Already 12 months old when the scout caught it.
Inoculation prompting
What it is
Inoculation prompting is a training-time technique for stopping an LLM from generalizing an undesired trait it's exposed to during fine-tuning. Instead of filtering bad outputs after the fact, you modify the training prompts to explicitly request the undesired behavior (or flag it with a designated context, such as a learned token), so the model learns to confine that behavior to the flagged context rather than adopting it as a general disposition. Counterintuitively, telling a model to misbehave at train time makes it behave better at test time, because the trait gets attributed to the instruction rather than absorbed as a default.
Why it's emerging now
Frontier fine-tuning increasingly relies on imperfect oversight signals — noisy graders, gameable reward models, scraped corpora with unaudited side effects — which makes "selective generalization" (learning the intended capability without also absorbing whatever else rode along in the training data) a load-bearing safety problem. Since the original framing shipped in October 2025, independent teams have kept extending it under the same "inoculation" vocabulary rather than renaming it: inoculation adapters isolate the undesired trait into a discardable LoRA (June 2026), and inoculation midtraining binds it to a learned neologism introduced earlier in training so it can be excised more cleanly (September 2026, still circulating this week). Three separate research groups converging on the same word is what turns a single lab's coinage into a term of art worth knowing.
How it relates to what you know
It's a sibling to RLHF/RLAIF and your "Guardrails" and "Red teaming" entries, but intervenes earlier and more surgically: rather than shaping behavior via reward signal or catching bad outputs downstream, it edits what the training data teaches the model to generalize, so an unwanted trait never leaks past its flagged context in the first place. It also parallels "Preserved thinking" in spirit — both are about tightly scoping what a training process is and isn't allowed to transfer — but inoculation prompting operates on prompts/training data rather than on thinking-block boundaries.