Technology & AI
Oracle agent memory cuts token use 10.7x
Oracle agent memory hits 93.8% on LongMemEval with 10.7x fewer tokens than flat-history baselines, per a new arXiv preprint on long-horizon AI agents.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
Self-improving AI agents: survey maps how agents edit themselves
A new arXiv survey formalises how AI agents update their own prompts, memory and tools with minimal human input, and what breaks when they do.
Technology & AI
PriEval-Protect scores hospital privacy risk with legal LLM
PriEval-Protect, a new arXiv preprint, uses a fine-tuned legal LLM and technical data analysis to automate GDPR and HIPAA privacy risk scoring for
Technology & AI
Geospatial AI paper maps path from satellite data to agents
A Google Public Sector preprint proposes splitting geospatial AI between big-compute pretraining and expert fine-tuning, with LLMs as orchestrators.
Technology & AI
Adversarial prompting framework finds encoded attacks bypass AI safety
New arXiv preprint from Microsoft and Walmart researchers finds encoded prompts bypass AI safety filters most, raising the stakes for enterprise AI
Technology & AI
PVDetector catches prompt injection attacks with under 1% miss rate
A new arXiv preprint proposes a training-free detector that reads LLM hidden states to catch prompt injection attacks on purpose-specific agents, reporting
Technology & AI
AI agents burn 91% of tokens on simple edits, study finds
New E3 framework cuts AI agent cost 85% by estimating task difficulty before acting, matching 100% success on a 121-edit benchmark with code released.
Technology & AI
LLM agents hallucinate 36.9% of skill names, study finds
A new arXiv preprint found every agent tested invented non-existent skill names, creating a supply-chain attack path through open skill registries.
Technology & AI
LLM rubrics for AI grading biased toward high scores
First meta-evaluation of LLM-generated rubrics for paper reproduction finds AI graders are overly generous and too detailed, but augmentation helps.
Technology & AI
OpenAI GPT-Red automates red teaming with self-play
OpenAI's automated red teaming system GPT-Red uses self-play to find model weaknesses like prompt injection gaps affecting every AI user.
Technology & AI
LLM judges flip 85% of verdicts when given a reference answer
When AI models grade other AI models without an answer key, they hand out passing marks too freely, new research from Potsdam and Ottawa finds.
Technology & AI
DiffusionGemma transcribes speech in 8 parallel steps, 6.6% error rate
A 26-billion-parameter diffusion model transcribes speech in eight parallel steps, training just 42 million parameters for 6.6% word error rate on