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 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.
Research explores ARC-AGI benchmark's human-level validation for fluid intelligence, linking performance to novel rule induction.
modelDNA is a new research tool to verify open-weight LLM lineage from sampled weight fingerprints, addressing the issue of unverified self-reported base models on platforms like Hugging Face.
PREF-Gate is a new auditable decision framework for graph-based fraud detection that addresses provenance-constrained relational evidence use.
STAMP introduces a provenance-guided credit assignment method for deep-search agents, improving reward distribution beyond trajectory-level scoring.
Research explores device-cloud collaborative LLM inference to address computational, memory, and energy constraints for multi-modal, multi-task models.
Research explores a learning-based method to satisfy boundary conditions in variational problems, improving the Deep Ritz Method.
Research indicates LayerNorm in looped transformers acts as an implicit gain controller, stabilizing recurrence even when operator norm exceeds 1.
Research proposes Federated Multimodal Knowledge Graph Completion (FedMKGC) to train over decentralized multimodal knowledge graphs, predicting missing links securely.
Research identifies a flaw in class unlearning evaluations where underlying class geometry causes information leakage about forgotten data.
Research explores Adaptive Reinforcement Learning to handle unobservable random delays in agent-environment interactions, moving beyond standard MDP assumptions.
Research introduces a unified framework to improve context-aided forecasting with LLMs, addressing current performance, diagnostic, and cost challenges.
Research evaluates memory-efficient LLM fine-tuning alternatives to backpropagation, including FmAD and zero-order optimization, against checkpointed BP.
Research explores optimal continual fine-tuning strategies for foundation models on resource-limited devices, balancing compute cost and model performance.
Research identifies the "modality gap" in CLIP-style multimodal models, where image and text embeddings misalign due to InfoNCE loss with independent encoders.
Research proposes a comparison-based ordinal learning method for proactive driving risk assessment, addressing collision data scarcity and label imperfections.
LSTrans is a lightweight hybrid deep learning model proposed for efficient and sensitive ECG classification on resource-constrained wearable devices.
Research proposes temperature sampling and variance-corrective time shifting to diversify output and reduce mode collapse in diffusion models.
Research identifies 'phantom transfer' in agentic synthetic data, where filtering harmful actions during training still propagates underlying intent.
Research paper proposes WattCouncil, a framework using governed LLMs to generate context-aware household energy scenarios while protecting privacy.
Research proposes answer-conditioned chain-of-thought distillation to adapt small VLMs for industrial vision tasks using minimal labeled data.
Research characterizes ReLU neural network approximation rates as a joint function of width and depth, enabling greater architectural flexibility.
BucketKD, a new knowledge distillation framework, reduces end-to-end motion planning model size while maintaining safety for autonomous systems.
Research proposes a novel watermarking technique for tabular data, including discrete, categorical, and mixed types, addressing limitations of existing methods focused on numerical data.
Researchers propose UNIT, a two-stage estimator combining deep representation learning with G-estimation for causal mediation analysis under no essential heterogeneity.
Research proposes a contrastive learning method to detect inconsistencies between LLM agent skill descriptions and their actual runtime behavior.
NetInjectBench is a new 130-scenario benchmark for evaluating indirect prompt injection risks in tool-using LLM agents for network operations.
Research explores production-ready federated learning in healthcare, addressing privacy, orchestration, and governance in MLOps.
Research explores integrating expert background knowledge into causal discovery algorithms to enhance scalability, identifiability, and accuracy for large datasets.
Research explores emergent generalization through representation learning in neural networks, identifying low-dimensional structures and their functional advantages.
Research explores using intermediate Vision Transformer (ViT) layers to correct last-layer classification failures, introducing 'recoverability'.
New research proposes Gradient-Skipping Relevance Propagation (GSRP) to improve explainability for Vision Transformers by better modeling attention heads and residual connections.
ZoRRO, a zero-weight, training-free personalized news recommender system, shows performance comparable to deep learning models but operates 600x faster.
EquiFusion introduces a kinematics-agnostic latent diffusion model for 3D human motion prediction, overcoming prior limitations of skeleton hard-coding.
Research introduces NAILS, a method for aligning recommender systems with normative objectives like fairness and diversity via internal label shift.
Research identifies Fourier locking as a core optimization bottleneck in quantum data re-uploading classifiers, proposing spectral homotopy as a solution.
Research proposes implicit neural networks as static controllers, enabling formal analysis for well-posedness and Lyapunov/IQC stability.
Researchers introduced a new conformal prediction method for constructing calibrated prediction sets over spatial events like tropical cyclones and earthquakes.
Research explores AI and machine learning for diagnosing cardiac amyloidosis, from single-modality detection to multimodal integration.
Research presents a depth-efficient quantum topological data analysis method, adapting Pauli Correlation Encoding for financial stress detection.
Researchers demonstrated a hybrid quantum-classical sampling workflow for discrete Markov random fields, comparing against classical MCMC for small MRFs.
ShapKO proposes a method for robust multimodal learning in medical models that addresses missing inputs and modality dominance during training.
Research introduces DUNE (Diffusion Unified Network refiNEr), a method to refine diffusion model backbones by analyzing internal latents to reduce artifacts.
Research introduces MorphologyFM, a foundation model for ECG and pulse oximetry waveforms focused on preserving clinically meaningful morphology.
SupplyNetPy is an open-source Python library for high-fidelity modeling and discrete-event simulation of complex supply chain networks.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion