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.
Researchers introduced a hybrid framework combining Transformer models with Benders decomposition to accelerate stochastic optimization.
Researchers propose "program distillation" to convert opaque, costly LLM-as-a-judge evaluators into deterministic, code-based programmatic committees.
Researchers propose TokenMem, an architecture injecting knowledge into frozen LLMs via a cross-attention channel to avoid memory conflicts.
Researchers developed a pipeline to extract and verify the actual algorithms LLMs use internally to predict Hidden Markov Models.
Researchers demonstrate that visual token pruning in multimodal LLMs mitigates security vulnerabilities, jailbreaks, and hallucinations.
Researchers identify MLX scheduling bottlenecks on Apple Silicon, proving CPU+GPU co-execution is viable for the prefill phase, not decode.
Researchers formalize the theoretical limits of runtime safety gates for tool-using agents, proving safety policy decidability limits.
Academic study challenges whether code coverage and mutation scores of LLM-generated tests reliably correlate with actual bug detection.
Researchers defined and measured the "misguidance effect," where prompting LLMs with buggy code leads them to generate tests that validate the bug.
Researchers propose Agent Team Work Zone, a persistent workspace architecture to prevent state loss in multi-agent LLM coding workflows.
Researchers propose using Hidden Markov Models to assess LLM reliability and uncertainty, moving beyond static benchmark accuracy.
Researchers propose metamorphic testing to detect logical domain violations in ML models that traditional performance metrics fail to capture.
Researchers propose a conformalized selection method that maintains statistical guarantees even when calibration data has noisy labels.
Researchers proposed a model-agnostic framework combining multi-scale forecasting with gate-localized conformal prediction for nonstationary time series.
Researchers introduce a logit-coordinate framework to improve continuous generative models handling mixed tabular data.
Researchers propose using physical GPU WebGL render-timing dynamics as a passive, AI-resistant CAPTCHA signal to detect automated bots.
Research paper proposes an offline-to-online workflow using generative models and adaptive testing for ad creative optimization, leveraging historical A/B test data.
Research finds LLM-generated authentication code has security vulnerabilities, even with iterative reprompting, when assessed against NIST SP 800-63B.
Researchers propose a two-level Hierarchical Reinforcement Learning (HRL) framework using Soft Actor-Critic for sparse-reward, long-horizon tasks.
Research from Compass details how firms struggle to choose AI projects and highlights two contrasting examples of project success and failure.
Research explores deep learning techniques, including augmentation, to address bias in medical image classification for rare conditions due to long-tailed data distributions.
Research introduces SpecBox, a speculative scheduling technique for LLM agent sandboxes to improve resource utilization and reduce tail latency.
New research identifies a late-time relaxation mechanism for grokking in linear models, extending to nonlinear neural networks.
Research explores acoustic biomarkers for Alzheimer's detection, using AI models to predict clinical status and human perceptual scores across languages and genders.
HydroAgent formalizes expert knowledge into AI-driven flood forecasting workflows, integrating tacit rules, review checkpoints, and operational constraints.
Research proposes Legendre-regularized policies for contextual optimization, enforcing hard constraints while maintaining smoothness for gradient training.
Research introduces SpecFormer, a transformer architecture mitigating embedding and attention collapse for improved recommendation systems.
Research proposes a global, guideline-grounded evaluation method for post-hoc XAI, addressing misrepresentation of model behavior in critical applications like ECG.
Research explores variational quantum conditional Boltzmann machines (QCRBMs) for time-series forecasting, benchmarking against classical models.
Research explores contextual dynamic pricing using non-stationary generalized linear models to optimize revenue for sequentially arriving consumers.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion