Enhancing LLM-based Zero-Shot Conversational Recommendation via Reasoning Path

Heejin Kook, Seongmin Park, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.7.617

Conversational recommender systems provide personalized recommendations through bi-directional interactions with users. Traditional conversational recommender systems rely on external knowledge, such as knowledge graphs, to effectively capture user preferences. While recent rapid advancement of large language models has enabled zero-shot recommendations, challenges remain in understanding users' implicit preferences and designing optimal reasoning paths. To address these limitations, this study investigates the importance of appropriate reasoning path construction in zero-shot based conversational recommender systems and explores the potential of using a new approach based on this foundation. The proposed framework consists of two stages: (1) comprehensively extracting both explicit and implicit preferences from conversational context, and (2) constructing reasoning trees to select optimal reasoning paths based on these preferences. Experimental results on benchmark datasets INSPIRED and ReDial show that our proposed method achieves up to 11.77% improvement in Recall@10 compared to existing zero-shot methods, It even outperforms some learning-based models.

A Large Language Model-based Multi-domain Recommender System using Model Merging

Hyunsoo Kim, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.6.548

Recent research in recommender systems has increasingly focused on leveraging pre-trained large language models (LLMs) to effectively understand the natural language information associated with recommendation items. While these LLM-based recommender systems achieve high accuracy, they have a limitation in that they require training separate recommendation models for each domain. This increases the costs of storing and inferring multiple models and makes it difficult to share knowledge across domains. To address this issue, we propose an LLM-based recommendation model that effectively operates across diverse recommendation domains by applying task vector-based model merging. During the merging process, knowledge distillation is utilized from individually trained domain-specific recommendation models to learn optimal merging weights. Experimental results show that our proposed method improves recommendation accuracy by an average of 2.75% across eight domains compared to recommender models utilizing existing model merging methods, while also demonstrating strong generalization performance in previously unseen domains.

Linear Sequential Recommendation Models using Textual Side Information

Dongcheol Lee, Minjin Choi, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.6.529

Recently, research on leveraging auxiliary information in sequential recommendation systems is being actively conducted. Most approaches have focused on combining language models with deep neural networks. However, they often lead to high computational costs and latency issues. While linear recommendation models can serve as an efficient alternative, research on how to effectively incorporate auxiliary information is lacking. This study proposed a framework that could effectively utilize auxiliary information within a linear model. Since textual data cannot be directly used in linear model training, we transformed item texts into dense vectors using a pre-trained text encoder. Although these vectors contained rich information, they failed to capture relationships between items. To address this, we applied graph convolution to obtain enhanced item representations. These representations were then used alongside the user-item interaction matrix for linear model training. Extensive experiments showed that the proposed method improved the overall performance, particularly in recommending less popular items.

An Inference Framework for Text-Based Sequential Recommendation Model Using Nearest Neighbor Mechanism

Junyoung Kim, Hyunsoo, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.5.435

Sequential recommendation task aims to predict the next item to interact with based on users’ interaction history. Text-based recommendation models, which represent items as text, show improved performance in cold-start problems and zero-shot recommendation tasks. However, they suffer from textual bias and the lack of collaborative knowledge. To overcome these limitations, we propose a text-based recommendation model inference framework using the nearest neighbor mechanism. The proposed method leverages text-based recommendation models as a neighbor retriever model to search neighbors with similar preferences to the user and aggregate the neighbor information with existing recommendation results to improve recommendation performance. Experiments conducted on four datasets show that the proposed method consistently outperforms existing models, with performance improvement up to 25.27% on NDCG@50. Furthermore, the proposed method effectively complements collaborative knowledge and improves model explainability by providing recommendation rationale.

Efficient Large Language Model Based Passage Re-Ranking Using Single Token Representations

Jeongwoo Na, Jun Kwon, Eunseong Choi, Jongwuk Lee

http://doi.org/10.5626/JOK.2025.52.5.395

In information retrieval systems, document re-ranking involves reordering a set of candidate documents based on evaluation of their relevance to a given query. Leveraging extensive natural language understanding capabilities of large language models(LLMs), numerous studies on document re-ranking have been conducted, demonstrating groundbreaking performance. However, studies utilizing large language models focus solely on improving reranking performance, resulting in degraded efficiency due to excessively long input sequences and the need for repetitive inference. To address these limitations, we propose ListT5++, a novel model that represents the relevance between a query and a passage using single token embedding and significantly improves the efficiency of LLM-based reranking through a single-step decoding strategy that minimizes the decoding process. Experimental results showed that ListT5++ could maintain accuracy levels comparable to existing methods while reducing inference latency by a factor of 29.4 relative to the baseline. Moreover, our approach demonstrates robust characteristics by being insensitive to th initial ordering of candidate documents, thereby ensuring high practicality in real-time retrieval environments.

