Skip navigation
🚀 Introducing DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
0:00 / 0:40

🚀 Introducing DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

Ling Yang

@LingYang_PU5K followers

Visit on X
727 likesDownloadOriginal post

55,712 views

🚀 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. 📄 💻 🔴

17 Comments

Comments from the original post on X (Twitter)

  • MANIC_PANIC777's profile picture

    @MANIC_PANIC777

    @LingYang_PU It'd be super cool if you could actually build a hypergraph memory system with IA³ adapters, coupled with this . That way you could have a full autodidactic lifetime learning model.

    1 like

  • Deep's profile picture

    @deeponailabs

    Check how often random hypotheses match later papers, that sets your baseline

    1 like

  • ChainlinkR's profile picture

    @b4lee1981

    huge release 🔥

  • Randy Butternut's profile picture

    @vikneshmanohar

    Wait how are you testing against GPT-6?

  • Gabriele Lima's profile picture

    @samjo_00

    recursive learning for scientific discovery is the right direction imo

  • BreezeOg's profile picture

    @BreezeOg1

    hypotheses that actually got published later 👀 ok that's cool

  • basedcapital's profile picture

    @thebasedcapital

    the 6.3% number is cool but what about the other 93.7%. do the misses just get tossed or fed back as the bad ones?

  • Lukas Punished's profile picture

    @n_guseva67

    Congrats Ling!

  • Shardul's profile picture

    @ShardulAggarwal

    Very interesting! Congrats

  • callum_mtb's profile picture

    @callum_mtb

    6.3% vs 2.5% is pretty wild if real

  • TSOT's profile picture

    @Cattch11Tsot

    thought this was hype until I read the paper, benchmark setup is actually solid

  • SynthGen Studio's profile picture

    @RuthWinstanley

    27B beating Opus on hypothesis generation is insane

  • BeastTitanHunter's profile picture

    @Titan1Beast

    Hmmmmmmm

  • Marc X's profile picture

    @FaijanK21499117

    Ok the 2.3% baseline for claude opus makes sense given how conservative it is

  • Ivrqen's profile picture

    @Ivrqen

    filtering the 94% noise is where all the actual human time goes

  • 👺Tr_Bolado👺's profile picture

    @Tr_Boladoo

    what paper dataset was used for validation?

  • mo's profile picture

    @Allmight_1919

    how do you define matching later published work though