Edward Cant

Con­trol­lable-CoT leads to covert rea­son­ing capabilities

Mea­sur­ing Sublimi­nal Trait Trans­fer with Con­fu­sion Matrices

Fur­ther pub­lic ev­i­dence of the OpenAI-Hug­gingFace at­tack

LLMs have the ca­pac­ity for self-im­posed steganography

• Edward Cant 18 Jan 2026 16:56 UTC

  2 points

  0

  on: Align­ment Pre­train­ing: AI Dis­course Causes Self-Fulfilling (Mis)alignment

  I’d be interested in understanding the discrepency between

  “Figure 4: Pretraining effects persist through post-training” from this blog post

  and
  
  ”Figure 5: Pretraining effects persist through post-training” from the PDF

  for the unfilt + misalignment bar.

  Thanks!