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.
Meta founder Mark Zuckerberg advocated for open-source AI, positioning Meta's models as alternatives to closed systems from OpenAI and Anthropic.
OpenAI introduced GPT-5.6-Cyber, a specialized model available via Daybreak Red for authorized vulnerability research and exploit validation.
OpenAI is granting vetted Daybreak program partners access to specialized frontier cyber models to provide governed cybersecurity services.
Security breaches at Hugging Face involving OpenAI agents, alongside incidents at Anthropic and Meta, have intensified AI safety concerns.
Researchers define the Cross-Lingual Comprehension Gap, measuring how model performance drops when processing the same content in non-English languages.
Researchers propose GRASP, a method using Group Relative Policy Optimization to train lightweight, local LLM-based data anonymizers.
Research evaluates seven confidence estimators for financial vision-language models to route misread charts and documents to human reviewers.
Research shows that prompting LLMs to verify and correct false presuppositions in queries degrades overall QA accuracy on standard questions.
Researchers propose Factorized Hypothesis Search to map indirect evidence, like table cells, to large taxonomies, closing retrieval gaps.
A systematic survey of 1,547 papers defines the 'horizon gap' where LLM agents fail at multi-hour tasks due to memory and planning drift.
Researchers introduce TA-RAG, demonstrating that source document tone systematically overrides LLM system prompts in RAG applications.
Researchers propose Autonomy-of-Heads (AoH), a data-free sparse attention method to reduce long-context KV-cache costs without calibration.
Researchers introduce LLMRouter, an open-source unified framework to standardize the evaluation, development, and deployment of LLM routing models.
A research paper evaluates how large language models reinforce user biases expressed in prompts and explores the boundary of prompt manipulation.
Researchers propose a hybrid knowledge graph pipeline grounding LLMs in Wikidata using agentic reflection to organize unstandardized data.
Researchers introduce an evaluation framework for crystallization—retaining and reusing verified Text-to-SQL repair episodes to save compute.
Researchers expose flaws in standard LLM benchmark contamination metrics and propose a more accurate probability-based detection method.
Researchers introduce ADIAS, an issue-centric framework designed to automate the design and iterative repair of interactive agent systems.
Researchers introduce StepJack, a benchmark proving computer-use agents are vulnerable to multi-step indirect prompt injection attacks.
Researchers demonstrate that multimodal model confidence readouts can be manipulated via image attacks while preserving the exact text answer.
Researchers propose IB-RL, a reinforcement learning framework designed to train dialogue agents in strategic, adaptive environments.
Researchers introduced SABRE, an automated pipeline that generates stress tests for Vision-Language Models to identify model weaknesses.
Researchers introduce SkillProx, an academic framework that dynamically refines and deletes agent skills in-context using textual gradient descent.
Researchers analyzed how exposing long reasoning chains to LLM judges influences their evaluation of answer factuality and bias.
Researchers introduce Multi-Legal-Bench, a cross-jurisdictional benchmark evaluating LLMs across six European legal systems and languages.
Research demonstrates that LLM judge reliability falls as the number of verdicts per call increases, recommending a 'sharding' architecture.
Researchers proposed an adversarial framework to detect and falsify incorrect causal structures encoded within generative AI models.
Researchers find recurrent context compression in long-horizon agents causes execution instability, blocked actions, and repetitive loops.
Researchers identify "removal-budget confounding" in adaptive data cleaning, where automated filtering shifts model validation metrics.
Researchers introduced CertBind, a mathematical framework to guarantee the reliability of retrieval decisions when connecting frozen multimodal models.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion