The same technique, cited or deployed twice, doing different work each time
Across the corpus, a single mechanism sometimes gets asked to do more than one job, and comparing the jobs is informative. Cooperative inverse reinforcement learning is the clearest early case: it solves the shutdown problem and a demonstration-learning problem at once, purely as a side effect of never fixing a reward or policy in advance.
Inverse reinforcement learning shows the same pattern, repurposed once against reward hacking and once against unsafe exploration, and principle ensembling shows a more surprising version, doing nothing measurable at one training stage and paying off directly at the next. The page also follows two borrowed statistical tools reused across unrelated methods, and closes on the corpus's newest instance: a steering vector narrowed to a personality trait, then handed two further jobs beyond the original inference-time nudge, predicting behavior from a prompt's activations and steering during training itself rather than only after.
Across the corpus, a single mechanism gets asked to do more than one job, and comparing the jobs is informative. Inverse reinforcement learning is repurposed for use demonstrations after already being proposed under reward pretraining, the identical tool once defending against reward hacking and once bounding unsafe exploration. Cooperative inverse reinforcement learning supplies a candidate mechanism for the shutdown problem and, separately, is the interactive counterpart of use demonstrations, its single goal-inference structure solving two nominally unrelated problems as a free side effect of never fixing a reward or policy in advance. Ensembling principles diversifies use of SL-CAI principles at the supervised stage, where it fails to move the harmlessness score it seems aimed at but pays off one step downstream in RL exploration; the identical technique instead makes the feedback model more robust when applied at the RL stage, a direct payoff rather than a deferred one. The ensembling move recurs a third time, five years earlier, doing a third job again: the 2017 paper's reward-predictor ensemble supplies the disagreement signal that uncertainty-based query selection runs on, three independently bootstrap-trained predictors queried for their variance rather than diversified for coverage or hardened for robustness, an epistemic use of the same multiple-model trick that neither Constitutional AI deployment repeats. A direct-revision ablation tests the necessity of critique in critique-and-revision and finds a weak effect at scale, yet the paper keeps the step anyway, for the transparency a written critique provides. The mean-difference method recurs a further time outside the corpus's own papers rather than across them: already used by Zou et al. (2023) for representation extraction and by Li et al. (2023) for Inference-Time Intervention on truthfulness, the identical difference-of-means arithmetic supplies the entire extraction rule for CAA's steering-vector construction, so what CAA adds is not new mathematics but a new interface to it, a specific dataset, layer, and token position fed into a borrowed operation. PCA joins the corpus's other imported, off-the-shelf statistical tools, applied via a standard package rather than reimplemented, projecting high-dimensional activations down to two components; that so blunt and general a technique is sufficient to make refusal-behavior clustering visible by eye is itself evidence that CAA's behavioral directions are simple linear features rather than complex, entangled ones. Two years later, the corpus's steering-vector concept itself gets narrowed and then re-widened this same way: persona vectors specializes the generic steering-vector concept for character traits specifically, evil, sycophancy, and hallucination as three main cases plus optimism, impoliteness, apathy, and humor in Appendix G, and having narrowed the target, puts the resulting object to jobs the generic concept was never built around. Beyond the inference-time nudge steering vectors elsewhere supply, persona vectors confirms the same direction reads behavior it did not cause, projecting a prompt's final-token activation onto the trait direction predicts the trait expression of a response not yet generated, correlating at r=0.75-0.83 across system-prompt and many-shot elicitation, and the same vector is injected during training itself through preventative steering rather than only at inference. Where the corpus's other reuses of a single mechanism keep the same job specification and move it to a new target, persona vectors keeps the same target, a scoped subclass of steering vector, and gives it two new jobs, a monitor operating before generation begins and a regularizer operating during weight updates, beyond the single lever the general concept is otherwise centered on.