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.
Research defines 'prompt privilege,' measuring how LLM output quality varies based on user literacy, phrasing, and prompting capability.
Research identifies temporal misgrounding in legal RAG systems, where LLMs fail to retrieve historical laws applicable to past periods.
New research highlights LLM vulnerabilities where implicit natural language context bypasses standard alignment and safety filters.
Audit reveals nearly 60% of unsolved SWE-bench Verified tests are flawed, prompting a new multilingual refactoring benchmark.
Research demonstrates standard AI disclaimers in high-stakes domains like finance paradoxically induce user overconfidence and habituation.
Researchers introduced Intent Violation Rate (IVR) and a pilot benchmark to measure how often LLM-generated code misses implicit constraints.
Researchers present IntelliAudit, a retrieval-grounded multi-agent system evaluating IT security and compliance controls from enterprise evidence.
Updated Financial Touchstone benchmark evaluates 20 models, including Chinese open-weight options, across 495 international annual reports.
Researchers introduce SAVOR, a method enabling one-shot indirect prompt injection attacks on tool-using LLM agents without repeated queries.
PolicyKG uses a LangGraph pipeline to extract obligations, permissions, and prohibitions from policy PDFs into executable SHACL constraints.
New research models probabilistic re-identification risks in LLM query delegation, exposing gaps in standard PII-redaction gateways.
Research shows a rule-gated 7B analytics agent outperforms a 32B model on enterprise data tasks by enforcing business constraints.
PAM introduces a framework for training custom LLM moderation filters tailored to user-defined enterprise policy guidelines.
Researchers introduced OBCache, a KV cache pruning method using Optimal Brain Surgery principles to reduce memory in long-context LLMs.
Research identifies information disclosure failures in multi-turn LLM agents where internal reasoning and state are not adequately shared with users.
Autorubric introduces a framework to mitigate position bias, conflation, and calibration errors in LLM judges for non-verifiable tasks.
CC-OCR V2 introduces a benchmark evaluating Large Multimodal Model failure modes under real-world visual document acquisition conditions.
Research demonstrates standardized Q&A benchmarks for LLM fairness are unreliable, as prompt construction drives score variance.
New research shows automated legal citation-grounding metrics measure underlying graph coverage depth rather than actual LLM hallucination rates.
Academic paper proposes provenance analysis to replace inconsistent LLM-as-a-judge guardrails for preventing misaligned agent actions.
Researchers propose Retrieval-Augmented Defense (RAD) to dynamically block evolving LLM jailbreak attacks without model retraining.
Research shows LLMs experience dual-process interference and retrieval failures when handling conflicting or superseding contextual updates.
Research demonstrates tool-augmented LLM agents harbor implicit state that persists across sessions and propagates across boundaries.
New research uses quality-diversity search to stress-test Process Reward Models by finding edits that trick step-wise scoring.
A reproducibility study evaluates whether lightweight MLP probes on LLM latent activations reliably detect harmful prompts across model families.
Research reveals fidelity gaps and narrative fragility when using LLMs to translate credit model feature attributions into narratives.
arXiv paper reveals confidence-based data cleaning in credit default models introduces systematic income bias against low-income defaulters.
SAGE introduces learned adaptive retrieval for production RAG, dynamically adjusting retrieval budgets based on query difficulty and latency SLOs.
Research demonstrates that offline log replay fails to accurately evaluate model switching inside multi-step agentic workflows.
Research demonstrates that standard multi-LLM routing benchmarks overstate deployable gains by relying on unobservable oracle outcomes.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion