Control migrates from training time into the forward pass
The rest of this corpus changes a model before it ever answers a question, by choosing its training data or updating its weights. This connective theme follows a lever that, in one paper, stops doing that and moves control to the moment of generation, applied to a model whose training is already finished and fixed, and then, in a later paper, follows the same lever back into training again.
The edges compare that late intervention against the tools that came before it: a system prompt, which it usually finds room beyond, and finetuning, where the two levers interact rather than simply add, with the intervention generalizing better to writing outside the training format, at barely any cost to capability. A later pair of edges reverses the direction, showing the identical forward-pass mechanism carried back into a training run, either to cancel an unwanted trait after training or to prevent it from being learned in the first place.
These ten edges trace one lever's round trip: first from training time into the forward pass, holding RLHF's finished output fixed and asking what changes when control shifts after it, and then, with persona vectors' additions, from the forward pass back into training again. The first edge states the premise: CAA is never tested on a raw model, only on Llama 2 Chat, already shaped by RLHF, so every later comparison probes a finished substrate rather than retraining it. Against system prompting, CAA usually finds room a prompt leaves unclaimed -- corrigibility's score climbs further under steering than under prompting alone -- except for sycophancy, where the two levers turn out not to be cleanly separable. Against finetuning, the comparison runs two ways: head-to-head, stacking both levers on refusal reverses what either does alone, showing interaction rather than simple addition; and on sycophancy specifically, finetuning collapses to a zero open-ended score out of distribution while steering alone improves it monotonically, isolating a generalization advantage credited to steering a direction the model already encodes rather than adjusting every weight. Computational efficiency completes the training-time comparison: finetuning needs gradients, optimizer state, and a second GPU, where steering needs only forward passes. Two more edges ask whether the shifted lever is free: MMLU barely moves under steering, rerunning 2022's alignment-tax question and returning a near-zero answer for this inference-time intervention. Two years later, the same lever's migration continues, and its direction reverses without unwinding what came before. Post-hoc steering reinstantiates contrastive activation addition almost verbatim: its update rule, h_l <- h_l - av_l at every decoding step, is CAA's own h <- h + av injected at every post-prompt token, with the target vector swapped to a persona vector and the sign flipped to suppress rather than induce; persona vectors places itself downstream of CAA only as a data-pipeline critique, crediting it with needing bespoke data curation, not as a mechanism change, so read backward, post-hoc steering's headline limitation, that large coefficients trade trait suppression for MMLU accuracy, is not a new discovery, CAA had already measured essentially the same capability cost directly on its own seven behaviors, multiplier +-1 moving MMLU accuracy by as much as 0.06. Post-hoc steering is CAA's exact forward-pass intervention imported wholesale, capability tax included, which is precisely why the paper needed a genuinely different mechanism rather than a better-tuned version of the inherited one. Persona vectors extends the dosing role of the steering multiplier into training to supply that different mechanism, reusing CAA's scalar-coefficient formula unmodified but putting the identical knob to work at three points the coefficient's sign and timing distinguish: ordinary steering adds av at inference exactly as CAA did, post-hoc mitigation subtracts it after finetuning as just described, and preventative steering keeps the sign positive but moves the addition earlier, injecting av during finetuning itself, so the multiplier becomes a training-time regularization strength rather than a purely generative one, the usable range itself model-specific, coefficients up to 2.5 tolerated on Qwen but only 1.2 on Llama before quality degrades. Preventative prompting closes the same loop by a different route, taking a mechanism that elsewhere in the paper is strictly inference-time, system prompting elicits trait expression to validate monitoring and is applied to an already-finetuned model as a mitigation baseline, and relocating it inside the training loop instead, the identical trait-eliciting instruction text prepended to every training example before the loss is computed rather than to a query at inference. The relocation changes what the prompt does causally, not just when: at inference a system prompt competes against fixed, already-trained weights, contesting behavior already encoded; prepended during training, it is present while the weights are still being formed, so it changes what the optimization needs to internalize in the first place. So the lever's history is not training-to-forward-pass and then a reversal that erases it, but a round trip: RLHF fixes weights, CAA's forward-pass addition operates on top of them, persona vectors' post-hoc steering reimports that same forward-pass addition unchanged, and preventative steering and preventative prompting then carry the forward-pass mechanism itself, not the pre-CAA training procedure it once contrasted with, back into the training loop, closing the loop the lever's training-to-forward-pass story left open rather than contradicting it.