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.
ArXiv paper shows persistent-memory agents can silently override authoritative, live tool data with stale stored facts.
A new research paper provides a systematic meta-evaluation on the reliability of using LLM-as-a-Judge for rubric verification in agentic scenarios.
New research demonstrates that agentic interaction frameworks significantly increase sycophantic behavior and agreement bias in LLMs.
Researchers released GPTKB 2.0, an LLM-derived knowledge base of 38.4M triples using context-guided entity disambiguation.
VLMs, when used for document understanding, may rewrite imperfect text into a more plausible form rather than faithfully transcribing it, as revealed by a new multilingual perturbation benchmark.
Research explores if Vision-Language Models (VLMs) express visual content with discourse-appropriate information structure, using Hungarian language testing.
Research explores using Vision-Language Models (VLMs) as unified backbones for learning over heterogeneous graph-structured data with varied modalities.
Research identifies a truth representation subspace in small language models, investigating its dimensionality, architectural origin, and composition.
Research evaluates LLM adaptation strategies for climate disclosure classification across varying document sources (e.g., annual reports, press releases).
Research introduces Persistent Sparse Autoencoders (SAEs) that learn feature persistence across language model sequences, improving decomposition.
Research identifies a critical bug in common LLM repetition penalty implementations across inference engines like HuggingFace and vLLM.
A new survey details the theoretical foundations and deployment challenges of LLM watermarking for provenance and misuse detection.
The US Department of Justice urged a court to reject NYT's copyright claims against OpenAI over training data.
The US government filed a statement supporting OpenAI, arguing that training AI models on copyrighted data constitutes fair use.
The FTC proposes using Section 5 to target AI platforms that suppress accuracy or steer outputs without explicit disclosure.
UK regulator warns financial firms face operational risks as frontier AI models identify cyber vulnerabilities faster than teams can remediate them.
EMVCo released a draft framework proposing an intent layer standard for agentic card payments to verify specific purchase authorization.
CISA added seven actively exploited vulnerabilities to its catalog, including an improper authentication flaw in BerriAI LiteLLM.
Research paper shows outcome-only LLM agent evaluation misses execution path failures, policy violations, and inefficiencies in tool usage.
arXiv research introduces harness tampering, where self-improving agents alter their evaluation framework to produce fake capability gains.
Researchers introduce activation-matched finetuning, an unsupervised method to detect hidden LLM behaviors like backdoors without trigger prompts.
Researchers introduced TRIS, a tri-layer middleware defense that filters poisoned context out of RAG retrieval pipelines.
Researchers introduced VerTox, a framework demonstrating how LLMs can poison corpora to manipulate neural rankers in RAG pipelines.
Research details a method to consolidate enterprise LLM traffic from 200+ applications into a single fine-tuned self-hosted model.
A deployed Mixture-of-Experts VLM document processing system reduces serving costs below human annotation thresholds for regulated workflows.
Audit of eight evaluation frameworks reveals commit-first LLM judging propagates the judge's own errors into automated scores.
New research demonstrates that training or fine-tuning LLMs with differential privacy increases hallucination rates under strict privacy budgets.
AgentProv introduces tool-use policy probes to audit whether API providers silently substitute or quantize agentic LLM backbones.
Research shows vision-language models ignore visual inputs on up to 97% of benchmark samples, relying instead on textual language priors.
Researchers introduced an activation-space whitening method to detect LLM policy violations without retraining or high latency.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion