Evidence-led coverage tied to exact public sources, with separate news bylines and dates, visible uncertainty, limitations and conflicts.
Mathematics · Artificial Intelligence · Research Integrity
A reasoning model and coding agent helped narrow the Grothendieck constant enough to fix its previously unknown tenths digit. The documented research run also showed why producing mathematics, choosing the right research direction and validating a result remain different jobs.
Read the full story
Artificial Intelligence · Uncertainty · Research Methods
A new study finds that an AI model's stated confidence and its internal token probabilities can share a signal—but the same connection may also transmit misleading certainty from training data.
Read the full story
AI Safety · Model Monitoring · Research Methods
A COLM 2026 paper reports that complementary detectors can catch backdoored models when individual monitors fail. Its strongest results—and an internal reporting inconsistency—deserve close scrutiny.
Read the full story
Artificial Intelligence · Medicine · Research Methods
A new language model fits one hospital's oncology notes remarkably well—but low perplexity measures familiarity with text, not safety, accuracy or clinical value.
Read the full story
Mathematics · Computer Science · Algorithms
A 76-page preprint claims the first polynomial speedups over the classic n² and n³ algorithms for 3SUM and all-pairs shortest paths—results that, if correct, would overturn conjectures underpinning years of conditional lower bounds.
Read the full story
Mathematics · Artificial Intelligence · Research Integrity
A frontier model has produced 722 mathematical manuscripts in one release, creating a verification burden at unusual scale. Public release is only the start of turning those claims into knowledge.
Read the full storyReporting preserves source status and exact public version context. Featured researchers and organisations did not commission or endorse this coverage, and no story implies a Haivian audit result.