Large Language Models · Research
Perplexity Trains Computer Agent on Real Mistakes With Hint-Guided Self-Distillation
Perplexity trained its Computer agent on real failed sessions, cutting live tool-call failures by 21.2% relative in an A/B test.
Latest news, developments and reporting on Research.
Perplexity trained its Computer agent on real failed sessions, cutting live tool-call failures by 21.2% relative in an A/B test.
A review of 640 contested trust cases published in the Washington University Law Review found that giving one child control of a family trust while siblings also benefit is a leading cause of inheritance litigation, with disputed cases lasting an average of 481 days.
Researchers in London have created the UK's first canine DNA database covering 80 breeds, aiming to help police identify dogs or breeds linked to crime scenes.
A rapid genomic test for brain tumours can diagnose patients during surgery, cutting NHS waits from up to eight weeks to two hours, according to The Guardian.
Leiphone reports on DeepSeek's DSec sandbox, which can run 3 million environments daily, and traces 11 papers signed by Liang Wenfeng.
Donor hearts adopt recipients' biological age, study finds, undercutting the idea that young organs reverse aging.
Tsinghua University researchers have published a paper on Cache-to-Cache, a technique that passes one AI model's internal memory directly into another instead of converting it to text. The authors report 100% to 150% faster collaborative inference, but the method works only with open-weight models.
According to QbitAI, NIO and USTC researchers, with Ren Shaoqing as corresponding author, propose MM-Future, a multi-mode world-action model for autonomous driving planning that generates paired scene-action hypotheses.
New classical method cuts RSA security by orders of magnitude via signature forgery, but widely used keys remain safe.
Newcastle University has appointed oil and gas heir Anant Ambani as a professor of practice in sustainability, drawing criticism from conservationists and academics.
A World Bank-funded experiment in Liberia gave couples equal cash and joint financial planning. It improved finances for those who planned, but among couples already at high risk, intimate partner violence increased, according to results released in July.
诺因发布面向通用具身智能的生成式学习架构GLOW技术报告，称机器人可通过一次人类演示实现跨物体、跨环境、跨任务复用，并在RoboDojo、LIBERO-Pro及Embodied Arena等评测中取得领先成绩。
Tsinghua University and Infinigence AI have open-sourced RLark, a cloud-native platform for managing embodied devices, cloud and edge resources by task. QbitAI reports it cuts robot onboarding to about five minutes and starts cross-cluster tasks within 10 seconds.
Delft University's sub-100g drone uses artificial whiskers and 34KB software to navigate by touch in dark or dusty areas.
A study of 1.4m adults in England found that 59% of hypertension cases are undiagnosed and half treated by a GP remain uncontrolled.
Anthropic says its Bay Area lab found a CRISPR-like enzyme system in bacteriophage DNA, mostly via Claude, with humans running all lab work.
A French-led team has sent a crewed submersible to the Atlantic seafloor where more than 200,000 barrels of low-level radioactive waste were dumped between 1949 and 1982, finding many containers corroded and leaking. Instruments detected cobalt-60 and niobium-94, but radiation levels around the barrels were not high enough to endanger the researchers.
Phil and Penny Knight have donated $1bn to the University of Oregon to create an engineering school in Eugene, believed to be the largest known gift to a flagship university.
Melbourne Institute releases annual Hilda survey snapshot, tracking 17,000 Australians in 9,000 households since 2001.
An international team examined the 2022 Hunga eruption in Tonga, focusing on a rapid caldera collapse that may have amplified its tsunami, as scientists work to gauge when undersea volcanoes turn deadly.
A Gallo-Roman wall found beneath the Île de la Cité may confirm the island, not Nanterre, as the site of ancient Lutetia, France's culture ministry says.
Le Monde reports that an ancient wall has been discovered in Paris and may confirm the city's original settlement.
DeepSeek's new paper, with Liang Wenfeng listed as an author, details DSec, a system that builds more than 5,000 sandboxes per second for agent training and documents the infrastructure and security challenges involved.
Kyutai released two open-weight speech-to-speech models that solve spoken math, reaching 77.1% on GSM8K after RL.