Do misalignment directions transfer between different emergent models?

The abstract says "there is not a general misalignment direction that transfers across different EM models," and the conclusion says the framework "rebuts previous interpretations such as convergent misalignment directions." The account being rebutted, as the abstract puts it, appeals to "general misalignment directions": the idea that EM models trained on different narrow data converge on one shared internal direction for being misaligned, so a direction found in one would describe or control the behavior in another.

On the paper's account, what an EM model does is tied to its own training data through distance in the base model's representation (Does representational distance predict where misalignment emerges?). Different datasets have different centroids, so different EM models have no reason to share one direction. The two results are consistent; the excerpt does not derive one from the other.

Two vault readings. If the result holds, a monitor or steering vector calibrated on one EM model cannot be assumed to work on another; the excerpt says nothing about monitors or steering, only about transfer. That is the piece of the portability claim in Can cheap model organisms reveal misalignment threats in frontier models? that this result would bear on: "applied across frontier models" is asserted there without a test, and a direction-based countermeasure built on one organism could not be assumed to carry. Countermeasures that use no direction are outside this result. And the vault's trait directions are not obviously the target: Can we track and steer personality shifts during model finetuning? finds directions that track finetuning shifts, and How stable is the trained Assistant personality in language models? finds a dominant axis of persona space. Neither is described in the excerpt as a "general misalignment direction," and it does not say which directions it tested.

Within a model, the vault holds directions that look like the opposite result. Do reward hacking behaviors share a single direction in activation space? finds one direction per model spanning varied hacks and transferring across settings, and Does sandbagging use a single residual stream axis? finds one axis for one behavior in installed organisms. Neither is a claim about transfer between models: the first excerpt does not say whether a vector is per model, and the second does not say whether its axis is shared across locks or models. The sandbagging pair is filed as a tension against this result (A single residual-stream axis carries sandbagging while no general misalignment direction transfers across emergent misalignment models — whether the axis is shared across locks and models may decide); reading the two as within-model versus across-model is the vault's, not any paper's. One non-transfer result in a different channel sits beside it: Can language models transmit hidden behavioral traits through unrelated data? finds transmission fails when teacher and student have different base models. That concerns whether a trait rides in the data, not whether an internal direction carries over, and neither excerpt relates the two.

What the excerpt does not give. How directions were extracted, how transfer was tested, across which models, and how little transfer counts as "not general."