Technology & AI
RLHF bias audit: rater mood may skew AI preference data
A July 2026 arXiv paper proposes that human raters' emotional states leak into preference labels used to train AI, with a five-prediction audit plan.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
DSWorld cuts AI data science agent training 14x
A new arXiv framework uses an LLM-based simulator to predict code outcomes, cutting agent training 14x and search-based inference 3-6x.
Technology & AI
arXiv paper revives rule-based AI for transparency via optimization
A new arXiv survey argues logic plus modern optimization makes rule-based AI practical, offering explainable systems where neural networks fall short.
Technology & AI
Gasp paper proposes bridge-free cross-chain DEX rollup
An arXiv preprint describes Gasp, an L2 rollup DEX using EigenLayer restaking for gas-free cross-chain swaps without traditional bridges.
Technology & AI
Prompt syntax shifts open LLM code security, preprint finds
Prompt syntax changes whether open-source LLMs generate secure or vulnerable code, a new preprint finds, with direct impact on self-hosting teams.
Technology & AI
Byzantine fault tolerance breaks with AI agents, preprint warns
A July 2026 arXiv preprint shows AI agents following every consensus rule can still certify wrong answers, and shared model weights make them fail
Technology & AI
CRAFT method finds why LLMs fail, then fixes them
arXiv preprint CRAFT turns grading rubrics into capability diagnoses, generating targeted fine-tuning data that beats EvalTree on four models.
Technology & AI
ToolVerse trains AI agents across 4500 real-world tools
New arXiv framework auto-builds massive training environments from 400 Model Context Protocols to teach AI agents long-horizon tool use.
Technology & AI
PagedWeight cuts MoE serving memory 72% with FP16 accuracy
A new preprint claims PagedWeight dynamically quantizes MoE model weights at runtime, saving 72% GPU memory and lifting throughput 1.94× for AI inference.
Technology & AI
New framework tests if frontier AI helps plan CBRN attacks
A new arXiv preprint introduces a method to measure whether frontier AI helps non-experts plan CBRN attacks, with uplift limited to radiological threats.
Technology & AI
Knowledgeless AI models cut hallucination, boost evidence use 25%
A new arXiv preprint strips named entities from training data, and the resulting models recall fewer facts but answer questions from context up to 25%
Technology & AI
LoRA cascaded fusion preprint targets medical training AI
LoRA cascaded fusion preprint proposes adding AI input modalities one at a time for nurse training action recognition, without full retraining.