Enhancing Molecular Understanding in LLMs through Multimodal Graph-SMILES Representations

http://doi.org/10.5626/JOK.2025.52.5.379

Recent advancements in large language models (LLMs) have shown remarkable performace across various tasks, with increasing focus on multimodal research. Notably, BLIP-2 can enhance performance by efficiently aligning images and text using a Q-Former, aided by an image encoder pre-trained on multimodal data. Inspired by this, the MolCA model extends BLIP-2 to the molecular domain to improve performance. However, the graph encoder in MolCA is pre-trained on unimodal data, necessitating updates during model training, which is a limitation. Therefore, this paper replaced it with a graph encoder pre-trained on multimodal data and frozen while training the model. Experimental results showed that using the graph encoder pre-trained on multimodal data generally enhanced performance. Additionally, unlike the graph encoder pre-trained on unimodal data, which performed better when updated, the graph encoder pre-trained on multimodal data achieved superior results across all metrics when frozen.

SASRec vs. BERT4Rec: Performance Analysis of Transformer-based Sequential Recommendation Models

Hye-young Kim, Mincheol Yoon, Jongwuk Lee

http://doi.org/10.5626/JOK.2024.51.4.352

Sequential recommender systems extract interests from user logs and use them to recommend items the user might like next. SASRec and BERT4Rec are widely used as representative sequential recommendation models. Existing studies have utilized these two models as baselines in various studies, but their performance is not consistent due to differences in experimental environments. This research compares and analyzes the performance of SASRec and BERT4Rec on six representative sequential recommendation datasets. The experimental result shows that the number of user-item interactions has the largest impact on BERT4Rec training, which in turn leads to the performance difference between the two models. Furthermore, this research finds that the two learning methods, which are widely utilized in sequential recommendation settings, can also have different effects depending on the popularity bias and sequence length. This shows that considering dataset characteristics is essential for improving recommendation performance.

Learning with Noisy Labels using Sample Selection based on Language-Image Pre-trained Model

Bonggeon Cha, Minjin Choi, Jongwuk Lee

http://doi.org/10.5626/JOK.2023.50.6.511

Deep neural networks have significantly degraded generalization performance when learning with noisy labels. To address this problem, previous studies observed that the model learns clean samples first in the early learning stage, and based on this, sample selection methods that selectively train data by considering small-loss samples as clean samples have been used to improve performance. However, when noisy labels are similar to their ground truth(e.g., seal vs. otter), sample selection is not effective because the model learns noisy data in the early learning stage. In this paper, we propose a Sample selection with Language-Image Pre-trained model (SLIP) which effectively distinguishes and learns clean samples without the early learning stage by leveraging zero-shot predictions from a pre-trained language-image model. Our proposed method shows up to 18.45%p improved performance over previously proposed methods on CIFAR-10, CIFAR-100, and WebVision.

Improving the Performance of Knowledge Tracing Models using Quantized Correctness Embeddings

Yoonjin Im, Jaewan Moon, Eunseong Choi, Jongwuk Lee

http://doi.org/10.5626/JOK.2023.50.4.329

Knowledge tracing is a task of monitoring the proficiency of knowledge based on learners" interaction records. Despite the flexible usage of deep neural network-based models for this task, the existing methods disregard the difficulty of each question and result in poor performance for learners who get the easy question wrong or the hard question correct. In this paper, we propose quantizing the learners’ response information based on the question difficulty so that the knowledge tracing models can learn both the response and the difficulty of the question in order to improve the performance. We design a method that can effectively discriminate between negative samples with a high percentage of correct answer rate and positive samples with a low percentage of correct answer rate. Toward this end, we use sinusoidal positional encoding (SPE) that can maximize the distance difference between embedding representations in the latent space. Experiments show that the AUC value is improved to a maximum of 17.89% in the target section compared to the existing method.

Data Augmentation Methods for Improving the Performance of Machine Reading Comprehension

Sunkyung Lee, Eunseong Choi, Seonho Jeong, Jongwuk Lee

http://doi.org/10.5626/JOK.2021.48.12.1298

Machine reading comprehension is a method of understanding the meaning and performing inference over a given text by computers, and it is one of the most essential techniques for understanding natural language. The question answering task yields a way to test the reasoning ability of intelligent systems. Nowadays, machine reading comprehension techniques performance has significantly improved following the recent progress of deep neural networks. Nevertheless, there may be challenges in improving performance when data is sparse. To address this issue, we leverage word-level and sentence-level data augmentation techniques through text editing, while minimizing changes to the existing models and cost. In this work, we propose data augmentation methods for a pre-trained language model, which is most widely used in English question answering tasks, to confirm the improved performance over the existing models.