Aditya Raj

small tar­geted train­ing as a mechanis­tic “am­plifier” for dis­cov­er­ing la­tent/​learned de­cep­tive circuitry

• Aditya Raj 5 Aug 2026 17:52 UTC

  1 point

  0

  in reply to: Brendan Long’s comment on: Chain-of-Thought as Con­tex­tual Sta­bi­liza­tion and As­so­ci­a­tive Retrieval

  Yes, I have to test the hypothesis with bigger models and multiple question sets, I will be posting the results and will surely tag you.
  
  Note: I believe that reasoning is something which is needed even for small tasks(in human context), Even a statement,
  
  ”A=2, B=3, A+B=?”
  “A+B=?, A=2, B=3”
  
  In the 2nd statement, a human will stumble a bit or will think, where is A and B?, to answer the question, which was not same with the LLM.

Chain-of-Thought as Con­tex­tual Sta­bi­liza­tion and As­so­ci­a­tive Retrieval

Why Safety Con­straints in LLMs Are Easily Break­able? Knowl­edge as a Net­work of Gated Circuits