The steering vector doubles as a measuring instrument
A tool built to change a model's behavior does not have to stop being useful once the changing is done; the same object can be read, rather than applied, as evidence about what the model was already doing. This connective theme traces edges where a steering vector, or the activations it is built from, does double duty as a measuring instrument.
Some edges stay close to the vector itself, checking whether it acts like a detector when nothing is being added and whether vectors from nearby layers resemble each other closely enough to be called portable. Others put it to a specific use, comparing vectors from before and after RLHF to see what that training changed, or reading a sudden separation in activations as evidence a behavior is one direction. The corpus's newest paper pushes the pattern further, running several quantitative correlational tests of the same idea and decomposing one trait direction into smaller, interpretable parts.
Twelve edges show the same object doing double duty: built as a control mechanism, it also turns out to be a measuring instrument. Token-level cosine similarity is the cleanest case -- take the dot product between a steering vector and ordinary forward-pass activations, no addition at all, and the vector reads behavior off individual tokens as a passive detector. Inter-layer similarity and transfer asks whether that reading generalizes across depth, and finds vectors from nearby layers cluster together and a vector built at one layer still steers behavior applied at another -- construction captures something portable, not a layer-specific artifact. Base-chat representation comparison turns the same read-twice logic on RLHF itself: because Llama 2 Chat is the RLHF-tuned counterpart of Llama 2 with nothing else changed, comparing vectors extracted from each becomes a clean measurement of what RLHF left alone internally, revealing a mid-layer band where the two representations stay unusually close. Behavioral clustering supplies a third reading: PCA on activations shows a behavior emerging suddenly at one residual-stream layer, offered as direct evidence for the linear representation hypothesis, which in turn is the same hypothesis licensing the vector arithmetic steering depends on. The red-teaming edge closes the loop: the identical vector, aimed at eliciting rather than suppressing, becomes a probe for what RLHF failed to remove. The pattern recurs a final time, in 2025, at higher resolution than any earlier instance. Persona vectors quantifies empirical support for the linear representation hypothesis the same way behavioral clustering offered PCA evidence for it, but where the earlier evidence was largely qualitative, visible clustering under projection, persona vectors runs three independent quantitative correlational tests against the same hypothesis: steering along the extracted direction reliably shifts trait scores across layers and coefficients, projecting a prompt's final-token activation onto the direction predicts the trait score of a response not yet generated at r=0.75-0.83, and projecting a finetuning-induced activation shift onto the direction predicts post-finetuning trait expression at r=0.76-0.97, markedly higher than cross-trait control correlations of r=0.34-0.86. The persona vector itself then submits to the same doubling the corpus's other instruments do: it admits a fine-grained decomposition via SAE decomposition of persona vectors, the paper's own limitations section calling the vector a coarse average that may miss fine-grained behavioral distinctions, a claim the decomposition makes legible by ranking Qwen's SAE features against the evil vector and surfacing five separate sub-concepts among the top matches, insulting or derogatory language, deliberate cruelty and sadism, jailbreak-related content, discussion of bad people or actions, and malicious code, no single feature's cosine similarity exceeding 0.336, confirming the coarse vector is a weighted composite of several independently steerable levers rather than a restatement of one dominant SAE feature. Projection-based persona monitoring reads instead of writing to that same vector: the vector is built and layer-selected purely for steering effectiveness, Appendix B.4 sweeping every layer at a fixed coefficient and choosing whichever elicits the highest trait score, yet monitoring takes that steering-optimized vector, at that same layer, and reads it with a plain inner product against the last-prompt-token activation, no addition or subtraction to the residual stream at all, a capability the extraction pipeline never targets or guarantees but that the paper discovers downstream once the vector already exists. That return to reading revives the reading half of representation engineering, the paper's own related-work section naming the projection move linear probing and citing Zou et al. among the works establishing that linear representations exist at all, though monitoring's use of the paradigm is narrower than RepE's original program, one specific forward-looking question, will the next response express this trait, rather than a general campaign to locate and characterize concept representations.