Recoverable but Not Stationary:Local Linear Structures in Weights and Activations
Researchers challenge the assumption that learned behaviors in neural networks align to fixed linear subspaces. Testing task vectors, LoRA, and activation steering across synthetic and real models, they find that while low-rank task structure exists locally, the basis for recovery drifts significantly over training steps rather than remaining stationary. A new Gaussian local-linear theorem explains why random search succeeds in high dimensions, offering theoretical grounding for parameter-space exploration methods. This work reshapes how practitioners should think about mechanistic control of model behavior: linear directions matter, but their geometry is dynamic, not frozen.
Modelwire context
ExplainerThe practical punchline here is not just that linear structure exists, but that the coordinates of that structure shift during training, meaning any intervention calibrated at one checkpoint may degrade silently when applied to another. The Gaussian local-linear theorem is the theoretical scaffolding that explains why random search in weight space works at all, which has been treated as an empirical curiosity until now.
This connects most directly to the 'One-Shot GRPO' paper from the same day, which showed that alignment interventions applied through post-training can be overridden by a single adversarial example. If the linear directions used by techniques like activation steering are themselves non-stationary across training steps, that fragility has a geometric explanation: practitioners may be targeting a subspace that has already drifted by the time the intervention is applied. Together, these two papers suggest that the instability of behavioral control is not purely a data or optimization problem but may be partly structural.
Watch whether teams building activation steering pipelines for production models begin reporting checkpoint-sensitivity in their evaluations. If drift rates correlate with model scale or training compute, that would give practitioners a concrete signal for when recalibration is necessary.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsLoRA · DistilGPT-2 · GPT-2 · task vectors · activation steering
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.