AI article

SFT vs. RL: What Changes Inside the Model?

Community description: An intuitive guide to SFT, RLVR, and weight spectra: fine-tuning rewrites the singular values, RL keeps them and rotates the frames around them.

Dev.to | Sep 24, 2026 | Aleksei Romanov

Automated excerpt

RL adaptation occurs almost entirely through rotations of the singular coordinate frames (U and V). SVD Decomposition: Factoring a weight matrix W into U · Σ · Vᵀ (output frame, singular values spectrum, input frame). Singular Values (Σ, the Diagonal Spectrum): A non-negative diagonal matrix containing singular values σ₁ ≥ σ₂ ≥ … ≥ σᵣ ≥ 0.

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.

Read the original article

More AI news