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.
A research survey reviews empirical studies on LLM-based persuasion, categorizing applications and examining ethical implications.
Research claims adding simple supervision improves LLM alignment with diverse population groups across public health, public opinion, and values data.
Research identifies 'owner-harm' as a critical, under-addressed AI agent threat where agents harm their own deployers, citing real-world incidents.
Research identifies and measures "insider-outsider bias" in LLMs, where models default to mainstream cultural perspectives when generating interview scripts.
Research compared consistency of exercise prescriptions from GPT-4.1, Claude Sonnet 4.6, and Gemini 2.5 Flash across six scenarios, 20 generations each.
Research proposes personalized LLM benchmarks, arguing current aggregate evaluation methods overlook individual user preferences in real-world deployment.
Research investigates if GPT-5 and DeepSeek-R1 exploit gaps between valid proofs and faithful formalizations (formalization gaming) in logical reasoning.
Research explores integrating Sparse Autoencoders (SAEs) into LLM inference to understand robustness against gradient-based jailbreak attacks.
Research proposes Visual Contrastive Editing (VCE) to mitigate object hallucinations in LVLMs by leveraging visual contrastive pairs.
XpertBench introduces a new benchmark for LLMs on complex, expert-level tasks using rubrics-based evaluation to counter plateauing performance.
Research introduces AskBench, an interactive benchmark to evaluate and improve LLMs' ability to ask for clarification, reducing hallucinations.
Research introduces Persuaficial benchmark to detect AI-generated persuasive text, analyzing linguistic differences between AI and human persuasion.
Research indicates LLMs exhibit performance degradation when processing multiple instances, affected by instance count and context length.
Research identifies and evaluates 'temperature-constrained Non-Deterministic Machine Translation' (ND-MT) as a distinct phenomenon in modern MT systems.
Anthropic briefly updated and then reverted its Claude.com pricing page, suggesting a move of 'Claude Code' from the $20/month Pro plan to higher tiers.
OpenAI reportedly launched GPT-Image-2. Cursor secured a $10B contract with xAI, with a $60B acquisition right, as per Latent Space.
OpenAI introduced an open-weight model, OpenAI Privacy Filter, for PII detection and redaction in text with high accuracy.
OpenAI launched ChatGPT Images 2.0, with Sam Altman claiming a performance leap from 1.0 equivalent to GPT-3 to GPT-5. User testing showed improved object recognition and scene composition.
Google DeepMind is collaborating with global consulting firms to expand the deployment of its frontier AI models across various organizations.
Hugging Face launched QIMMA, a quality-first leaderboard for Arabic Large Language Models, evaluating various models on multiple Arabic NLP tasks.
METR Research is measuring how much AI agents can accelerate AI R&D and how this impact is changing over time.
Research indicates LLMs persuade psychologically susceptible individuals on societal issues via emotional appeals and perceived AI trust, despite logical fallacies.
New additive feature-group-aware stacking framework (STRIKE) proposed for credit default prediction, combining interpretability with performance.
SPaRSe-TIME introduces a low-rank temporal modeling technique for time series prediction, aiming for efficiency and interpretability over traditional RNNs.
Research identifies collaboration gaps in human-LLM interactions, noting users must frequently correct misunderstandings and misaligned responses.
Research finds state-of-the-art vision-language models (VLMs) exhibit strong biases in objective visual tasks like counting and identification.
REALM proposes fine-tuning LLMs with noisy human annotations by jointly learning model parameters and annotator reliability, surpassing standard aggregation.
Research benchmarks cloud and local LLMs on system dynamics tasks, specifically causal loop diagram extraction and interactive model discussion.
Researchers propose a single-sequence method for LLM uncertainty estimation, aiming to reduce computational cost versus multi-sequence approaches.
Research paper proposes methods for reliable testing and quality assurance of machine unlearning algorithms, addressing regulatory compliance.
Research proposes using dropout as a compression tool to reduce the computational and memory costs of influence functions for ML models.
Research details Fission-GRPO, a reinforcement learning method enabling LLMs to recover from tool-call errors, improving multi-turn task reliability.
Research finds differentially private SGD (DP-SGD) in neural networks harms model fairness and adversarial robustness due to feature learning degradation.
Benchmarking of 17 multimodal models on a challenging handwritten form achieved 85% accuracy with latest Google and OpenAI models.
Research clarifies non-stationarity in time series foundation model embedding spaces, distinguishing it from distribution shift, crucial for SPC.
Research identifies 'XOXO' cross-origin context poisoning, enabling attackers to subtly compromise AI coding assistants by injecting malicious context.
UniComp introduces a unified evaluation framework for LLM compression techniques (pruning, quantization, distillation) across performance, reliability, and efficiency.
Research identifies a bit-flip vulnerability in shared KV-cache blocks in LLM serving systems, specifically vLLM's Prefix Caching.
Research identifies "Unlearning Aftermaths" in Vision-Language Models (LVLMs) after privacy-driven unlearning, leading to degenerate or hallucinated outputs.
Research identifies a mechanistic explanation for catastrophic loss explosions during low-precision transformer training with Flash Attention.
Research introduces a linearity score (λ(f)) to diagnose neural network input-output behavior, claiming fidelity to models is insufficient for XAI.
Research explores active sequential prediction-powered mean estimation, deciding when to query ground-truth labels versus using model predictions.
Research claims ML-enhanced Monte Carlo outperforms classical methods for some Quadratic Unconstrained Binary Optimization (QUBO) problems.
SeekerGym is a new academic benchmark evaluating AI agents for reliable information seeking, focusing on completeness and bias in retrieval.
New benchmark, MMErroR, evaluates Vision-Language Models' ability to detect and categorize reasoning errors in multi-modal inputs.
Research identifies logit suppression vulnerabilities in LLM safety alignment, enabling manipulation despite current safeguards.
A research survey from arXiv explores methods to address distribution shifts in deep learning models for medical image analysis, enhancing deployment reliability.
Research proposes ASTRA, an automated framework to autonomously discover, retrieve, and evolve LLM jailbreak attack strategies through continuous learning.
CaTS-Bench introduces a new benchmark for evaluating language models' ability to describe time series data across 11 diverse domains.
Research evaluates LLM alignment with human moral values in high-stakes kidney allocation, identifying deviations from human preferences.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion