Evidence-Grounded Guardrail Extraction for Activity Recognition Using Small Language Models
Romero Victor II, Tomokazu Matsui, Yuki Matsuda, Hirohiko Suwa, Keiichi Yasumoto: “Evidence-Grounded Guardrail Extraction for Activity Recognition Using Small Language Models,” 研究報告モバイルコンピューティングと新社会システム（MBL）, Vol.2026-MBL-119, No.31, pp.1-8, 沖縄県宮古島市, 2026年5月.
Abstract
Activity recognition remains a significant challenge in pervasive computing, where models must infer user actions from sparse signals but often fail to enforce the contextual constraints required for consistent predictions. This study proposes a failure-triggered method for extracting decision guardrails using Small Language Models (SLMs) to improve the reliability of activity recognition systems. The framework constructs a Knowledge Graph of guardrails from model feedback, which serves as a grounded evidence base for subsequent inference. During inference, these grounded constraints are incorporated to guide predictions toward more contextually consistent activity recognition. We implement a pipeline that generates and reuses these constraints, and evaluate its impact on classification performance and reasoning behaviour. Experiments show that the proposed approach improves top-1 accuracy from 46.4% to 69.7%, with reduced class-level confusion and more consistent predictions. This work offers a training-free mechanism for transitioning from purely pattern-based activity recognition toward more constraint-aware systems.
Links
BibTeX
code:references.bib
@article{bib:victor_optimist_mbl119,
author={Romero II, Victor and Matsui, Tomokazu and Matsuda, Yuki and Suwa, Hirohiko and Yasumoto, Keiichi},
title={{Evidence-Grounded Guardrail Extraction for Activity Recognition Using Small Language Models}},
booktitle={研究報告モバイルコンピューティングと新社会システム（MBL）},
volume={2026-MBL-119},
number={31},
year={2026},
pages={1--7},
}
https://scrapbox.io/files/6a0535345c3c134ed4361276.png
Category
Conference
Keywords