Users often modify their interaction patterns before explicitly asking for help, giving systems a chance to anticipate future assistance needs. In high‑risk domains, delayed help can cause operational, financial, or safety issues. Large language models are usually too heavy for continuous on‑device inference, while behavioral patterns are highly personal and device‑specific, calling for local processing that safeguards privacy and data sovereignty. Small language models are lightweight and fast, yet their ability to reason over low‑level sequential events remains under‑examined.
We introduce Evidence‑Grounded Local Behavioral Prediction (ELBP), a four‑layer pipeline:
- Deterministic interaction processing that normalizes raw events;
- Dual‑view behavioral representation capturing temporal dynamics and contextual semantics;
- SLM‑based prediction that consumes the dual view and outputs probability scores;
- Evidence verification with abstention, where the model’s generated evidence determines confidence and triggers abstention when needed.
Three instruction‑tuned models with ≤3 B parameters were tested under raw‑event, structured, dual‑view, and full ELBP settings. The top performer, Qwen2.5‑3B with ELBP, achieved an overall F1 of 0.80 and PR‑AUC of 0.84, a 15.9% gain over raw‑event prompting and a 17.6% gain over XGBoost, while full‑set ECE dropped by 43.2%. With evidence verification, selective F1 rose to 0.84 at 91% coverage and evidence support reached 96%.
These findings demonstrate that structured behavioral representations combined with evidence‑verified selective prediction markedly improve the reliability of behavioral forecasting using small language models.
Review