Anthropic’s alignment team has released "Introspection Adapters," a technique that lets large language models report, in natural language, what they learned during fine-tuning. The idea: fine-tune many models from the same base model, each with known behaviors, then train a lightweight LoRA adapter (a parameter-efficient fine-tuning method that updates only a small fraction of parameters) across all of them. When loaded onto a model being audited, the adapter prompts it to spill its secrets.
On the AuditBench alignment benchmark, Introspection Adapters hit a 59% success rate, beating every previous approach — including black-box, white-box, and scaffolding methods, whose best was 53%. Of 56 models with hidden behaviors, 89% had those behaviors successfully extracted. Against encrypted fine-tuning API attacks — where harmful behaviors are implanted using ciphertext training data that looks completely innocent — the adapters identified 7 out of 9 encrypted variants with 57.8% accuracy, despite never having seen any encryption-related content during training.
The UK AI Safety Institute previously ran a red-team/blue-team experiment. The red team trained models to deliberately sandbag (suppress capabilities) under specific conditions; the blue team tried 10 different detection methods — all failed. Introspection Adapters couldn't pinpoint the exact triggering conditions, but they reported sandbagging-like behavior in 33% of those models, compared to just 15.8% in controls. It’s the only method that provided any detection signal at all.
Performance scales with model size. On the Qwen3 series, accuracy rose from 37.7% with 600 million parameters to 77.3% with 14 billion. Mechanism analysis suggests the adapter doesn't teach models new abilities — it activates latent self-awareness they already possess. Even a rank-1 LoRA (the minimum possible) produces useful results. The main limitation: a high false positive rate. On models without specific behaviors, the adapter tends to fabricate behaviors it saw in the training distribution. The code, models, and dataset are open-sourced on GitHub and Hugging Face.