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 benchmarks nine progressive sampling strategies for data quality profiling, aiming for efficient near-real-time monitoring in data-centric AI.
A new benchmark, HANDBOOK.md, evaluates long-context agentic instruction following, testing if agents adhere to lengthy policy documents during tool use.
PatientAgentBench introduces a new benchmark framework to evaluate patient-facing health AI agents, focusing on conversational and reasoning capabilities.
Research explores using LLM-verbalized confidence to route retrieval-augmented generation queries, improving efficiency and reducing irrelevant data.
Researchers propose MemSFT, a method using an external parametric memory to mitigate catastrophic forgetting and 'alignment tax' during LLM fine-tuning.
AIriskEval-edu Demo is a platform for auditing pedagogical risks in educational explanations across five dimensions: accuracy, depth, focus, appropriateness, and bias.
Research investigates how fine-tuning Transformers by adapting specific layers, rather than the whole model, impacts learning generalization and selectivity.
New research introduces Stemma, a method to determine LLM provenance by mapping 'induced decision regions,' improving reliability over response-level analysis.
RSIBench-Data proposes a new benchmark to evaluate LLM agents' research capabilities in data-centric recursive self-improvement.
Research explores 'input-only suppression' of LLM evaluation-awareness latents to prevent models from detecting safety evaluations.
Researchers propose ClinPRISM, a cost-effective multimodal LLM reasoning framework for question answering over irregular clinical time series data.
Research identifies how and where distinct human persona representations are encoded within large language models, analyzing specific model layers.
Research explores if Large Reasoning Models (LRMs) exhibit human-like cognitive habits by analyzing recurring Chain of Thought (CoT) patterns.
PilotRL introduces a new reinforcement learning method for training language model agents, improving long-term strategic planning over ReAct.
Research finds public AI-art discourse marginalizes artists' concerns, despite their work training models and reshaping the creative economy.
A new arXiv survey reviews current challenges in harmful content generation by LLMs and explores safety mitigation techniques, highlighting the dual role of these models.
Research explores contrastive weak-to-strong generalization to train stronger LLMs from aligned weaker models without human feedback, addressing noise and bias.
VisRAG2.0 proposes an evidence-guided multi-image reasoning framework to mitigate visual hallucinations in Visual Retrieval-Augmented Generation (VRAG) systems.
New research introduces $M^2PO$, a multi-perspective preference optimization method addressing blind spots in current Machine Translation (MT) quality estimation for LLMs.
Research introduces controlled synthetic pretraining tasks and identifies 'Canon Layers' to evaluate and improve language model architectures.
Researchers introduced LogicScore, a new evaluation method for Retrieval Augmented Generation (RAG) systems focused on global logical integrity.
Research investigates how syntactic properties of a first language (L1) influence processing of a second language (L2) in modern language models.
PatchWorld introduces a method for inducing executable code as a world model in black-box environments for agent decision-making without gradient-based learning.
KletterMix introduces a high-quality, reusable German pretraining corpus for language models, addressing the scarcity of well-curated German linguistic resources.
Researchers introduce SPO (Stochastic Prompt Optimization), a framework for black-box search over prompt space to improve AI systems.
Researchers introduce NormWorlds-CF, a solver-verified environment for evaluating LLM normative reasoning, providing proofs and falsification certificates without LLM judges.
Research audits eight generative AI platforms for resume screening, revealing competence gaps and intersectional bias beyond traditional fairness metrics.
Researchers introduced Deep Delta Learning (DDL), a new structured residual update method for transformer models designed for targeted state edits.
Researchers propose a sparse autoencoder method to control LLM reasoning strategies, aiming for more reliable and flexible path selection.
Blackstone-owned AirTrunk secured a $2.3 billion green loan from 30 lenders for a new data center in Malaysia, its largest single asset financing.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion