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.
Samsung semiconductor engineers are reportedly moving to rival SK Hynix, citing better work conditions and opportunities in the competitive chip industry.
METR proposes frameworks for independent post-incident investigations of autonomous AI agents that violate developer intent or safety constraints.
Expert commentary notes widespread discussion on open-weight models, but only Kimi K3 has commercially shipped today.
Researchers found top image editing models on Hugging Face easily create explicit deepfakes, with evidence from 1,000 user prompts.
Nvidia employee detained in Taiwan over alleged AI chip smuggling to China, implicating the company in black market activities.
Nvidia is reportedly backing a $50bn lease for a Texas data center, leveraging its balance sheet to support the growth of the AI computing market.
Researchers introduce CORVUS, a framework that optimizes LLM coding agent context by synchronizing changing files instead of appending static snapshots.
Researchers proposed an influence-based data auditing pipeline to identify safety risks and annotation errors in LLM fine-tuning datasets.
Researchers propose optimizing Transformer inference on FPGAs for real-time anomaly detection in high-frequency financial time series.
Research shows post-training quantization preserves classification accuracy but alters the internal reasoning and explanation outputs of deep learning models.
Researchers propose expanding Pearl's causal framework with Causal Zeros and Causal Differential Equations to handle economic feedback loops.
Researchers propose a mathematical method to prevent harmful fine-tuning of open-weight models while retaining benign adaptation.
Researchers propose MixQuant, an adaptive mixed-precision quantization method that optimizes LLM memory budgets post-calibration.
Researchers propose DiffTilt, a diffusion-guided search method to identify rare safety-critical failures in autonomous cyber-physical systems.
Researchers propose foundation models and fine-tuning paradigms for zero-shot time series forecasting across diverse datasets.
Researchers introduced AlloBench, a benchmark evaluating LLM agents on their ability to efficiently allocate and reuse tools under budgets.
Researchers prove mathematically that GRPO-based models, like DeepSeek-R1, suffer from inherent length bias and mathematical optimization flaws.
Researchers propose a Directional Influence Function to track how training data affects models trained under explicit safety and fairness constraints.
Researchers establish mathematical limits on when increasing model depth can compensate for precision loss in quantized neural networks.
Researchers identify a systematic bias in on-policy model distillation, where token-level teacher feedback is confounded by trajectory outcomes.
Researchers proposed WISERouter, an LLM routing framework that optimizes query allocation to balance model utility against global budget constraints.
Researchers introduce DECAF, a machine unlearning framework designed to defend against clustering attacks that reconstruct deleted data.
A research paper introduces a budget-aware evaluation framework for active retrieval-augmented generation (RAG) systems to optimize costs.
Researchers introduce MAPLE, a deep learning stock-ranking method that explicitly generates diverse, low-correlation multi-alpha signals.
Researchers propose KAP, a framework to optimize LLM serving by linking structured knowledge-selection signals directly to backend KV state consumption.
Researchers introduce DynaCalKV, a Key-Value cache compression method using adaptive rank allocation to optimize LLM long-context inference.
Researchers demonstrate that LLM conceptual geometry and relationships are dynamically rewritten by in-context instructions rather than static world-models.
Researchers present a GPU-accelerated method for Symmetric Non-negative Matrix Factorization to scale portfolio risk-factor estimation.
Researchers evaluate fuzz testing methodologies to systematically detect failures and edge-case crashes in Reinforcement Learning agents.
Researchers prove that correcting distribution drift in temporal graph generative models is mathematically impossible from observations alone.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion