Figure Gallery
PI Distillation Research - All Generated Visualizations
Main Null Result
PI conditions (online RL, offline distillation, hybrid) yield equivalent final performance across domains.
Mechanism Decomposition
Comparison of training with and without failure signal, isolating the critic's contribution.
4-Level Method Hierarchy
Hierarchy of PI methods from simple rejection sampling through full async critic loops.
Cross-Domain Scatter
Performance gains scatter across Lean, MATH, and code domains showing domain-agnostic patterns.
Retry Count Inverted-U
Performance as a function of retry count K (1, 2, 3, 5), showing diminishing and then negative returns.
Baseline Competence Inverted-U
7-point curve showing PI gains peak at intermediate baseline competence, vanishing at extremes.
Mechanism Quantification Waterfall
Waterfall chart decomposing total PI gain into constituent mechanism contributions.
Cross-Seed Agreement
Overlap analysis showing per-problem solve consistency across independent training seeds.
Amortized Inference Gap
Analysis of the gap between online search cost and amortized (distilled) inference performance.