Science Open source

DiscoveryRSI: a model that recursively learns to discover valuable hypotheses

Ling Yang introduced DiscoveryRSI, which matches 6.3% of later published work versus 2.5% for GPT-6 and 2.3% for Claude Opus 5, with code and a live demo.

View on X
391likes
55reposts
340saves
24Kviews
Made with
DiscoveryRSI-27B, GPT-6, Claude Opus 5
Role
Incoming Assistant Professor, Peking University
Posted
Ling Yang@LingYang_PU

🚀 Introducing DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

6.3% of DiscoveryRSI-27B's hypotheses match later published work, vs 2.5% GPT-6 & 2.3% Claude Opus 5.

📄 t.co/aBpR5GcChg
💻 t.co/XjUmq6xt6a
🔴

View on X
Still from the video

More in Science

Thirty games where two models negotiated to avoid civil warScience@spakhm1.5K114719
A robot's phone video became a full-dynamics digital twin in an hourScience@keplerccccc41849314
AI proving classic distributed-systems protocols in TLA+Science@tianyin_xu27635231
An AI bug-bounty pipeline earned $5,000 while its owner did nothingScience@AdamShao46439499