Already 5 months old when the scout caught it.
Introspection adapters
What it is
An introspection adapter (IA) is a small LoRA adapter trained to make a language model accurately self-report behaviors it picked up during fine-tuning — even behaviors the fine-tuning process never explicitly labeled. A single IA, once trained, generalizes across many differently fine-tuned versions of a model, so it works as a reusable "ask the model what it learned" probe rather than a one-off audit built for a specific behavior.
Why it's emerging now
Anthropic's alignment team published the technique this week, building on earlier "Diff Interpretation Tuning" work that only generalized to narrow behavior categories. IAs extend that to broad generalization across eight categories of fine-tuning, combining a joint LoRA adapter with a DPO-based refinement stage that trains the adapter to prefer accurate self-reports over hallucinated ones. It's landing now because it addresses a concrete auditing problem: models often can't reliably say what a round of fine-tuning changed about them, which makes catching problematic behavior — or detecting attacks hidden inside a fine-tuning API — much harder.
How it relates to what you know
It's a companion technique to your existing "Red teaming" and "Hallucination" entries — IAs exist to catch fine-tuning-induced behaviors before red-teaming has to find them the hard way, and the DPO refinement stage specifically exists to suppress the confident-but-wrong self-reports that hallucination research studies elsewhere. It's also a narrower, more mechanistic cousin of "model cards": a model card documents behavior after the fact, while an IA is meant to let a model report unexpected new behavior itself.