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 an inference-time single-neuron intervention to continuously adjust investment stance and bias in LLMs.
Research identifies 'answer churn' in RAG systems, where expanding retrieval indices changes outputs despite static configurations.
Research shows LLM decision-making conflates protected proxy attributes with predictive evidence, undermining standard demographic bias audits.
Research demonstrates Best-of-N inference scaling causes 'safety hacking' when imperfect safety filters are paired with reward maximization.
AgentWeave proposes a pre-routing strategy to reduce candidate tool sets before LLM inference, lowering prompt token costs and latency.
Research identifies critical diagnostic flaws in RAG evaluation and proposes a causal leave-one-out probe for context optimization.
Researchers introduce Deep Contrastive Unlearning to selectively remove private or copyrighted data from LLMs without full retraining.
Researchers introduced SAFE, a framework using LLMs as verifiers to check intermediate step-by-step reasoning in multi-hop QA.
Researchers introduced GRACE, a benchmark evaluating step-level hallucinations in Chain-of-Thought reasoning over provided source context.
ArXiv research demonstrates safety training via DPO can persist during helpfulness optimization in multi-step agentic tool-use environments.
Research introduces the Safety Asymmetry Score to measure how AI agent susceptibility to malicious prompts varies by input channel.
Research demonstrates commercial LLMs exhibit systematic incumbent brand bias and cognitive manipulation dynamics in product recommendations.
TraceSQL introduces a reference-free method to verify Text-to-SQL query accuracy at inference time without requiring ground-truth SQL.
New research shows frontier LLMs exhibit poor confidence calibration when hallucinating catalog items in recommendation tasks.
Researchers identify security vulnerabilities in multi-turn LLM search agents, exposing them to prompt injection via retrieved web content.
A new framework predicts vision-language model performance before training, bypassing costly compute-based scaling trial-and-error.
Researchers propose RSMeM, a knowledge-enhanced memory evolution framework for remote sensing agents to improve domain-specific analysis.
Research finds behavioral detection of unfaithful Chain-of-Thought (CoT) explanations fails when LLM answers are incorrect, hindering oversight.
Telco-GAIA introduces a bilingual, multi-modal benchmark for evaluating tool-using agents on real-world telecom data with complex reasoning.
CMI-Mem proposes an RL-based memory manager for agent systems, using a hybrid reward combining QA correctness and intrinsic Conditional Mutual Information.
Research finds model performance improves when deeper transformer layers learn context-free value vectors, challenging standard attention paradigms.
New research proposes JoLT, a near-lossless KV cache compression method for LLMs, addressing memory limits and improving inference throughput.
Research proposes LoID, a method to extract informative prior distributions from LLMs for Bayesian logistic regression, improving generalization on small datasets.
Research proposes 'Verbalized Assumptions' framework to elicit and control LLM sycophancy by making implicit user assumptions explicit.
Research identifies a unified mechanism for harmful content generation in LLMs, indicating current alignment training is brittle and jailbreaks exploit a common vulnerability.
The SEC is probing AI-focused hedge fund Situational Awareness following its near-implosion.
Deutsche Bank announces collaboration with Google Cloud to co-design AI solutions tailored for financial services operations.
The SEC has subpoenaed Wall Street banks regarding Leopold Aschenbrenner's AI hedge fund, Situational Awareness, following its near-collapse.
Stanford research reveals employment for entry-level workers in AI-exposed occupations dropped 19% compared to non-exposed roles.
Nvidia customers face potential 15% price increases on Blackwell and Rubin AI systems due to rising memory costs.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion