SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning

TL;DR

SAFE-QAQ addresses the shortcomings of transcription-driven fraud detection by processing raw audio end-to-end and integrating slow-thinking via rule-based reinforcement learning with efficiency-focused fine-tuning. The method unfolds in three stages—SAFE-RL for slow reasoning, SAFE-RS and SAFE-LS for concise reasoning, and SAFE-Real for real-time detection—achieving state-of-the-art performance on TeleAntiFraud-Bench while dramatically reducing reasoning length and runtime. Through end-to-end audio processing and phase-aware prompting, SAFE-QAQ effectively captures paralinguistic cues and contextual acoustics essential for spotting sophisticated deception, enabling timely live interventions. The approach demonstrates strong practical impact, including deployment in production pipelines handling over 70,000 calls daily, with reduced human effort and financial losses, while noting dataset limitations and ethical/privacy safeguards.

Abstract

Existing fraud detection methods predominantly rely on transcribed text, suffering from ASR errors and missing crucial acoustic cues like vocal tone and environmental context. This limits their effectiveness against complex deceptive strategies. To address these challenges, we first propose \textbf{SAFE-QAQ}, an end-to-end comprehensive framework for audio-based slow-thinking fraud detection. First, the SAFE-QAQ framework eliminates the impact of transcription errors on detection performance. Secondly, we propose rule-based slow-thinking reward mechanisms that systematically guide the system to identify fraud-indicative patterns by accurately capturing fine-grained audio details, through hierarchical reasoning processes. Besides, our framework introduces a dynamic risk assessment framework during live calls, enabling early detection and prevention of fraud. Experiments on the TeleAntiFraud-Bench demonstrate that SAFE-QAQ achieves dramatic improvements over existing methods in multiple key dimensions, including accuracy, inference efficiency, and real-time processing capabilities. Currently deployed and analyzing over 70,000 calls daily, SAFE-QAQ effectively automates complex fraud detection, reducing human workload and financial losses. Code: https://anonymous.4open.science/r/SAFE-QAQ.

Figures (9)

• Figure 1: Comparison of (a) Previous Method and (b) Our Proposed Method: End-to-End Call Fraud Detection via Reinforcement Learning (RL). Our approach trains LALMs in three stages: (i) developing slow-thinking reasoning through RL, (ii) optimizing thought length using rejection sampling fine-tuning and length-constrained RL, and (iii) achieving real-time detection with audio chunks training.
• Figure 2: Overview of Our Method. Starting from an LALM, we: (i) apply rule-based RL to obtain SAFE-RL with slow-thinking capabilities; (ii) refine it using rejection sampling (SAFE-RS) and length-constrained RL (SAFE-LS) to improve reasoning efficiency; and (iii) perform real-time fine-tuning on audio chunks to derive SAFE-Real.
• Figure 3: Performance-Efficiency Trade-off: Scatter Plot of Average Thinking Tokens vs. Average Classification Performance. Models closer to the top-left corner achieve a better balance of higher efficiency (fewer thinking tokens) and superior performance (higher classification scores). The points representing the best trade-offs for the baselines and our model are highlighted with star markers.
• Figure 4: Model Output Case Study: Input with Text Instruction and Audio (ASR Results for Clarity, Left), Reasoning Process of SAFE-QAQ Series (Right). Key reasoning points are highlighted in purple, and inference results are marked in orange.
• Figure 5: Prompt for Non-Real-Time.