RESEARCH · RESEARCH · #2539
ELBP: Evidence-Grounded SLM architecture improves on-device behavioral prediction (arXiv:2610.32193v1)
This arXiv preprint introduces Evidence-Grounded Local Behavioral Prediction (ELBP), a four-layer architecture that combines deterministic event processing, dual-view behavioral representations, SLM-based prediction, and evidence verification with abstention to anticipate future assistance use from interaction events. In experiments with instruction-tuned models ≤3B parameters, Qwen2.5-3B plus ELBP reached a full-set F1 of 0.80 and PR-AUC of 0.84 (≈15.9% relative gain over raw-event prompting and 17.6% over XGBoost) and substantially reduced ECE; selective prediction with evidence verification achieved F1 0.84 at 91% coverage with 96% evidence support.
KEY POINTS
- This arXiv preprint introduces Evidence-Grounded Local Behavioral Prediction (ELBP), a four-layer architecture that combines deterministic event processing, dual-view behavioral representations, SLM-based prediction, and evidence verification with abstention to anticipate future assistance use from interaction events.
- In experiments with instruction-tuned models ≤3B parameters, Qwen2.5-3B plus ELBP reached a full-set F1 of 0.80 and PR-AUC of 0.84 (≈15.9% relative gain over raw-event prompting and 17.6% over XGBoost) and substantially reduced ECE; selective prediction with evidence verification achieved F1 0.84 at 91% coverage with 96% evidence support.
- Shows that structured behavioral representations and evidence-verified selective prediction can make small LMs more reliable and better calibrated for privacy-preserving, on-device anticipation of assistance needs.
WHY IT MATTERS
Shows that structured behavioral representations and evidence-verified selective prediction can make small LMs more reliable and better calibrated for privacy-preserving, on-device anticipation of assistance needs.