Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
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 introduced a new benchmark, the Metacognitive Monitoring Battery, to evaluate LLM self-monitoring across six cognitive domains using human psychometric methods.
Research proposes PolicyBank, a framework for LLM agents to evolve policy understanding via pre-deployment interaction and corrective feedback.
Hyatt deploys ChatGPT Enterprise with GPT-5.4 and Codex for global workforce productivity and operations, according to OpenAI.
An industry observer argues that the current state of AI is analogous to the early 20th century's electrification, not the dot-com bubble.
Anthropic updated Claude.ai's system prompt for Opus 4.7, marking an ongoing evolution in model instruction transparency.
A research workflow for deep understanding of open-weight LLM architectures, focusing on technical papers and implementation details.
PyCon US 2026, a major Python developer conference, will be held in Long Beach, CA, introducing new AI and security tracks.
Research introduces IF-RewardBench, a new benchmark to evaluate judge models' reliability in assessing LLM instruction-following, addressing current benchmark deficiencies.
Research proposes SecureGate, a token-gated dual-adapter method for federated LLMs to selectively reveal PII, aiming to mitigate privacy leakage.
Research explores 'effective abstention' for multimodal AI, allowing systems to decline answers when evidence is insufficient, underexplored in current benchmarks.
Research explored domain fine-tuning of Finnish BERT on medical text, observing embedding changes to predict pre-training benefits with limited labeled data.
Research found Chinese prompts are not more token-efficient than English for LLM coding tasks, refuting social media claims of 40% cost savings.
New research proposes XMark, a multi-bit watermarking method for LLM-generated text, aiming for improved message length, text quality, and decoding accuracy.
Fact4ac won a financial misinformation detection challenge using fine-tuned and few-shot LLMs for reference-free verification.
Research examines how retrieved context, especially irrelevant documents, affects internal representations within RAG models, beyond just output behavior.
Research finds schema key wording acts as an instruction channel in LLM structured generation, impacting performance beyond just structural constraints.
MARCA, a new benchmark, evaluates LLMs on multilingual web search and synthesis, focusing on English and Portuguese for reliability assessment.
Research claims graph neural networks (GNNs) match or surpass LLMs for specific close-ended human simulation tasks, introducing Graph-basEd Models for Human Simulation (GEMS).
Researchers propose IF-CRITIC, a fine-grained LLM critic to improve instruction-following evaluation, addressing deficiencies in existing LLM-as-a-Judge methods.
Research introduces Semantic Resonance Architecture (SRA) for MoE models, routing tokens based on cosine similarity to semantic anchors for interpretable decisions.
Research uncovers large language models' (LLMs) vulnerability to textual ambiguity, specifically in Chinese, via a new benchmark dataset.
Research claims 42% of turn-level findings in LLM conversation analysis are spurious due to uncorrected autocorrelation.
Research proposes RoPE-Perturbed Self-Distillation for long-context adaptation, addressing positional bias in LLMs fine-tuned for extended sequences.
OmniCompliance-100K is a new, rule-grounded, multi-domain dataset designed to enhance LLM safety and compliance evaluation using real-world cases.
Research proposes latent-geometric denoising to improve LLM knowledge boundary awareness, reducing hallucinations and excessive abstentions.
Research proposes CausalDetox, a method to identify and intervene on specific attention heads in LLMs responsible for toxic content generation.
Research identifies the 'LLM fallacy,' where users misattribute AI-assisted cognitive improvements to their own abilities, impacting self-perception.
Research studies speculative decoding's token acceptance rates across different cognitive tasks, revealing performance variations in LLM inference.
Research proposes Multi-Persona Thinking (MPT), an inference-time framework, to reduce social bias in LLMs by prompting reasoning from multiple perspectives.
Research identifies LLM overgenerations in machine translation, distinguishing between self-explanations, confabulations, and appropriate explanations.
New arXiv research introduces QuantCode-Bench, a benchmark to evaluate LLMs generating executable algorithmic trading strategies, focusing on domain-specific logic and API knowledge.
Research identifies segment-level coherence as a method to reduce false positives in LLM harmful intent detection, especially in CBRN contexts.
Research introduces SPAGBias, a framework to systematically evaluate spatial gender bias in LLMs, combining a taxonomy of urban micro-spaces and a prompt library.
Researchers propose a multiple-choice evaluation protocol with up to 100 options to better assess LLM competence beyond shortcut strategies, applying it to Korean orthography.
Research proposes combining LLMs with encoder-decoder translation models to improve multilingual performance, especially for low-resource languages.
New benchmark, MADE, for multi-label text classification in medical device adverse event reporting emphasizes uncertainty quantification (UQ).
Research explores Hierarchical Recurrent Memory (HRM-LM) as an alternative to flat Transformer layers, aiming for efficient, quality-matched representation.
Researchers developed HARNESS, a family of lightweight, distilled Arabic speech models achieving strong performance on ASR and dialect ID.
Research introduces DharmaOCR Full and Lite, specialized small language models for structured OCR, claiming superior transcription and stability over baselines.
Research finds multimodal LLMs underperform on visual tasks, with text centroid structure more critical than visual for accuracy across models.
Research introduces a method, "Faithfulness Serum," to improve the factual accuracy of textual explanations generated by LLMs for their decisions.
Research used GPT-4.1 to predict fan experience ratings from unstructured text, achieving two-thirds accuracy against survey scores across 10,000 baseball fan responses.
Research finds LLM reasoning errors often stem from early, specific transition points, leading to coherent but globally incorrect paths.
Research finds prompt optimization for compound AI systems often fails, with 49% of methods performing worse than zero-shot on Claude Haiku.
Research on AIMO 3 competition shows advanced prompting and diverse voter strategies fail to significantly improve LLM math reasoning; model capability dominates.
Research finds advanced LLMs with strong reasoning capabilities demonstrate less cooperative behavior in social dilemma games like Prisoner's Dilemma.
Research explores using reinforcement learning for prompt warmup to improve small language models (SLMs) for reranking in retrieval-augmented generation.
A research survey consolidates fragmented approaches to evidence-based text generation with LLMs, focusing on attribution, citation, and quotation.
Research identifies >80% redundant computation in parallel Chain-of-Thought LLM reasoning; proposes DeepPrune to mitigate inefficiency.
Research proposes MemGround, a new benchmark for evaluating LLM long-term memory in dynamic, gamified interactive scenarios, moving beyond static retrieval tests.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion