@article{ART003397396},
author={Tae-Yeon Eom and Jae-Woong Kim and Won-Young Jeon},
title={Evaluating LLM Recommendation Safety under Hard Constraints},
journal={Journal of The Korea Society of Computer and Information},
issn={1598-849X},
year={2026},
volume={31},
number={10},
pages={107-113}
TY - JOUR
AU - Tae-Yeon Eom
AU - Jae-Woong Kim
AU - Won-Young Jeon
TI - Evaluating LLM Recommendation Safety under Hard Constraints
JO - Journal of The Korea Society of Computer and Information
PY - 2026
VL - 31
IS - 10
PB - The Korean Society Of Computer And Information
SP - 107
EP - 113
SN - 1598-849X
AB - This study compared final-action performance and rejection feedback conditions across four hard-constraint product recommendation procedures using a single local LLM (qwen3:8b). A fixed 40-monitor catalog, based on official manufacturer specifications and domestic prices, mapped 19 hard constraints to 46 product fields.
With deterministic ground truth precomputed for 60 queries, we compared single pass (M1), self-refinement (M2), and Proposer–Critic with rejected-product-ID feedback (M3) or detailed rejection history containing violated constraint keys, required values, actual values, and natural-language reasons (M4). Under fixed evaluation conditions, final-action accuracy on REJECTION_STRESS was 40.0% (M1), 46.7% (M2), and 93.3% (M3 and M4). Both M3 and M4 recorded zero unsafe acceptances and correctly abstained on all 10 infeasible queries. Repeated same-constraint violation rates were 61.11% (M3) and 57.14% (M4), with a difference (M4−M3) of −3.97 percentage points (95% confidence interval: [−10.77, 0.00] percentage points). M4 showed a lower observed repeated-violation rate, although the confidence interval included zero, and slightly higher input-token use and response latency than M3.
KW - LLMs;hard-constraint recommendation;proposer–critic;abstention;deterministic ground truth
DO -
UR -
ER -
Tae-Yeon Eom, Jae-Woong Kim and Won-Young Jeon. (2026). Evaluating LLM Recommendation Safety under Hard Constraints. Journal of The Korea Society of Computer and Information, 31(10), 107-113.
Tae-Yeon Eom, Jae-Woong Kim and Won-Young Jeon. 2026, "Evaluating LLM Recommendation Safety under Hard Constraints", Journal of The Korea Society of Computer and Information, vol.31, no.10 pp.107-113.
Tae-Yeon Eom, Jae-Woong Kim, Won-Young Jeon "Evaluating LLM Recommendation Safety under Hard Constraints" Journal of The Korea Society of Computer and Information 31.10 pp.107-113 (2026) : 107.
Tae-Yeon Eom, Jae-Woong Kim, Won-Young Jeon. Evaluating LLM Recommendation Safety under Hard Constraints. 2026; 31(10), 107-113.
Tae-Yeon Eom, Jae-Woong Kim and Won-Young Jeon. "Evaluating LLM Recommendation Safety under Hard Constraints" Journal of The Korea Society of Computer and Information 31, no.10 (2026) : 107-113.
Tae-Yeon Eom; Jae-Woong Kim; Won-Young Jeon. Evaluating LLM Recommendation Safety under Hard Constraints. Journal of The Korea Society of Computer and Information, 31(10), 107-113.
Tae-Yeon Eom; Jae-Woong Kim; Won-Young Jeon. Evaluating LLM Recommendation Safety under Hard Constraints. Journal of The Korea Society of Computer and Information. 2026; 31(10) 107-113.
Tae-Yeon Eom, Jae-Woong Kim, Won-Young Jeon. Evaluating LLM Recommendation Safety under Hard Constraints. 2026; 31(10), 107-113.
Tae-Yeon Eom, Jae-Woong Kim and Won-Young Jeon. "Evaluating LLM Recommendation Safety under Hard Constraints" Journal of The Korea Society of Computer and Information 31, no.10 (2026) : 107-113.
