FRAUDSkill Boosts Audio Anti-Fraud Macro-F1 to 73.5% With Frozen Weights

PPSUCTeleantifraudCommunity · hf · 2026-09-21

The PPSU TeleAntiFraud community released FRAUDSkill, a structured frozen-weight adaptation framework for audio anti-fraud detection. Instead of fine-tuning, it optimizes an external layer of skill programs, route-specific policies, and decision rules on top of an unchanged audio-language model, combining structured output control with validation-guided multi-path inference to enforce a decision protocol of scenario identification, fraud detection, and conditional fraud-type classification.

On the TeleAntiFraud benchmark it reaches 73.50% Macro-F1, outperforming the shared frozen-model baseline by 31.96 points while cutting invalid outputs to 1.94%. Code is available.

Original post →

More from Research

Research channel →