Sharpening Tax in Post-Training¶
L1 core-mechanism diagnostic. Computes the coverage gap under a fixed budget, without running Meta models or agent environments.
Paper information¶
| Field | Details |
|---|---|
| Paper | Sharpening Tax in Post-Training |
| Organization | First-author affiliations: Meta Superintelligence Labs / University of Wisconsin–Madison |
| First public release | 2026-10-01 (arXiv v1) |
| Authors' code | changdaeoh/sharpening-tax is public. |
| Adapter / method | sharpening-tax |
| Local reproduction code | src/auto_research/post_training/latest_20261003.py |
Original paper¶
Motivation and method¶
Compare coverage of the base and post-trained policies under the same sampling budget. Posterior-tempered group sampling adjusts temperature by estimated difficulty to balance pass@1 and coverage.
flowchart LR
I[Public input and current state] --> M[sharpening-tax mechanism]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 1: Training pipeline and key optimization steps. From the original paper; copyright remains with the authors. Click the image for the source.
Core equation¶
\(Tax_K=Coverage_K(base)-Coverage_K(post)\).
Paper results¶
The tax appears broadly across 14 checkpoint pairs and three agent benchmarks. PTGS improves both single-attempt accuracy and coverage.
Local reproduction¶
Local comparison: The baseline disables the mechanism or uses default state; the treatment executes the defining operator. L1 reports no formal relative improvement, and relative percentages are not applicable.
- Three-seed diagnostic
diagnostic_only=true, ineligible for formal capability ranking.
Reproduction boundaries¶
The fixed-budget coverage-gap metric runs. Meta models and agent environments are not executed.