The Journal for Information Professionals
Asia Pacific Journal of Information Systems (APJIS), a Scopus and ABDC indexed journal, is a
flagship journal of the information systems (IS) field in the Asia Pacific region.
ISSN 2288-5404 (Print) / ISSN 2288-6818 (Online)
Editor : Youngsok Bang
Share this page
Current Issue
| Date | September 2026 |
|---|---|
| Vol. No. | Vol. 36 No. 3 |
| DOI | https://doi.org/10.14329/apjis.2026.36.3.694 |
| Page | 694~724 |
| Title | Stepwise Post-Training of Compact Language Models for Business Applications: A Design-Oriented Study of Reasoning and Conversational Retention |
| Author | Jinsoo Park, Seungho Jeong |
| Keyword | Small Language Models, Business AI Applications, Stepwise Post-Training, Resource-Aware AI Design |
| Abstract | As organizations explore generative AI for customer-facing and support processes, task-specialized adaptation can improve downstream performance but weaken general conversational capabilities inherited from the base model. This trade-off is especially salient for small language models (SLMs), which are attractive for cost, latency, governance, and deployment flexibility but remain capacity-constrained on reasoning-intensive tasks. This study develops and evaluates a design-oriented stepwise post-training framework for calibrating a 1B-scale language model for bounded business applications while maintaining usable conversational quality. Using Llama-3.2-1B as the student model, the framework integrates warm-up supervised fine-tuning, stepwise reinforcement learning, rationale-consistent distillation, and selective reasoning control. The post-trained model was evaluated on customer intent inference, conversational preference, and selected PISA mathematical-reasoning items. Results suggest improvement over the base model and the same observed Pass@3 accuracy as Llama-3.2-3B on selected PISA mathematics items, with no clear evidence of degradation in limited human-preference tests. The observed gains were task-dependent, appearing more clearly in reasoning-oriented tasks than in routine classification. The study frames compact models as deployment-aware business artifacts and identifies a design logic for balancing task specialization, conversational retention, and resource-conscious model design. |
Home l Site Map l Abstracting/Indexing l FAQ l Publisher l Contact Us l Admin Login
© 2013 The Korean Society of Management Information Systems. All rights reserved.