Neil Shah

Two More Meth­ods for Con­sis­tency Train­ing and Some New Ways to Ap­ply It

• Neil Shah 20 May 2026 9:24 UTC

  6 points

  0

  in reply to: Clément Dumas’s comment on: Seal­ing Con­di­tional Misal­ign­ment in Inoc­u­la­tion Prompt­ing with Con­sis­tency Training

  Hey, we ran it for the base models with the “malicious evil” prompt and see extremely high misalignment rates. For the 3 models that we use in our experiments the average rate is 60.1% (Llama-3.1-8B − 89.8%, Qwen3-8B − 51.4% and Qwen3-32B − 39.4%). So the models after IP + CT(consistency training) are substantially lower than this baseline hovering around (11%-17%).

Seal­ing Con­di­tional Misal­ign­ment in Inoc­u­la­tion Prompt­ing with Con­sis­tency Training

Gemma Gets Help: Miti­gat­ing Frus­tra­tion and Self-Dele­tion with Con­sis­tency Training