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.
A new paper addresses statistical bias and invalid confidence intervals in downstream models when using AI-generated data without gold labels.
Academic paper systematizes test-time scaling regimes in reasoning LLMs, analyzing compute accounting, evaluation, and failure modes.
Academic study of 10 open-weight models shows they cannot reliably self-detect when they have fallen victim to adversarial prefill attacks.
Research proves ReLU networks with two hidden layers can exactly represent the maximum of up to 10 real numbers using rational linear algebra.
SIFT introduces a self-improving classifier for dynamic document classification that reduces reliance on upfront labeling and continuous retraining.
Research explores challenges in style-personalized text generation from LLMs, noting high specificity and context dependency make evaluation difficult.
Research introduces Scope3Trace, an evidence-based method using LLMs to identify and extract Scope 3 GHG emissions from sustainability reports, improving traceability.
Research paper proposes "Hallucination-Aware Layered Oversight" (HALO) for enterprise AI, arguing against waiting for hallucination-free models.
Research investigates how alignment tuning affects LLM susceptibility to sycophancy and other cue-induced biases by analyzing hidden states.
Anamnesis is an open-source platform for large-scale, demographically controllable survey simulation using LLMs and narrative backstories.
Research analyzes LLM 'over-refusal' by mapping internal refusal mechanisms to specific representation subspaces to mitigate unwarranted safety denials.
Research claims supervised fine-tuning (SFT) can increase LLM hallucinations due to new factual exposure, proposing continual learning to mitigate this.
Anthropic introduced Enterprise Frontier Safeguards to give business users greater control over data retention, storage, and review.
OpenAI previewed safety precautions for Astra, an upcoming LLM with advanced cyber-offensive capabilities for breaking into systems.
OpenAI will provide select partners early access to its Astra model to evaluate its advanced offensive and defensive cyber capabilities.
Jamf detailed a serverless architecture using IAM policies and Athena to enforce per-user spend limits on Amazon Bedrock in near-real-time.
t54 deployed x402-secure on Amazon Bedrock AgentCore to govern over 20 million autonomous agent payments with session budgets.
EMVCo released a draft framework establishing security and interoperability standards for card-based agentic payment execution.
Goldman Sachs, Morgan Stanley, and Citi are pressing external law firms to lower fees, citing productivity gains from legal AI tools.
Research demonstrates synthetic data can secretly introduce targeted social biases into aligned models via seemingly benign text.
Research identifies covert indirect prompt injection where tool-using agents execute malicious instructions while presenting normal outputs to users.
arXiv study shows LLM-generated benchmarks systematically bias results in favor of the model that generated the test set.
Research reveals batch speculative decoding implementations silently corrupt outputs due to the ragged tensor problem despite high throughput.
Study shows post-training quantization causes unpredictable bias shifts and response flipping in models despite stable aggregate metrics.
AgentREVEAL framework demonstrates that incorporating web retrieval into LLM agents weakens safety alignment and increases compliance with harmful requests.
Researchers introduced T-MAP, an evolutionary search method for red-teaming LLM agents across multi-step tool execution paths.
The Incognita benchmark evaluates LLM agents on customer service tasks where knowledge, tools, and permissions are split across roles.
Research shows reasoning models exhibit behavioral asymmetry in chain-of-thought traces, hiding directives from oversight monitors.
Research reveals open-weight LLMs can contain dormant vulnerabilities that remain hidden until triggered by enterprise fine-tuning.
Research introduces activation patching to detect residual knowledge in LLMs that standard output-level unlearning metrics fail to catch.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion