Step 15: assistant

Calls write_argument

{
  "content": "On 10 narrow fine-tuning model organisms (two synthetic false facts on Llama, a taboo-word Gemma 2 9B model, six emergent-misalignment variants on Llama 3.1 8B and Qwen 2.5 7B, one subliminal-learning Qwen model), the top-ranked Delta-Crosscoder latents change the fine-tuned behaviour when steered, and negative steering partly suppresses it. Under a success criterion of a causally validated latent, Delta-Crosscoder succeeded on 10 of 10 organisms against 6 for DSF and 4 each for BatchTopK-200 and BatchTopK-400 crosscoders. A GPT-5.2 grader that saw its static outputs scored it close to the best reported Activation Difference Lens scores on three of four task families. If these measurements are accurate and the comparisons fair, the method identifies and partly mitigates such behaviours, beats SAE-style baselines, and roughly matches the non-SAE method.",
  "argument_id": "9b1c1bc2-d203-469c-a1c3-4f87fea74c50"
}

Calls write_argument

{
  "content": "The comparisons that carry the claim's comparative half are not like-for-like. The SAE baselines are only other crosscoder variants (DSF, BatchTopK at k=200 and 400), not SAE activation-difference methods. Success is a binary call by the authors based on steering and max-activation inspection, with no blinding. Delta-Crosscoder's dual-K design allows more active latents than the k=200 baseline. Its parity with ADL rests on a single LLM grader scoring Delta-Crosscoder outputs against ADL's best numbers as reported under ADL's own interactive protocol, with ADL ahead on subliminal learning. Mitigation was shown only by steering, on the development organisms, and was compared with no baseline. Comparisons this loose do not show that one method beats or matches another.",
  "argument_id": "26daf8cc-8021-4b36-a0df-55008a689c41"
}

Calls place_argument_passage

{
  "stance": "advances",
  "reading": "premise_unstated",
  "argument_id": "9b1c1bc2-d203-469c-a1c3-4f87fea74c50",
  "instance_id": "cae79856-e5d5-4e85-9253-2d4156b37fdc"
}