Entropy traces predict SLM hallucination risks

This framework provides a much-needed "internal telemetry" for edge AI deployment, where raw benchmark scores often mask dangerous failure modes.

• –Deterministic models like DeepSeek-1.5B show decreasing entropy traces, suggesting a tendency toward rigid, repetitive inaccuracies
• –"Exploratory" models like Gemma-3-1B exhibit rising uncertainty, highlighting a structural predisposition toward drifting off-topic
• –Qwen-2.5-1.7B's stable entropy traces indicate its deeper architecture manages reasoning more effectively than shallower rivals
• –Real-time monitoring of these traces could enable "fail-safe" inference on mobile devices by flagging uncertain outputs before they are generated
• –The study reinforces that SLM truthfulness is a function of architectural depth and internal dynamics, not just training data scale