• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Announcement

Eight hidden-state numbers are claimed to predict 68 later measurements in Hermes 8B

A post says the numbers came from one layer of a frozen Hermes 8B and predicted measurements several layers later, with error about 69–76% lower than a simple average baseline.

Daniel Khashabi 🕊️DK
Mahyar FazlyabMF
Jack Jingyu Zhang @ COLM 2026JJ
7 Sources, 11d ago, first seen 11d ago

TLDR

A user says eight numbers from one layer of a frozen Hermes 8B predicted 68 measurements several layers later. The user reports prediction error about 69–76% lower than a simple average baseline and says targeted edits can pull an LLM’s internal state back on course.

Combined views

2.5K

7 Sources, first seen 11d ago

47 likes1 comments10 saves23 reposts

Combined views

2.5K

7 Sources, first seen 11d ago

47 likes1 comments10 saves23 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Related

OpenScience launches as an open-source agent, with claimed benchmark leads over Codex and Claude Code

A post describes Synthetic Sciences’ OpenScience as a free, Apache 2.0-licensed workbench for any provider’s model. It reports that OpenScience solved 53 of 70 Terminal-Bench-Science tasks, scoring 75.7%.

Commentator Urges Safety Testing For All Capable AI Models Regardless Of License

7 Sources

Rohan Paul@rohanpaul_aiFull paper https://proprioceptiveai.com/From_Internal_Prediction_to_Measured_Control.pdf11d
Mahyar Fazlyab@MFazlyabFirst post on X! 🎉 Excited to share our #NeurIPS2026 paper: “Minimally Invasive Steering of Language Models.” How can we steer LLMs toward high reward at inference time, without fine-tuning or unnecessarily changing their behavior? We introduce MISVO. 🧵11d
Daniel Khashabi 🕊️@DanielKhashabiRT @MFazlyab: Joint work with my student Taha Entesari and collaborators @jackjingyuzhang and @DanielKhashabi! This is part of a broader q…11d
Jack Jingyu Zhang @ COLM 2026@jackjingyuzhangExcited to share our #NeurIPS2026 paper! ✨ We study how to steer frozen LLMs toward higher reward while limiting changes to their output distribution, introducing MISVO 🎯, which optimizes steering vectors with a KL-inspired penalty.10d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    Rohan Paul

    7 Sources

    Rohan Paul@rohanpaul_aiFull paper https://proprioceptiveai.com/From_Internal_Prediction_to_Measured_Control.pdf11d
    Mahyar Fazlyab@MFazlyabFirst post on X! 🎉 Excited to share our #NeurIPS2026 paper: “Minimally Invasive Steering of Language Models.” How can we steer LLMs toward high reward at inference time, without fine-tuning or unnecessarily changing their behavior? We introduce MISVO. 🧵11d
    Daniel Khashabi 🕊️@DanielKhashabiRT @MFazlyab: Joint work with my student Taha Entesari and collaborators @jackjingyuzhang and @DanielKhashabi! This is part of a broader q…11d
    Jack Jingyu Zhang @ COLM 2026@jackjingyuzhangExcited to share our #NeurIPS2026 paper! ✨ We study how to steer frozen LLMs toward higher reward while limiting changes to their output distribution, introducing MISVO 🎯, which optimizes steering vectors with a KL-inspired penalty.10d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet