This work was motivated by following publication Mechanistically Eliciting Latent Behaviors — rely primarily on static steering vectors: h′=h+α⋅v When i get known about steering vectors as conceptual possibility i had idea to try to change knowledge i llm using only math and statistic and avoid uses gradient descend. And...
Most current activation engineering techniques — building on foundational work like Mechanistically Eliciting Latent Behaviors — rely primarily on static steering vectors: h′=h+α⋅v. While effective for global attributes (sentiment, refusal), static vectors struggle with structural tasks. They apply a constant "force" regardless of the token's context—pushing nouns and verbs in...
This work was motivated by following publication Mechanistically Eliciting Latent Behaviors — rely primarily on static steering vectors:
When i get known about steering vectors as conceptual possibility i had idea to try to change knowledge i llm using only math and statistic and avoid uses gradient descend.
And after long research i got quite interesting result.
While approach described in publication mentioned above effective for global attributes (sentiment, refusal), static vectors struggle with structural tasks. Proposed method apply a constant "force" regardless of the token's context—pushing nouns and verbs in the same direction, often leading to semantic drift or syntax degradation in long-form generation.
During research i found that exist different amount of neurons that... (read 926 more words →)
Most current activation engineering techniques — building on foundational work like
Mechanistically Eliciting Latent Behaviors — rely primarily on static steering vectors:
.
While effective for global attributes (sentiment, refusal), static vectors struggle with structural tasks. They apply a constant "force" regardless of the token's context—pushing nouns and verbs in the same direction, often leading to semantic drift or syntax degradation in long-form generation.
I’ve been working on a method called Iterative Sparse Matrix Steering, which replaces the static vector with a learned affine transformation:
Instead of using SGD (which is heavy), I solve this analytically using Ridge Regression on the CPU. This treats the steering problem as Subspace Alignment: mapping the model's internal "English geometry" onto... (read 436 more words →)
Thanks for these questions!
To be honest, the grammatical errors was unintentional because I am not native speaker, so i do mistakes time to time. And yes, they were acidental in the very beginning.
However I noticed that when i fix them, it degraded the steering performance slightly, so I decided to kept them.
I can't explain why it work, but here is my asumption:
Grammar without mistakes might allow model to overfit to syntactic patterns, "broken" or slightly incorrect syntax acts like form of noise injection/data augmentation and forcing Ridge Regression to focus on invariant semantic mapping (Moon -> Cheese) rather than sentence structure.
And i assume that it push solver to find more robust vector.... (read 380 more words →)