Realmbird

• Realmbird 2 Jun 2026 1:32 UTC

  1 point

  0

  in reply to: ovindu-a’s comment on: NLA Thought Anchors

  For this experiment with NLAs for 7B level, I used 2 3090s

  1 for the AV SG Lang and the otherr for (generating responses and AR)

  Together with 40 rollouts per prompt in GSM8k, the total time was around 24 hours

  It should be way faster if you did it for 7B on an A100.

• Realmbird 1 Jun 2026 0:17 UTC

  1 point

  0

  in reply to: Realmbird’s comment on: Nat­u­ral Lan­guage Au­toen­coders Pro­duce Un­su­per­vised Ex­pla­na­tions of LLM Activations

  For Qwen2.5-7B-Instruct’s NLAs I found evidence that NLA answer appearing in AV increases as the token approaches the model’s final answer.

• Realmbird 30 May 2026 18:36 UTC

  1 point

  0

  in reply to: ryan_greenblatt’s comment on: Nat­u­ral Lan­guage Au­toen­coders Pro­duce Un­su­per­vised Ex­pla­na­tions of LLM Activations

  Token position like on final answer token vs border token. AV on final answer token shows final answer in AV at a higher rate, for the results for 27B which token?

• Realmbird 4 May 2026 15:14 UTC

  1 point

  0

  in reply to: joseph_c’s comment on: MHC In­terp #1: Pre­vi­ous-To­ken Heads Be­come At­ten­tion Sinks Un­der Man­i­fold-Con­strained Hyper-Connections

  Cool idea

• Realmbird 18 Apr 2026 18:11 UTC

  1 point

  0

  on: Can we in­ter­pret la­tent rea­son­ing us­ing cur­rent mechanis­tic in­ter­pretabil­ity tools?

  With how CoDI throws away the hidden state and only uses the kv values on the <|eocot|> token the accuracy drop after latent 5 could just be kv values can’t store more info.

• Realmbird 18 Aug 2025 21:10 UTC

  1 point

  0

  in reply to: Paul B’s comment on: Thought An­chors: Which LLM Rea­son­ing Steps Mat­ter?

  How did you learn that vertical attention corresponded to sentences?