Ji-Hoon Park

Yeong-Joon Ju, Ji-Hoon Park, Seong-Whan Lee

Despite deep learning (DL) has achieved remarkable progress in various domains, the DL models are still prone to making mistakes. This issue necessitates effective debugging tools for DL practitioners to interpret the decision-making process within the networks. However, existing debugging methods often demand extra data or adjustments to the decision process, limiting their applicability. To tackle this problem, we present NeuroInspect, an interpretable neuron-based debugging framework with three key stages: counterfactual explanations, feature visualizations, and false correlation mitigation. Our debugging framework first pinpoints neurons responsible for mistakes in the network and then visualizes features embedded in the neurons to be human-interpretable. To provide these explanations, we introduce CLIP-Illusion, a novel feature visualization method that generates images representing features conditioned on classes to examine the connection between neurons and the decision layer. We alleviate convoluted explanations of the conventional visualization approach by employing class information, thereby isolating mixed properties. This process offers more human-interpretable explanations for model errors without altering the trained network or requiring additional data. Furthermore, our framework mitigates false correlations learned from a dataset under a stochastic perspective, modifying decisions for the neurons considered as the main causes. We validate the effectiveness of our framework by addressing false correlations and improving inferences for classes with the worst performance in real-world settings. Moreover, we demonstrate that NeuroInspect helps debug the mistakes of DL models through evaluation for human understanding. The code is openly available at https://github.com/yeongjoonJu/NeuroInspect.

Geon Park, Ji-Hoon Park, Seong-Whan Lee

Composed image retrieval (CIR) addresses the task of retrieving a target image by jointly interpreting a reference image and a modification text that specifies the intended change. Most existing methods are still built upon contrastive learning frameworks that treat the ground truth image as the only positive instance and all remaining images as negatives. This strategy inevitably introduces relevance suppression, where semantically related yet valid images are incorrectly pushed away, and semantic confusion, where different modification intents collapse into overlapping regions of the embedding space. As a result, the learned query representations often lack discriminativeness, particularly at fine-grained attribute modifications. To overcome these limitations, we propose distinctive query embeddings through learnable attribute weights and target relative negative sampling (DQE-CIR), a method designed to learn distinctive query embeddings by explicitly modeling target relative relevance during training. DQE-CIR incorporates learnable attribute weighting to emphasize distinctive visual features conditioned on the modification text, enabling more precise feature alignment between language and vision. Furthermore, we introduce target relative negative sampling, which constructs a target relative similarity distribution and selects informative negatives from a mid-zone region that excludes both easy negatives and ambiguous false negatives. This strategy enables more reliable retrieval for fine-grained attribute changes by improving query discriminativeness and reducing confusion caused by semantically similar but irrelevant candidates.

Ji-Hoon Park, Yeong-Joon Ju, Seong-Whan Lee

Diffusion models have demonstrated remarkable performance in generation tasks. Nevertheless, explaining the diffusion process remains challenging due to it being a sequence of denoising noisy images that are difficult for experts to interpret. To address this issue, we propose the three research questions to interpret the diffusion process from the perspective of the visual concepts generated by the model and the region where the model attends in each time step. We devise tools for visualizing the diffusion process and answering the aforementioned research questions to render the diffusion process human-understandable. We show how the output is progressively generated in the diffusion process by explaining the level of denoising and highlighting relationships to foundational visual concepts at each time step through the results of experiments with various visual analyses using the tools. Throughout the training of the diffusion model, the model learns diverse visual concepts corresponding to each time-step, enabling the model to predict varying levels of visual concepts at different stages. We substantiate our tools using Area Under Cover (AUC) score, correlation quantification, and cross-attention mapping. Our findings provide insights into the diffusion process and pave the way for further research into explainable diffusion mechanisms.

Hanna Zubkova, Ji-Hoon Park, Seong-Whan Lee

Bearing in mind the limited parametric knowledge of Large Language Models (LLMs), retrieval-augmented generation (RAG) which supplies them with the relevant external knowledge has served as an approach to mitigate the issue of hallucinations to a certain extent. However, uniformly retrieving supporting context makes response generation source-inefficient, as triggering the retriever is not always necessary, or even inaccurate, when a model gets distracted by noisy retrieved content and produces an unhelpful answer. Motivated by these issues, we introduce Semantic Uncertainty Guided Adaptive Retrieval (SUGAR), where we leverage context-based entropy to actively decide whether to retrieve and to further determine between single-step and multi-step retrieval. Our empirical results show that selective retrieval guided by semantic uncertainty estimation improves the performance across diverse question answering tasks, as well as achieves a more efficient inference.

Kun-Woo Kim, Ji-Hoon Park, Ju-Min Han et al.

Large Language Models (LLMs) trained on extensive datasets often learn sensitive information, which raises significant social and legal concerns under principles such as the "Right to be forgotten." Retraining entire models from scratch to remove undesired information is both costly and impractical. Furthermore, existing single-domain unlearning methods fail to address multi-domain scenarios, where knowledge is interwoven across domains such as privacy and copyright, creating overlapping representations that lead to excessive knowledge removal or degraded performance. To tackle these issues, we propose GRAIL (GRadient-based AdaptIve unLearning), a novel multi-domain unlearning framework. GRAIL leverages gradient information from multiple domains to precisely distinguish the unlearning scope from the retention scope, and applies an adaptive parameter-wise localization strategy to selectively remove targeted knowledge while preserving critical parameters for each domain. Experimental results on unlearning benchmarks show that GRAIL achieves unlearning success on par with the existing approaches, while also demonstrating up to 17% stronger knowledge retention success compared to the previous state-of-art method. Our findings establish a new paradigm for effectively managing and regulating sensitive information in large-scale pre-trained language models.

L'ea Dubois, Klaus Schmidt, Chengyu Wang et al.

Current video understanding models excel at recognizing "what" is happening but fall short in high-level cognitive tasks like causal reasoning and future prediction, a limitation rooted in their lack of commonsense world knowledge. To bridge this cognitive gap, we propose a novel framework that synergistically fuses a powerful Vision Foundation Model (VFM) for deep visual perception with a Large Language Model (LLM) serving as a knowledge-driven reasoning core. Our key technical innovation is a sophisticated fusion module, inspired by the Q-Former architecture, which distills complex spatiotemporal and object-centric visual features into a concise, language-aligned representation. This enables the LLM to effectively ground its inferential processes in direct visual evidence. The model is trained via a two-stage strategy, beginning with large-scale alignment pre-training on video-text data, followed by targeted instruction fine-tuning on a curated dataset designed to elicit advanced reasoning and prediction skills. Extensive experiments demonstrate that our model achieves state-of-the-art performance on multiple challenging benchmarks. Notably, it exhibits remarkable zero-shot generalization to unseen reasoning tasks, and our in-depth ablation studies validate the critical contribution of each architectural component. This work pushes the boundary of machine perception from simple recognition towards genuine cognitive understanding, paving the way for more intelligent and capable AI systems in robotics, human-computer interaction, and beyond.

Jung-Woo Shim, Yeong-Joon Ju, Ji-Hoon Park et al.

Recent advancements in large language models (LLMs) highlight their fluency in generating responses to diverse prompts. However, these models sometimes generate plausible yet incorrect ``hallucinated" facts, undermining trust. A frequent but often overlooked cause of such errors is the use of poorly structured or vague prompts by users, leading LLMs to base responses on assumed rather than actual intentions. To mitigate hallucinations induced by these ill-formed prompts, we introduce Curative Prompt Refinement (CPR), a plug-and-play framework for curative prompt refinement that 1) cleans ill-formed prompts, and 2) generates additional informative task descriptions to align the intention of the user and the prompt using a fine-tuned small language model. When applied to language models, we discover that CPR significantly increases the quality of generation while also mitigating hallucination. Empirical studies show that prompts with CPR applied achieves over a 90\% win rate over the original prompts without any external knowledge.

Jung-Woo Shim, Yeong-Joon Ju, Ji-Hoon Park et al.

Recent advancements in large language models (LLMs) have shown strong performance in natural language understanding and generation tasks. However, LLMs continue to encounter challenges with hallucinations, where models generate plausible but incorrect information. While several factors contribute to hallucinations, the impact of ill-formed prompts, prompts with ambiguous wording, incorrect grammar, or incomplete information, was relatively under explored. To address this, we introduce Multi-stage Prompt Refinement (MPR), a framework designed to systematically improve these ill-formed prompts across multiple stages. Each stage addresses specific errors such as punctuation, typographical mistakes, and misuse of key terms, using small language models (SLMs) fine-tuned for these tasks. MPR iteratively enhances the clarity of prompts with additional context and employs a self-reflection mechanism with ranking to prioritize the most relevant input. Experimental results on hallucination benchmarks show that prompts refined by MPR achieve over an 85~\% win rate compared to their original forms, demonstrating its effectiveness in reducing hallucinations and improving LLM output accuracy. Interestingly, we reveal that MPR can be combined with existing post-hoc hallucination mitigation frameworks, further enhancing its versatility. MPR provides a lightweight and adaptable solution for enhancing LLM reliability across various domains.

Woo-Chan Kim, Ji-Hoon Park, Seong-Whan Lee

Large language models (LLMs) have demonstrated state-of-the-art performance across a wide range of natural language processing tasks. However, high-performing models are typically accessible only via APIs, incurring substantial inference costs. Cascade methods address this by initially employing a cheaper model and escalating to a stronger one only when necessary. Nevertheless, existing cascade approaches struggle to select a reliable representative response and assess the overall reliability of free-form outputs, as they rely on exact text matching. To overcome these limitations, we propose Keyword-inspired Cascade (KiC), a novel framework for cost-efficient free-form text generation. KiC identifies the most representative answer among multiple outputs from a weaker model and evaluates the semantic alignment of other responses with it. Based on the degree of alignment, KiC determines whether to accept the weaker model's output or escalate to a stronger model. Experiments on three free-form text generation benchmarks show that KiC achieves 97.53 percent of GPT-4's accuracy while reducing API costs by 28.81 percent on average, and even outperforms GPT-4 in a specific benchmark.