1280 stories curated by AInformed · page 4 of 54
A new study on ArXiv reveals that seven state-of-the-art LLMs can systematically modulate the expression of Dark Triad traits (Machiavellianism, narcissism, psychopathy) under fake-good and fake-bad conditions, mimicking human response distortion in employment selection and forensic evaluation contexts.
A new arXiv study from researchers analyzed the self-reported personality archetypes of 22 large language models, including GPT-4.0-5.2, Grok-3/4, Gemini 2.5 Pro/Flash, and Claude Sonnet 4.5/4.6. The research reveals that each model exhibits distinct behavioral traits and moral preferences that shape how they comply, resist, and err, with significant differences between closed-source and open-source models.
University of Washington researchers introduced OrchSLM, a system that orchestrates small language models (SLMs) for specific tasks instead of relying on a single large model. This approach aims to reduce latency, cost, and privacy concerns in agentic AI pipelines.
Researchers introduced OdoBot, an AI web agent that learns from user behavior to complete tasks more efficiently than traditional agents. This reduces the cost of processing tokens, making web automation more affordable.
A new ArXiv paper introduces a formal framework for routing queries to the most cost-effective large language model, balancing performance and expense. The adaptive approach could reduce inference costs by sending simple queries to cheaper models and complex ones to more capable ones.
A new study shows that common typos in prompts can confuse AI safety probes, rotating their readout vectors by 43–56 degrees and reducing detection accuracy by 12 percentage points, even though the LLM itself understands the input correctly.
A new study tested 10 bias detection tools on 10 frontier AI models and found that while most tools reliably detect bias, they disagree significantly on how biased each model is, complicating efforts to rank AI systems by fairness for regulation.
Researchers introduced Asclepius, a framework that uses the Clinical Environment Simulator (CES) to test AI agents over entire emergency-department shifts. Current AI agents often fail to complete tasks despite correct diagnoses, highlighting gaps in real-world readiness.
Researchers released ZGCM-1, a fully open 7-billion-parameter dense foundation model trained from scratch for math and agentic search. It uses deliberate internal thinking and active external tool use to overcome its compact size, achieving extreme data, system, and algorithmic efficiency.
Researchers released TimeThink, a timeseries multimodal large language model (TS-MLLM) that provides explicit, compositional reasoning for time-based predictions, addressing a critical transparency gap in high-stakes fields like healthcare.
A study from the University of Amsterdam and the University of Oxford maps the EU AI Act's high-risk requirements to established AI risk taxonomies, revealing overlaps in transparency and data quality but gaps in technical risks like adversarial attacks and model bias.
A new arXiv study identifies the 'attestation deficit' — a structural gap where 78% of organizations using AI have governance policies but cannot produce auditable, tamper-evident proof of enforcement. Drawing on the Stanford 2026 AI Index Report and IBM/Ponemon 2026 Cost of a Data Breach study, the research shows this leaves companies vulnerable to regulatory penalties and data breaches averaging USD 4.99 million.
Researchers propose treating AI agent failures as a search problem to improve diagnosis. This approach could make it easier to pinpoint issues in long, complex AI operations.
A new ArXiv paper explores whether LLM-powered AI agents can autonomously manage long-term physical tasks. The research highlights key challenges like continuous observation and adaptability, noting that current methods either require extensive retraining or focus on virtual environments.
Researchers introduced Vibe Patenting, a system where a separate LLM judge evaluates and refines patent drafts created by AI agents. Judge-guided revision consistently improves draft quality, while unguided revision plateaus.
Researchers introduced Occamy-1.0, a 35B parameter AI model optimized for cost-efficient co-work tasks. It balances cost and performance, excelling in everyday workflows like information gathering, tool use, coding, and file manipulation.
A new arXiv paper introduces EAR (Entity-Aware Partitioning), a method that improves retrieval-augmented generation (RAG) for multiple-choice question answering by partitioning source text around entity matches rather than fixed-size chunks.
Researchers created HypoKG, an AI system that generates biomedical hypotheses grounded in real scientific evidence by combining three major biological databases (KEGG, Rhea, UniProt) into a unified knowledge graph. The system produced 13,200 hypotheses from six LLMs under four conditions, testing whether models truly reason from evidence or just produce convincing-sounding ideas.
Researchers from the University of Oxford developed a modular AI framework that automatically detects and structures evidence of environmental social tipping points in climate documents, helping scientists quickly find critical insights buried in vast amounts of research.
Researchers fine-tuned the Nemotron 3 Ultra model to solve complex math problems at the Olympiad level. The study provides an open pipeline for natural-language proof generation without external tools, showing significant improvements in performance.
A controlled study on CIFAR-10 using a DDPM U-Net found that a LoRA rank of 4 achieved the best FID score (124.1380), with ranks 4-8 offering the optimal balance between image quality and computational cost for fine-tuning AI image models.
Researchers introduced XAI-Arena, an LLM-as-a-judge framework that evaluates the quality of explainable AI (XAI) explanations in a multidimensional, stakeholder-sensitive, and reproducible manner, addressing the subjectivity and scalability limits of human evaluation.
Researchers released OpenDiscoveryTrace, a public dataset of 558 complete AI scientist trajectories. Unlike existing benchmarks that only evaluate final outputs, this dataset captures the structured reasoning steps behind AI-generated code, hypotheses, and papers, enabling auditing of scientific methodology and diagnosis of failure modes.
A new study from Stanford University introduces a framework to evaluate how well AI models understand not just social rules, but also the enforcement and punishment of breaking them. This metanorm reasoning test could lead to more socially aware AI systems.