AI Research
Nov 2, 2026 · 6 min read
AI Translation Hallucination Reaches 60%, Exposing Patient Safety Risks
LLMs hallucinate in translation at rates between 33% and nearly 60%, according to a 2025 Alibaba study covering…
AI Research
Nov 1, 2026 · 4 min read
AI Models Independently Refute 27-Year-Old ETH Conjecture
OpenAI and Anthropic each announced on October 26, 2026, that their models had independently produced machine-verified proofs refuting…
AI Research
Nov 1, 2026 · 5 min read
JoBS Co-Optimizes LLM Data and Model Config Without Full Training Runs
Four new frameworks presented at ICML 2026 and ICLR 2026 are converging on the same idea: stop running…
AI Research
Oct 29, 2026 · 3 min read
WorkFactory Automates 112,000 AI Tasks, Elevating WorkPilot Performance
Getting AI agents to do real professional work reliably means training them on real professional tasks, at scale,…
AI Research
Oct 26, 2026 · 5 min read
Hong Kong Researchers Unveil 500 Million Persona Simulation
Running a credible economic simulation has always meant choosing between scale and granularity. A University of Hong Kong…
AI Research
Oct 22, 2026 · 4 min read
How Nian Ran’s Protocol Diagnoses LLM’s ‘Principle-Level’ Optimization Traps
In 78.3% of 161 LLM optimization runs examined by Nian Ran et al. from the University of Manchester…
AI Research
Oct 20, 2026 · 4 min read
Warcraft III’s ‘Insane’ AI Falls to Self-Play RL on One Desktop
Warcraft III's "insane" AI sees the entire map, ignores the fog of war and collects double the normal…
AI Research
Oct 19, 2026 · 7 min read
56.5% Score Drift: ICLR 2026 AI Peer Review Flaws Exposed
The best AI model tested could only reproduce its own score for just over half the papers it…
AI Research
Oct 17, 2026 · 4 min read
Microsoft’s PBC Framework Doubles 2-6 Week Weather Forecast Accuracy
A week of overlapping AI weather announcements produced one result that stands out: a machine-learning model predicted a…
AI Research
Oct 13, 2026 · 5 min read
35% of AI Chatbots Fail Mental Crisis Interventions
In about 35% of test conversations, AI models tested by Scale AI in October 2026 recognised a user…