Small Language Models Assess Elder Financial Scam Risk Turn by Turn

Summary

Researchers introduce a cumulative turn-based framework for assessing progressive financial scams against older adults. The system updates risk after every conversational turn and produces a score, rationale, and safety recommendation. Tests across investment, charity, and technical-support scam dialogues show that fine-tuned Phi-4, LLaMA-3.2, DeepSeek-R1, and Qwen3 can identify fraud cues and escalation patterns in compact deployments. Phi-4 and LLaMA-3.2 perform particularly well relative to their parameter scale, supporting privacy-aware, on-device fraud monitoring.