f(θ) = ½ θᵀHθ
θ⋆
κ(H) ≫ 1
gradient descent
momentum
∇θ L(θ)
x ∈ ℝⁿ
hθ
Mθ = { hθ(x) : x ∈ X }
∂hθ / ∂z₁
∂hθ / ∂z₂
hθ(x₀)
Home
Work
Blog
Note
Research
Projects
mechanistic_interpretability
2 entries
Tags
mechanistic_interpretability
A Rank-One Representation Edit Recovers Much of RMU’s WMDP Accuracy Loss
Aug 7, 2026
Stable
Slop
Residual Stream is Key to Transformer Interpretability
Jul 18, 2026
Stable
Nearly slop