Researchers from Meta AI and the University of Illinois Urbana-Champaign built EvoHarness-RL, a training framework that teaches a small 8-billion-parameter model to decide when to read, update, or consolidate information it gathers from its execution environment during long, multi-step tasks. Rather than following rigid, human-written scripts for tool use, the model learns to manage its own memory and error recovery, reportedly reaching performance comparable to Claude Opus 4.5 on complex agentic workflows.
venturebeat.com
· 2026-08-28
Stanford researchers, working with the Terminal-Bench team and scientists across disciplines, launched Terminal-Bench-Science, a benchmark testing AI agents on real scientific research workflows. The initial version includes 70 tasks spanning life, physical, Earth, mathematical, and engineering sciences, with Claude Opus 5 running Claude Code topping the leaderboard at a 30% resolution rate.
terminal-bench-science.ai
· 2026-08-28
Researchers at the University of Zurich recruited 1,000 volunteers to bury cotton underpants in soil across Switzerland for two months as part of a citizen science project called 'Proof by Underpants.' After digging them up, scientists measured how much fabric had decomposed to gauge microbial activity and overall soil quality, publishing the results in Plants People Planet.
arstechnica.com
· 2026-08-27
Apple is expanding its $50 million Supplier Employee Development Fund by launching a new Apple Education Center in Vietnam, run in partnership with Hanoi University of Science and Technology. The center will offer technical training in smart manufacturing, robotics and automation, plus leadership and soft-skills courses, mirroring a similar program Apple rolled out in India earlier this year.
9to5mac.com
· 2026-08-27
Surgeons at the National Hospital for Neurology and Neurosurgery in London used an AI-powered video analysis system during a real operation to remove a pituitary tumor from patient Rhys Hibbert. The software, developed by UCL researcher Dr. Sophia Bano using footage from hundreds of past surgeries, color-coded critical nerves and blood vessels in real time to guide the surgical team. Hibbert, who faced possible blindness from the 11mm tumor, recovered quickly and has since returned to work.
techspot.com
· 2026-08-27
Biorobotics researchers at the University of Queensland and biomedical engineers at UNSW have created 'paraborgs'—giant burrowing cockroaches equipped with cameras and miniature auto-injection systems. In tests, the insects hit targets with injections 95% of the time when within 15cm, and completed the full navigate-and-inject sequence successfully in 72% of trials.
techspot.com
· 2026-08-27
Palaeoclimatologist Julia Cole and colleagues reconstructed a millennium of sea surface temperatures using Galápagos coral skeletons, published in Science. They found that strong El Niño events have occurred more often over the past century, since industrial-era greenhouse gas emissions began, than in the thousand years before it.
nature.com
· 2026-08-27
A mining-engineering professor in China describes being repeatedly denied underground access for research, despite designing systems meant to operate in those very environments. She notes that laws such as China's 2012 labor-protection statute, along with similar rules in dozens of countries, prohibit women from underground mine work without distinguishing between routine labor shifts and brief, supervised technical visits.
nature.com
· 2026-08-27
More than 600 teams, mostly from Chinese universities and robotics firms, brought humanoid machines to Beijing's second World Humanoid Robot Games from August 22 to 26. Alongside sprint, jump and combat contests, organizers added scenario-based challenges like shelving books and making beds, scoring autonomous performances higher than remotely operated ones. Robots set records in speed and jumping, but researchers noted that everyday chores remain far tougher for machines than flashy athletic feats.
nature.com
· 2026-08-27
Researchers from Google and the University of Illinois Urbana-Champaign tested large language models on variations of the classic Knights and Knaves puzzle, where truth-tellers and liars must be identified from their statements. The models frequently failed when puzzles were slightly altered from familiar training patterns, defaulting to memorized answers instead of reasoning through the new version. A similar pattern showed up on a benchmark called SimpleBench, where humans easily spot subtle twists that trip up even leading AI systems.
technologyreview.com
· 2026-08-26
The piece is a short fiction narrative told by an unnamed narrator recalling his childhood friend Aaron, a brilliant but unfocused classmate who chased fleeting obsessions instead of academics. Years later, after failing to finish school while his friends went to university, Aaron drifts to the margins of the group, working at a local pub where his old friends awkwardly avoid him.
rorz.io
· 2026-08-25
Northwestern University postdoc Reese Richardson discovered that images used by lab supply companies to demonstrate antibody performance had been digitally altered, including removed background noise and copy-pasted data meant to fabricate results. His broader search identified these manipulation issues across marketing materials for more than 17,000 commercially sold antibodies.
arstechnica.com
· 2026-08-25
GoKawiil