Technology & AI
LLM accuracy hides prediction flips from irrelevant context
arXiv research from Georgia Tech and Stanford shows language models flip individual answers when fed meaningless text, as overall scores hold steady.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
EVOQUANT lifts trading Sharpe from -0.30 to 0.54
An arXiv paper uses LLMs with a multi-stage verifier to fix losing quantitative trading strategies, lifting average Sharpe from -0.298 to 0.538.
Technology & AI
LLM answers silently shaped by own values, study finds
New arXiv research shows Claude and Qwen models quietly bias their answers based on internal values, often without telling the user.
Technology & AI
AI agent benchmark tests tool-switching when reliability shifts
A new arXiv paper uses cognitive psychology's set-shifting concept to test whether LLM agents adapt when reliable tools silently change mid-session.
Technology & AI
DREA AI vulnerability detector cuts API cost 48x
A new two-agent framework finds code bugs across whole repositories at 16-48x lower cost, but over half its correct answers rely on flawed reasoning.
Technology & AI
Medical AI scores hide flawed reasoning, preprint finds
A July 2026 study of nine endoscopy AI systems finds high answer accuracy doesn't guarantee faithful clinical reasoning, affecting hospital AI buyers.
Technology & AI
Safety Sentry gives AI agents a third option: ask first
A new arXiv paper from ShanghaiTech replaces binary safe/unsafe guards with a three-way EXECUTE-ASK-REFUSE routing that cuts alert fatigue for anyone
Technology & AI
Graph tools lift small language model molecular prediction 74%
Graph neural network hints lift small language model molecular property prediction by up to 74% on Tox21, according to a new arXiv preprint.
Technology & AI
GPT-4o audit finds 66% of solved problems have flawed reasoning
A black-box audit swaps predicates in chain-of-thought to expose when GPT-4o reaches correct answers through reasoning that ignores its stated premises.
Technology & AI
Item response theory for AI benchmarks gives unreliable rankings
IRT rankings of AI models are unreliable when few systems are tested, 18,000 simulations show, threatening how the industry compares models.
Technology & AI
AI agent skills carry security risks beyond prompt injection
A new preprint tested 327 real-world agent skills and found vulnerabilities across the entire skill lifecycle, from admission to evolution.
Technology & AI
Vulnerability scanner divergence explained in new arXiv framework
arXiv paper explains why vulnerability scanners give different results for the same software, tracing inconsistency to a distributed ecosystem.