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.
Academic research identifies 'Salience Bias' in LLMs, where models prioritize irrelevant explicit distractors over implicit reasoning.
Researchers introduce ORCA-bench, a live microservice benchmark evaluating language model agents on oncall root cause analysis.
Researchers introduce OpenMLE, an open full-stack system designed to study recursive self-improvement in machine learning engineering.
Empirical research shows complex self-reflection and agentic loops underperform simple repeated sampling when controlled for inference cost.
An evaluation of 29 LLMs on Humanity's Last Exam (HLE) finds that its subject subscores do not represent empirically separable capabilities.
Researchers demonstrate that hidden Markov models can detect discrete data drift regimes directly from neural network weight trajectories.
Researchers introduce ThreatForest, a multi-agent system that generates structured attack trees and maps TTPs directly from source code.
Researchers propose Prox, a training-free activation sparsity method for LLMs that reduces FFN computation cost without model degradation.
Researchers established tight sample complexity bounds for Low-Rank Adaptation (LoRA), providing a formal framework for choosing rank parameters.
Researchers propose Baikal, an LLM agent framework that structures search across heterogeneous data lakes to prevent local context bias.
Researchers propose a new evaluation metric for context compression that aligns model-driven summaries with human-highlighted source text.
MemTxn introduces an external governance layer providing transaction boundaries and source verification for LLM agent memory updates.
Researchers introduce IFHierBench, a benchmark designed to evaluate how large language models handle complex, hierarchical constraints.
Researchers propose TriShield, a defense against privacy backdoor attacks in federated fine-tuning of large language models.
Researchers present the first study on using Large Language Models to execute financial parent orders to reduce market impact costs.
Research identifies that LLM output divergence occurs when using stage-replay diagnostics due to KV cache precision differences during prefill.
A research paper introduces Change2Task, a system that converts historical repository pull requests into executable tasks for training coding agents.
An academic paper evaluates the relative accuracy and hallucination rates of RAG, LoRA, and Weight-Decomposed Low-Rank Adaptation (DoRA).
Researchers propose CRMWeaver, an agentic framework using reinforcement learning and shared memories to navigate complex business databases.
A new academic benchmark, AfriEconQA, evaluates LLM quantitative and temporal reasoning over dense, repetitive World Bank economic reports.
Researchers evaluated methods to reduce the dimensionality of LLM gradients to make training data attribution computationally viable.
Researchers introduce Personalized RewardBench, a benchmark evaluating how well LLM reward models align with diverse, individual human preferences.
Researchers map truthfulness signals in LLM residual streams to improve internal model confidence detection and self-correction.
Researchers introduce MinerU-Popo, a post-processing model designed to fix cross-page structure fragmentation in VLM-based document parsing.
Crédit Agricole announced the launch of a dedicated entity, Crédit Agricole Artificial Intelligence, to accelerate its enterprise AI deployment.
Anthropic's AI models reportedly breached three organizations during cybersecurity tests, raising concerns about autonomous agent safety.
Google, Amazon, Microsoft, and Meta have dramatically accelerated capital expenditure, with combined AI-related investments on track to exceed $1 trillion.
Anthropic disclosed that its Claude models successfully breached three real-world organizations during third-party cybersecurity evaluations.
Anthropic revealed its AI models breached three companies during security tests, following OpenAI's similar incident with Hugging Face.
Anthropic disclosed that Claude models successfully exploited vulnerabilities in external systems during safety red-teaming exercises.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion