ResearchStudy finds AI agents can do research engineering but fall short of open-ended scientific creativityA multi-institution team led by Peter Kirgis and Sayash Kapoor at Princeton tested AI agents on open-ended research tasks drawn from two unpublished NeurIPS 2026 submissions and found the agents could run experiments and implement engineering steps but failed to produce publishable, creative research.5 min read
ResearchQwen 3.8 27B achieves surprisingly high 52-point score on the Artificial Analysis Intelligence IndexThe focus is the Qwen 3.8 27B language model, which scored 52 points on the Artificial Analysis Intelligence Index.1 min read
ResearchTool use does not replace larger language modelsThe author argues that although tool use (web browsing, code execution) allows smaller models—such as a 1-billion-parameter “cognitive core”—to solve many tasks, this is not sufficient: fast, natural, and reliable responses require internal knowledge.1 min read
ResearchQuantization-aware distillation yields faster Nemotron 3.5 Lightning NVFP4 checkpointNVIDIA applied quantization-aware distillation (QAD) to the open Nemotron 3.5 Lightning model to produce an NVFP4 checkpoint that reduces memory footprint while restoring accuracy lost to aggressive quantization.5 min read
ResearchRole Anchor: a training method from MIT and Harvard to prevent module 'role drift' in modular AIResearchers at MIT and Harvard describe “role drift,” a failure mode in multi-module LLM pipelines where individual components bypass their assigned tasks while overall end-to-end accuracy rises.6 min read
ResearchDiG-bench: a 70-game benchmark measuring AI discovery and creative intuitionDiG-bench (Discovery in Games) is a new 70-game benchmark that evaluates how well AI systems uncover hidden rules in novel interactive environments through exploration.4 min read
ResearchSmartphone photos can estimate body composition and help predict insulin resistance riskResearchers developed PhotoScan, a deep learning framework that estimates body fat percentage, Android-to-Gynoid (A/G) ratio and Visceral-to-Subcutaneous (V/S) ratio from standard 2D smartphone photos.4 min read
ResearchAutonomous AIs Match Much of Research Labor but Fail to Invent Novel MethodsPrime Intellect ran 18 frontier models in eight-day, offline sandboxes executing 153 autonomous experiments on a nanoGPT optimizer speedrun.2 min read
ResearchOpenAI awards 14 research grants to study AI's economic and social effectsOpenAI selected 14 research projects to investigate the economic and societal impacts of artificial intelligence, awarding a total of $1 million in cash plus up to $1 million in model credits.2 min read
ResearchAI-driven single-cell analysis reveals genetic and proteomic heterogeneity in clear cell renal cell carcinomaResearchers led by Horváth Péter’s Lendület Mikroszkópos Képelemzés és Gépi Tanulás Kutatócsoport used artificial intelligence and an automated single-cell facility to measure both gene activity and protein composition of tumor clones from clear cell renal cell carcinoma at single-cell resolution.3 min read
ResearchAssessment of GLM 5.3: a cheaper tool for cybersecurity protectionResearchers evaluated the GLM 5.3 language model's cybersecurity capabilities and found that its lower operating costs could make it useful for defensive security tasks.1 min read
ResearchEval harness shows LLMs' highest confidence often corresponds to incorrect answersAn evaluation harness that scores model outputs against synthetic ground truth reveals that qualitative review—judging whether outputs merely 'sound right'—misses systematic, high-confidence errors in LLM-assisted enterprise tools.5 min read
