Research
Exploring the latest AI research coming out of Asia's universities and labs.
Google's ERA Writes Expert-Level Scientific Code — and Beat the CDC's COVID Forecaster
Published in Nature, Empirical Research Assistance uses Gemini and tree search to author research software, discovering 40 new single-cell analysis methods and topping human leaderboards.
SkillOpt: A Text-Space Optimizer That Lets AI Agents Improve Their Own Skills
A trending new paper proposes an 'executive strategy' for self-evolving agent skills — stable updates with zero deployment inference overhead — as the field races toward agents that learn on the job.
Apptronik Opens 90,000-Sq-Ft 'Robot Park' to Feed Real-World Data Into Gemini Robotics
The Google-backed humanoid maker unveils Apollo 2 and a flagship training facility built on a continuous learning loop with DeepMind — where simulation meets slipping feet and aging hardware.
Nature Publishes Twin AI 'Research Assistant' Systems That Design and Interpret Real Experiments
Google DeepMind's Co-Scientist and FutureHouse's Robin — both peer-reviewed this week — proposed drug candidates for leukemia and macular degeneration that held up at the lab bench.
DeepMind's AMIE Goes Multimodal: Diagnostic AI Now Reads Scans, Labs and Histories in Nature Medicine Study
The medical dialogue agent that outperformed physicians on text-based diagnosis can now reason over medical images and lab results — moving a step closer to clinical reality.
Meituan Open-Sources LongCat-2.0: A 1.6T-Parameter Coding Model Trained Entirely on Chinese Chips
The MIT-licensed agentic coding model was trained from scratch on a 50,000-card cluster of domestic accelerators — the strongest evidence yet that China's AI stack can reach near-frontier scale without Nvidia.
Meta Releases CWM, a 32B Open-Weights Model That Learns Code by Simulating It
The Code World Model treats programs as environments to simulate rather than text to predict — and Meta is releasing the weights for researchers to build on.
NatureBench Asks Whether Coding Agents Can Beat Published Science — Mostly, They Can't
The new 90-task benchmark distilled from Nature-family papers finds the strongest frontier agent surpasses published state-of-the-art on just 17.8% of tasks — succeeding by translation, not invention.
New Paper: Combinatorial Hardening Exposes Compositional Failures in Frontier LLMs
Researchers show that deterministically transforming multiple-choice questions into higher-order logical judgments cuts frontier model accuracy by 31-56% — even on material the models otherwise ace.
FrontierMath v2: Epoch's Error-Corrected Gauntlet for AI Mathematicians
After an audit found flaws in 42% of its problems, Epoch AI has rebuilt its hardest math benchmark — 338 problems, including 43 research-level Tier 4 questions that take experts days to solve.
Three Weeks On, Kimi K2.7-Code Shows the Promise — and the Problem — of Vendor Benchmarks
Moonshot AI's open-weight coding model posts double-digit gains and claims a win over Claude Opus 4.8, but every headline number comes from Moonshot's own benchmark suite.
ByteDance's Seed 2.1 Pro Claims Wins Over Claude Opus 4.6 at 80% Lower Cost
The agentic enterprise models lead on Terminal Bench 2.1, SWE-Pro, OSWorld and MMMU-Pro by ByteDance's count, with 256K context and pricing from $0.85 per million input tokens.