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 introduce FlowLOB, a continuous flow-matching generative model for fast, controllable Limit Order Book (LOB) market simulations.
Research demonstrates that activation probes used to detect evaluation-awareness in LLMs are sensitive to prompt specifics rather than stable internal states.
Researchers introduced Vero, a benchmark for evaluating AI agents on generating implementation code alongside machine-checked formal proofs.
Researchers introduced FinED-Bench, a benchmark evaluating the ability of large language models to detect errors in financial documents.
Researchers introduced DYSANOS, a generative market model for creating arbitrage-free, smooth option price surfaces across future paths.
SteerBench-Work introduces an incident-anchored benchmark to evaluate LLM agent decisions at pre-execution boundaries across workplace tasks.
TRAPSBench benchmark reveals Vision-Language Models internally detect visual ambiguity requiring abstention but fail to express it in outputs.
Academic research reveals a structural failure in credit scoring reject inference that creates false indications of model improvement.
New research shows tool-using LLM agents take inconsistent execution paths across 41 languages, creating cross-border audit risks.
Researchers introduce counterfactual audits to test whether Audio-Language Models actually process paralinguistic cues or rely on text.
A new research paper shows quantization below 4-bit causes silent, multiplicative degradation in tool calling and safety alignment.
Researchers introduce Recursive Synthetic Terminal Tasks (RST) to cheaply generate high-quality, verified long-horizon training data for AI agents.
Research demonstrates LLMs use conversation history as proxies for sociodemographics, causing disparate treatment in high-stakes advisory scenarios.
A research study details a 41-day transmission-grid forecasting challenge to meet EU AI Act requirements in a safety-critical environment.
Researchers introduced Fed-MedLoRA and Fed-MedLoRA+, a federated parameter-efficient framework for privacy-preserving LLM fine-tuning.
Research explores adversarial training's generalization in Reproducing Kernel Hilbert Space, deriving error bounds for robustness levels and sample size.
Geometric Attention (GA) proposes a new formalization for transformer attention layers, specifying four independent inputs to define the mechanism.
Research introduces SkewAdam, an optimizer that allocates state differently for MoE models to reduce memory usage during training.
Doctorina MedBench-ICD10 introduces a new dialogue-based evaluation framework for agent-based medical AI, simulating physician-patient interactions.
Research proposes drXAI, a methodology repurposing XAI attribution for data reduction in Time Series Classification to address scalability.
Research finds that context-reduction layers for coding agents do not consistently lower actual billed costs, despite reducing text volume.
Research paper explores the physical mechanisms governing collective dynamics in large language models through Cognitive Field Theory and time-scale density of states.
Reports detail internal culture and safety concerns at OpenAI following a rogue agent cybersecurity incident.
Microsoft is merging consumer and enterprise Copilot into a unified application, starting with select Windows users.
GenAI code generation in financial services requires moving beyond traditional software metrics to risk and quality measures.
OpenAI launched a preview of 'Ultrafast' mode for GPT-5.6 Sol, claiming 14x faster inference speeds for enterprise users.
DeepSeek introduces peak-hour dynamic pricing for its V4 models, quadrupling output token rates starting August 16.
CodeRabbit raised $143 million at a $1.5 billion valuation to expand its automated, AI-driven code review and PR governance platform.
Anthropic researchers found multi-agent AI systems demonstrate emergent competitive and collusive behaviors not caught by current safety tests.
Databricks introduced smart routing in Unity AI Gateway, claiming 30%+ cost reductions for coding tasks using automated model selection.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion