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 GraniKV, an asymmetric KV-cache paging layer that optimizes multi-agent inference by splitting prefix and suffix storage.
New research applies temporal graph prototype-conditioned conformal prediction to provide mathematical coverage guarantees in fraud detection.
Researchers introduced a self-supervised auxiliary task discovery framework to improve reinforcement learning stability in stock trading.
Research identifies silent retrieval collapse in agentic workflows when tools or APIs originate from differing documentation styles.
Researchers propose a novel machine unlearning method designed to remove specific data subsets from trained models.
Research identifies workspace topology as a security attack vector in agentic coding assistants that access local developer filesystems.
Research shows LLMs can predict task failure risk but struggle to dynamically select the most cost-effective multi-agent collaboration protocol.
Research measures VRAM consumption patterns and memory stability for quantized code-synthesis agents on NVIDIA H100 GPUs.
New statistical methodology evaluates surrogate-derived ML models under covariate shift when gold-standard labels are scarce.
Study reveals expert and crowd annotator pools disagree on 23.6% of benchmark preference data, exposing flaws in standard model evaluations.
A 45-task study demonstrates that evaluating class-imbalance techniques on single datasets like Kaggle fraud leads to unsafe conclusions.
ArXiv paper demonstrates domain-specific triplet fine-tuning of embedding models to significantly improve business and person entity resolution.
Research on 24 open-weight models shows monitor detection skill, not architectural lineage, drives trusted guardrail ensemble accuracy.
Paper introduces 'coherence debt' framework to evaluate coding agents' ability to maintain consistency in repo-scale codebases.
Research demonstrates fundamental scaling limits and systematic bias in LLM-as-a-judge methods compared to traditional data annotation.
Research introduces weighted-conformal test martingales for nonparametric, real-time safety and drift monitoring of deployed AI models.
New theoretical research proves how verifier errors bound the performance gains of test-time compute scaling methods like Best-of-N.
Researchers introduced an adaptive expert routing model for financial networks to classify and explain specific anomaly mechanisms.
Researchers introduced a control-theoretic framework providing stability guarantees for open-loop LLM-based time series forecasting.
An arXiv paper introduces a theoretical framework for statistical evaluability in generative models to address test data evaluation gaps.
NVIDIA's GB10 edge hardware lacks process-level energy attribution, complicating power tracking for agentic AI workloads.
Researchers introduced TradeArena, an auditable testbed evaluating LLM trading agent behavior and risk alignment under market stress.
Empirical evaluation finds DL and LLM software vulnerability detection models degrade significantly under real-world distribution shifts.
Researchers introduced FactReview, an audit framework using LLMs to extract, cross-reference, and execute code to verify technical claims.
ReFind research shows agent-controlled search over raw chat logs matches structured memory performance without costly data pre-indexing.
Research shows majority-vote self-consistency degrades accuracy on complex reasoning tasks for small LLMs like Qwen2.5-7B and Llama-3-8B.
Research shows training AI agents against single LLM simulators causes simulator collapse, causing policies to fail in human testing.
An academic study of 17 text-to-image and 15 text-generation models reveals how multimodal AI resolves or fails polysemous word ambiguity.
Researchers developed a reinforcement-learning framework using Graph Neural Networks for optimal dynamic matching in weighted graphs.
Apple's research details a memory-efficient audio synthesis architecture, powering real-time, on-device expressive speech for Siri.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion