Subhash Kantamneni

• Subhash Kantamneni 9 May 2026 2:05 UTC

  4 points

  0

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

  We’re fans of the Consistency Lens and shout it out in our acknowledgements!

• Subhash Kantamneni 9 May 2026 2:04 UTC

  1 point

  0

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

  Yeah we present a variety of such transformations in our experiments on steganography in the subsection Measuring behavioral properties of NLAs. Seems reasonable to explore adding these into the training loop as a preventative measure.

• Subhash Kantamneni 8 May 2026 16:32 UTC

  5 points

  0

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

  I agree, I generally like the setting of multihop reasoning, and it’s one of the first we looked at to build confidence that NLAs are doing something reasonable. For instance, this result used an early version of NLAs on Haiku 3.5.

  

  It’s worth noting that NLAs struggle with numbers (they’re the type of specific detail that get confabulated). But I’m a little confused, the capital of France is Paris right? So the answer should be (8-5)*2=6 so we shouldn’t see 13 as an intermediate? We also in general might need to run the NLA on more token positions.

Nat­u­ral Lan­guage Au­toen­coders Pro­duce Un­su­per­vised Ex­pla­na­tions of LLM Activations

Ac­ti­va­tion Or­a­cles: Train­ing and Eval­u­at­ing LLMs as Gen­eral-Pur­pose Ac­ti­va­tion Explainers

Scal­ing Laws for Scal­able Oversight

Take­aways From Our Re­cent Work on SAE Probing

Lan­guage Models Use Tri­gonom­e­try to Do Addition

SAE Prob­ing: What is it good for?