Continuous Function

Legacy Concept Lab

Self-Improvement & Distillation Loops

Makes data generation and model improvement a closed loop

Concept 98 of 100Scaling & AlignmentPhase 13
Open DemoAll Foundations
#98Self-ImproveScaling & Alignment
key equation\theta_{k+1} = \arg\min_\theta \mathcal{L}(\theta; D_k \cup \text{self-gen})
Phase 13: Cutting-edge 2024-2025 researchConcept 98 of 100

Why It Matters for Modern Models

  • Makes data generation and model improvement a closed loop
  • DeepSeek-R1: RL → reasoning → distill to smaller models
  • Reduces reliance on scarce human labels

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • Model bootstraps on its own outputs (filtered by verifier)
  • Teacher-student paradigm: large model → small model
  • Risk: distribution shift, mode collapse

Interactive Visualization

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and the visualization. Come back here for the full walkthrough.

Key Equation
θk+1=arg⁡min⁡θL(θ;Dk∪self-gen)\theta_{k+1} = \arg\min_\theta \mathcal{L}(\theta; D_k \cup \text{self-gen})

Iterative self-training loop:

Dk+1=Dk∪{(x,y^):y^∼πθk(⋅∣x)}D_{k+1} = D_k \cup \{(x, \hat{y}) : \hat{y} \sim \pi_{\theta_k}(\cdot|x)\}
θk+1=arg⁡min⁡θE(x,y)∼Dk+1[−log⁡πθ(y∣x)]\theta_{k+1} = \arg\min_\theta \mathbb{E}_{(x,y) \sim D_{k+1}}[-\log \pi_\theta(y|x)]

Distillation: min⁡θsEx[KL(πθt(⋅∣x)∥πθs(⋅∣x))]\min_{\theta_s} \mathbb{E}_x[\mathrm{KL}(\pi_{\theta_t}(\cdot|x) \| \pi_{\theta_s}(\cdot|x))]

Canonical Papers

Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models

Chen et al.2024ICML
Read paper →

Connections

Prerequisites

🧪Distillation⚖RLHF

Next Moves

Explore this concept from different angles — like a mathematician would.