Zero Gap Is Not Restoration: SA-PPG Metric and RailCap for Benchmark Contamination

zju · hf · 2026-08-10

A paper from ZJU argues that current metrics for evaluating benchmark contamination mitigation (G-AP) are fundamentally flawed, as over- and under-suppression can cancel each other out.

The authors propose SA-PPG (Stratified Aggregate of Per-question Probability Gaps): it estimates each question's solve probability by sampling, differences it against a clean model per question, and aggregates within groups.

They also introduce a novel mitigation strategy called RailCap. Instead of pre-estimating contamination, RailCap judges it dynamically during generation: if a sample falls back onto the greedy trajectory, it caps the next token to the runner-up until the response distribution is sufficiently dispersed. Experiments reveal that prior strategies substantially overestimate restoration, while RailCap achieves the lowest SA-PPG.

Original post →

More from Research

Research channel →