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.
Chronofy, a new RAG architecture, introduces a temporal-logical decay mechanism to prevent LLMs from using obsolete facts, addressing 'temporal hallucination'.
Research introduces CT-Merging, a method to combine multiple LoRA adapters into one multi-task adapter, reducing storage and inference complexity.
Research paper introduces Fisher widths to measure local parameter fluctuations and provide anisotropic recovery in statistical manifolds.
Research shows Bayesian uncertainty estimation (Monte Carlo dropout) improves confidence calibration and error detection in medical image AI models.
Research identifies 'Spectral Drift'—frequency-domain instability in internal activations—as a signature of neural network misclassifications.
A research paper introduces STeMP, a spatio-temporal modeling protocol to address data sensitivity and methodological choices in ML model quality estimation.
Research investigates when weight-tied looped transformers implement algorithms, finding a linear computation frontier dictated by training budget.
Research proposes analytic federated learning for medical image classification, enabling collaboration across institutions with task-heterogeneous data in one round.
Research introduces 'Parity Bottleneck Layers' to enable interpretable transformer models by addressing memory and compute costs of per-layer bottlenecks.
Research proposes Adaptive Multi-Horizon Reinforcement Learning, allowing AI agents to flexibly adapt temporal discounting for decision-making.
Research presents a method for learning safe, high-dimensional feedback control using control barrier functions (CBFs) to enforce safety constraints.
Research explores explanation-based runtime verification for trustworthy ML models in optical network automation, mitigating risks from incorrect decisions.
Research proposes a Fisher-market formulation to allocate fractional credit between planned tasks and performed actions, moving beyond all-or-nothing links.
Researchers propose CEDAR, a constraint-based method for lagged causal edge discovery in sparse autoregressive time series data.
Research explores how architectural conditions in active inference agents lead to 'causal emergence,' linking predictive organization to information theory.
Research finds LLMs struggle with dynamic user intent in iterative tasks, as current evaluation focuses on single-turn, fully-specified settings.
Research proposes Cardinality-Decomposed Loss to better match training objectives to varied relation structures in graph neural network-based recommendation systems.
GaugeQuant proposes a method to learn quantization-optimal bases for transformers during training by breaking internal symmetries, reducing activation outliers.
Research explores mechanistic interpretability for fine-tuned weather forecasting foundation models in atmospheric chemistry, aiming to understand their internal processes.
New research from arXiv proposes a de-biased test for synthetic minority data validity, claiming current methods are redundant or invalid.
Research explores memory architectures for AI models, testing whether models should rewrite internal memory during inference to improve recall with interference.
New research proposes normalized scores for clustering homogeneity and parsimony, addressing a trade-off in external clustering validation.
New research explores the computational complexity of optimally training neural networks, focusing on lower bounds for linear and ReLU activation functions.
New research proposes BR-Async-Q, a robust Q-learning algorithm to defend against adversarial corruption of rewards and states in reinforcement learning.
Research proposes an unsupervised, context-augmented language model using perplexity for Advanced Persistent Threat (APT) detection in security logs.
Research explores offline goal-conditioned reinforcement learning with hierarchical action chunking to address long-horizon task challenges.
Research introduces 'inspect_permute' to quantify LLM position bias in multiple-choice benchmarks via exhaustive answer-order permutations.
Research proposes TwistedMerge, a method for certified diagnostics and abstention in merging independently trained or fine-tuned models.
CudaPerf, a new reflective RL framework, enhances LLM-based CUDA kernel generation by incorporating structural and performance-aware rewards beyond mere outcome signals.
Research proposes TRISHUL, a spectral-control framework to improve federated parameter-efficient fine-tuning (PEFT) robustness under non-IID data heterogeneity.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion