Sungguk Cha

Sungguk Cha, DongWook Kim, Mintae Kim et al.

Multi-vector embedding models have emerged as a powerful paradigm for document retrieval, preserving fine-grained visual and textual details through token-level representations. However, this expressiveness comes at a staggering cost: storing embeddings for every token inflates index sizes by over $1000\times$ compared to single-vector approaches, severely limiting scalability. We introduce \textbf{ReinPool}, a reinforcement learning framework that learns to dynamically filter and pool multi-vector embeddings into compact, retrieval-optimized representations. By training with an inverse retrieval objective and NDCG-based rewards, ReinPool identifies and retains only the most discriminative vectors without requiring manual importance annotations. On the Vidore V2 benchmark across three vision-language embedding models, ReinPool compresses multi-vector representations by $746$--$1249\times$ into single vectors while recovering 76--81\% of full multi-vector retrieval performance. Compared to static mean pooling baselines, ReinPool achieves 22--33\% absolute NDCG@3 improvement, demonstrating that learned selection significantly outperforms heuristic aggregation.

Sungguk Cha, DongWook Kim

Large language model (LLM) agents are increasingly deployed in social settings, yet little is known about how they interact in open-ended environments. We present the first comprehensive sociological analysis of Moltbook, a Facebook-inspired social platform populated entirely by LLM agents. Analyzing 184,203 posts and 465,136 comments across 14 daily snapshots (2026-04-14 to 2026-04-28), we examine agent sociality through six research questions grounded in classical social theory: bonding vs. bridging communities, status hierarchies, temporal coordination, information diffusion, identity performance, and norm enforcement. Our findings reveal a social world that both mirrors and diverges from human online communities. Reciprocity is strikingly low (3.8% multi-day vs. 1.6% single-day; below the 10-30% range typical of human baselines), suggesting "attention bonding without exchange bonding." Prestige is heavy-tailed (top score 104.4 across 2,090 qualified authors), and 31% of posts come from 136 anonymized "super-poster" accounts that lack exposed profiles. Temporal activity is broadly flat across the day with a sustained 12:00-20:00 UTC working-hour band; k-means recovers six distinct temporal communities. Of 458 bridge agents, 325 carry at least one tracked viral phrase; 99.7% of those (324/325) are late amplifiers, not early adopters. Identity performance shows no unconditional engagement payoff (-72%), but stratifying by post-volume quartile reverses the sign in the upper half of the distribution -- a Simpson's-paradox effect rather than a uniform penalty. Most remarkably, downvotes are rare (0.9%), and a comment-sentiment test rejects the alternative-channel hypothesis: textual sanction is also absent. We frame these patterns through a "parasocial simulators" construct.

Jusung Lee, Sungguk Cha, Younghyun Lee et al.

Having revolutionized natural language processing (NLP) applications, large language models (LLMs) are expanding into the realm of multimodal inputs. Owing to their ability to interpret images, multimodal LLMs (MLLMs) have been primarily used for vision-language tasks. Currently, MLLMs have not yet been extended for domain-specific visual tasks, which require a more explicit understanding of visual information. We developed a method to transform domain-specific visual and vision-language datasets into a unified question answering format called Visual Question Answering Instruction (VQA-IN), thereby extending MLLM to domain-specific tasks. The VQA-IN was applied to train multiple MLLM architectures using smaller versions of LLMs (sLLMs). The experimental results indicated that the proposed method achieved a high score metric on domainspecific visual tasks while also maintaining its performance on vision-language tasks in a multitask manner.

Sungguk Cha, Jusung Lee, Younghyun Lee et al.

"When did the emperor Napoleon invented iPhone?" Such hallucination-inducing question is well known challenge in generative language modeling. In this study, we present an innovative concept of visual hallucination, referred to as "I Know (IK)" hallucination, to address scenarios where "I Don't Know" is the desired response. To effectively tackle this issue, we propose the VQAv2-IDK benchmark, the subset of VQAv2 comprising unanswerable image-question pairs as determined by human annotators. Stepping further, we present the visually dehallucinative instruction generation method for IK hallucination and introduce the IDK-Instructions visual instruction database. Our experiments show that current methods struggle with IK hallucination. Yet, our approach effectively reduces these hallucinations, proving its versatility across different frameworks and datasets.

Sungguk Cha, Jusung Lee, Younghyun Lee et al.

In recent years, synthetic visual instructions by generative language model have demonstrated plausible text generation performance on the visual question-answering tasks. However, challenges persist in the hallucination of generative language models, i.e., the generated image-text data contains unintended contents. This paper presents a novel and scalable method for generating visually dehallucinative instructions, dubbed CAP2QA, that constrains the scope to only image contents. Our key contributions lie in introducing image-aligned instructive QA dataset CAP2QA-COCO and its scalable recipe. In our experiments, we compare synthetic visual instruction datasets that share the same source data by visual instruction tuning and conduct general visual recognition tasks. It shows that our proposed method significantly reduces visual hallucination while consistently improving visual recognition ability and expressiveness.

Sungguk Cha, DongWook Kim, Taeseung Hahn et al.

Retrieval-Augmented Generation (RAG) systems rely heavily on effective query formulation to unlock external knowledge, yet optimizing queries for diverse, unstructured real-world documents remains a challenge. We introduce \textbf{RL-QR}, a reinforcement learning framework for retriever-specific query rewriting that eliminates the need for human-annotated datasets and extends applicability to both text-only and multi-modal databases. By synthesizing scenario-question pairs and leveraging Generalized Reward Policy Optimization (GRPO), RL-QR trains query rewriters tailored to specific retrievers, enhancing retrieval performance across varied domains. Experiments on industrial in-house data demonstrate significant improvements, with $\text{RL-QR}_{\text{multi-modal}}$ achieving an 11\% relative gain in NDCG@3 for multi-modal RAG and $\text{RL-QR}_{\text{lexical}}$ yielding a 9\% gain for lexical retrievers. However, challenges persist with semantic and hybrid retrievers, where rewriters failed to improve performance, likely due to training misalignments. Our findings highlight RL-QR's potential to revolutionize query optimization for RAG systems, offering a scalable, annotation-free solution for real-world retrieval tasks, while identifying avenues for further refinement in semantic retrieval contexts.

Sungguk Cha, Yooseung Wang

Fully supervised semantic segmentation technologies bring a paradigm shift in scene understanding. However, the burden of expensive labeling cost remains as a challenge. To solve the cost problem, recent studies proposed language model based zero-shot semantic segmentation (L-ZSSS) approaches. In this paper, we address L-ZSSS has a limitation in generalization which is a virtue of zero-shot learning. Tackling the limitation, we propose a language-model-free zero-shot semantic segmentation framework, Spatial and Multi-scale aware Visual Class Embedding Network (SM-VCENet). Furthermore, leveraging vision-oriented class embedding SM-VCENet enriches visual information of the class embedding by multi-scale attention and spatial attention. We also propose a novel benchmark (PASCAL2COCO) for zero-shot semantic segmentation, which provides generalization evaluation by domain adaptation and contains visually challenging samples. In experiments, our SM-VCENet outperforms zero-shot semantic segmentation state-of-the-art by a relative margin in PASCAL-5i benchmark and shows generalization-robustness in PASCAL2COCO benchmark.