distribution shifts impact graph neural network (GNN) calibration and proposes a theoretical framework for understanding it.
Research explores Graph Neural Networks for RFID-based spatial geometry inference to improve indoor spatial understanding in intelligent systems.
Researchers introduced Weight-Adjusted Gradients (WAG) to identify influential parameters and failure modes in large language models.
Research evaluates LLMs' ability to generate executable Unity C# code for game scenes in a single pass, without iterative repair loops, across 26 concepts.
Generative AI augments small biomedical datasets (58 samples) to improve machine learning classification of glioma using Raman spectroscopy.
Research identifies two confounds—response determinism and access harness—in cross-model value comparisons, proposing a new separation protocol.
Research explores how unsupervised autoencoders learn macroscopic variables from microscopic spin configurations, identifying distinct learning dynamics.
Researchers introduced AVDC, a large-scale dataset for multimodal models to improve audio-visual understanding by decoupling semantics.
Research challenges the assumption that lower prediction error in latent world models leads to better control, finding it unreliable for planning.
Research explores using reinforcement learning with continuous, physics-based rewards to post-train LLMs for solving partial differential equations (PDEs).
Research explores Restricted Boltzmann Machines' (RBMs) vulnerability to out-of-distribution (OOD) inputs, analyzing effective-rank collapse as a failure mode.
New research proposes a foundation for online active learning tasks based on partial monitoring, optimizing costly information acquisition and prediction errors.
Research explores distributionally robust reinforcement learning (RL) for reducing the sim-to-real gap, addressing performance under worst-case scenarios.
Research introduces a theoretical framework for sequential learning with incomplete feedback, leveraging randomized confidence bounds for stochastic partial monitoring.
Research explores scheduling in parallel-server queuing systems with multi-class jobs, uncertain rewards, and bilinear reward models.
Research proposes Conditional Optimal Bridge for Riemannian Activation Steering, a principled method to control LLMs at inference time.
Research investigates adaptive normalization in Transformers, addressing gradient variance challenges in differentiable gating for improved training stability.
Research challenges static game theory's applicability to multi-agent learning, arguing it obscures dynamic disequilibrium and game-theoretic bounds.
Research identifies 'distributed backdoors' in multi-agent LLM systems where local runtime monitors fail to detect harmful payloads split across agents.
Research audits distributional reinforcement learning agents, finding that their inherent risk claims may not be accurate for interpretability and safety monitoring.
Researchers developed a training-free method to impute off-screen player positions in football broadcast data, improving spatial analytics accuracy.
Research details method to attribute LLM serving speedups to runtime, kernel, and quantization components on NVIDIA RTX A5000 GPUs using FP16 intermediates.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion