Steering the model from inside: activation engineering and what it reveals

Every other technique in this corpus works on a model at training time, by choosing its data or updating its weights. The 2023 paper adds a lever that works at inference time instead, directly on activations, and this supertheme is organized around that distinction, its consequences, and, with the corpus's fourth paper, its collapse. A behavior becomes a direction in activation space is the representational move underneath everything else: a behavior is isolated as a single residual-stream direction and dosed with a signed multiplier, machinery this theme traces from construction to normalization to proposed refinements. The activation engineering family situates that move historically, reading Contrastive Activation Addition as the latest in a 2022-25 lineage of techniques that read or perturb activations, a lineage whose newest members reach into training as well. Seven alignment worries become seven dials and writing behavior down as A/B questions supply the substance being steered: the corpus's failure modes, operationalized as measurable behaviors, each requiring its own contrastive dataset before it can be dialed. Three levers on one model reads the same technique from outside the network, showing steering composes with system-prompting and finetuning rather than replacing them, at a fraction of their computational cost. Evaluating steering supplies the accounting that claim needs: behavior-score evidence that steering worked, and capability-floor evidence that nothing else broke. The last three themes turn the lever around and use it as an instrument. Steering doubles as interpretability reads steering artifacts as measurements of where and how behaviors are represented, evidence for the linear representation hypothesis. Base versus chat: what RLHF changes inside extends that instrument to RLHF itself, finding it reweights behavior while leaving mid-layer representations largely intact. Elicitation as audit closes the earlier supertheme's arc by inverting the lever once more, from steering a behavior on to using that capability to probe whether a model can be made to misbehave at all. Two further themes, added with the corpus's fourth paper, push the lever in the two directions that opening inference/training split didn't anticipate. Steering grows out of inference time carries the vector across that boundary directly: the same object CAA injects at inference time, the 2025 paper injects during finetuning instead, canceling drift a training run has already introduced when applied post-hoc, or preventing that drift from taking hold in the first place when applied preventatively, layer by layer, and weighed against regularization, prompting, and concept-ablation alternatives, with a dedicated fact-acquisition case study confirming the intervention cancels an unwanted trait without also canceling the run's intended learning. Projection as early warning turns the same construction into a read-only instrument instead of a write one: the dot product of a model's activations with a trait direction predicts trait expression before generation happens, giving the supertheme's steering apparatus a second use as a monitor, watching for prompt-induced shifts at deployment time and weight-induced ones during training, reading the persona rather than writing it.