Kentaro Ueda

Kentaro Ueda, François Portet, Hirohiko Suwa et al.

While LLMs excel at general tasks, they struggle in specialized domains like finance, requiring diverse skills in domain knowledge, mathematical reasoning, and multilingual processing. Merging domain-specific Continual Pre-training (CPT) "experts" offers a practical alternative to costly and unstable multi-skill training. However, unlike established Supervised Fine-Tuning (SFT) model-based merging, CPT model merging remains largely unexplored. We address this gap by creating financial LLMs from experts in finance, math, and Japanese. We propose a three-stage evaluation focusing on knowledge recovery, complementarity, and emergence, and assess three merging methods (Task Arithmetic, TIES, and DARE-TIES) on a comprehensive financial benchmark curated from 18 tasks across 8 established datasets. Results show that merging an expert with its base model recovers general knowledge lost during CPT, while merging experts improves performance and can yield emergent cross-domain skills. Among the methods, Task Arithmetic performs strongly but is hyperparameter-sensitive, whereas TIES is more robust. Our findings also suggest that while model similarity correlates with merging success, emergent skills depend on more complex factors. This work presents the first foundational analysis of CPT model merging, establishing a principled framework and providing clear guidance for building multi-skill LLMs from existing assets.

Kentaro Ueda, Takehiro Takayanagi

While recent advances in Large Language Models (LLMs) have improved the quality of creative text generation, significant challenges remain in producing personalized stories that reflect individual user preferences. Conventional approaches rely on explicit feedback or fine-tuning, which presents practical issues regarding user burden, data collection, computational costs, and privacy. In this work, we propose PREFINE (Persona-and-Rubric Guided Critique-and-Refine), a novel framework that extends the Critique-and-Refine paradigm to personalization. PREFINE constructs a pseudo-user agent from a user's interaction history and generates user-specific rubrics (evaluation criteria). By having this agent critique and refine outputs on the user's behalf based on these tailored rubrics, our method achieves personalized generation without requiring parameter updates or direct user feedback. We conducted a comprehensive evaluation on the PerDOC and PerMPST story datasets. We designed three baseline methods and several model variants to verify the contribution of each component of our framework. In automatic evaluations (LLM-as-a-Judge), PREFINE achieved higher win rates and statistically significant scores than the baselines, without compromising general story quality. Analysis of the model variants confirmed that both the pseudo-user agent and the user-specific rubrics are crucial for enhancing personalization performance. Beyond story generation, our approach holds potential for enabling efficient personalization in broader applications, such as dialogue systems, education, and recommendation.