arXiv subjects

Jihwan Bang

Publications and source records attributed to Jihwan Bang.

At least 19 recordsLinked to original sources

arXiv2026

Training-Free Active Test-Time Adaptation for Vision-Language Models

Test-time adaptation (TTA) for vision-language models (VLMs) is usually studied in a fully unsupervised regime, where adaptation relies on pseudo-labels, confidence scores, or entropy computed by the model itself. This is a severe restriction on hard shifted samples: the model is asked to repair its own uncertain predictions without any external evidence. At the other extreme, supervised online adaptation assumes labels can directly train or correct the model. We study the intermediate deployment regime between these two extremes. In Active Test-time Adaptation for VLMs (Active VLM-TTA), a pretrained VLM may spend an explicit labeling budget on selected test samples, but the prediction for the current sample must be emitted before its label can be used; queried labels are therefore costed, delayed evidence for future samples rather than free correction for the present one. Under this future-only constraint, the central question is how to reuse sparse labels conservatively. We instantiate this setting with ACTive VLM-TTA for Online Retrieval (ACTOR), a training-free plug-in module that routes uncertain samples among base prediction, label query, and memory-based correction using a consistency-and-reliability check over previously queried labels. Across ten cross-domain benchmarks and a domain-transfer benchmark based on ImageNet, ACTOR consistently improves strong VLM-TTA baselines on ResNet and ViT backbones with negligible inference overhead. Our implementation is publicly available at https://github.com/kaist-dmlab/ACTOR.

cs.CV↗

arXiv2026

How Much Can We Prune? Exploring the Performance Limits of Visual Token Pruning

Visual token pruning is increasingly used to reduce the memory and inference costs of vision-language models. Existing work has largely focused on designing method-specific selectors and demonstrating incremental improvements over prior methods at selected token budgets. It remains unclear whether the resulting performance reflects an inherent limit of visual token pruning or the limitations of current token-selection algorithms. We take a complementary perspective and ask: \emph{how well can visual token pruning perform when offline search cost is not the primary constraint?} We formulate pruning for general image understanding as selecting, under a fixed budget, one shared visual-token subset that supports a broad distribution of questions about an image. We approximate this utility with a finite auxiliary QA set generated by a frontier VLM and optimize a teacher-forced answer-probability objective with the evaluated model frozen. Motivated by the conceptual monotonicity and diminishing returns of adding visual evidence, we construct nested subsets greedily and introduce \emph{QA mini-batch greedy} (QA-MBG), which estimates token marginal gains from randomly sampled QA components at each step. Empirically, QA-MBG is both cheaper and more effective than full greedy. With one frontier-VLM call and approximately $0.5$ H100 GPU-hours per image, it produces a nested subset trajectory covering keep ratios up to $0.3$. Across Qwen3-VL-4B and Gemma~4 E4B on $900$ images from nine benchmarks, we find significant headroom over evaluated pruning baselines on held-out QA and image-grounded long-form generation, while the searched subsets substantially narrow the gap to full-token inference.

cs.CV↗

arXiv2026

Rethinking RAG in Long Videos: What to Retrieve and How to Use It?

Retrieval-augmented generation is extending beyond text to long videos, where query-relevant chunks can be represented across multiple modalities and temporal granularities. Progress in this setting, VideoRAG, is limited by two gaps: existing benchmarks allow queries to be answered without the video, obscuring retrieval errors, and prior methods apply a single modality-granularity configuration per query, ignoring chunk-level variability. We address both by introducing V-RAGBench, a benchmark of $\langle$query, evidence chunk, answer$\rangle$ triplets over hour-scale videos, in which each answer depends on a unique evidence chunk, enabling decoupled evaluation of retrieval and generation, and CARVE, a training-free method that runs parallel retrievers across configurations and uses chunk-adaptive reranking to select a winning configuration for each chunk, which is then carried into generation. On V-RAGBench, CARVE outperforms eight recent VideoRAG baselines on both stages, with the evidence it supplies to the generator interleaving multiple configurations rather than sharing one. Its gains hold across egocentric and third-person long videos and extend to an expanded configuration space.

cs.AI↗

arXiv2026

Feedback Adaptation for Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) systems are typically evaluated under static assumptions, despite being frequently corrected through user or expert feedback in deployment. Existing evaluation protocols focus on overall accuracy and fail to capture how systems adapt after feedback is introduced. We introduce feedback adaptation as a problem setting for RAG systems, which asks how effectively and how quickly corrective feedback propagates to future queries. To make this behavior measurable, we propose two evaluation axes: correction lag, which captures the delay between feedback provision and behavioral change, and post-feedback performance, which measures reliability on semantically related queries after feedback. Using these metrics, we show that training-based approaches exhibit a trade-off between delayed correction and reliable adaptation. We further propose PatchRAG, a minimal inference-time instantiation that incorporates feedback without retraining, demonstrating immediate correction and strong post-feedback generalization under the proposed evaluation. Our results highlight feedback adaptation as a previously overlooked dimension of RAG system behavior in interactive settings.

cs.CL↗

arXiv2025

Think Straight, Stop Smart: Structured Reasoning for Efficient Multi-Hop RAG

Multi-hop retrieval-augmented generation (RAG) is a promising strategy for complex reasoning, yet existing iterative prompting approaches remain inefficient. They often regenerate predictable token sequences at every step and rely on stochastic stopping, leading to excessive token usage and unstable termination. We propose TSSS (Think Straight, Stop Smart), a structured multi-hop RAG framework designed for efficiency. TSSS introduces (i) a template-based reasoning that caches recurring prefixes and anchors sub-queries to the main question, reducing token generation cost while promoting stable reasoning, and (ii) a retriever-based terminator, which deterministically halts reasoning once additional sub-queries collapse into repetition. This separation of structured reasoning and termination control enables both faster inference and more reliable answers. On HotpotQA, 2WikiMultiHop, and MuSiQue, TSSS achieves state-of-the-art accuracy and competitive efficiency among RAG-CoT approaches, highlighting its effectiveness in efficiency-constrained scenarios such as on-device inference.

cs.CL↗

arXiv2025

References Indeed Matter? Reference-Free Preference Optimization for Conversational Query Reformulation

Conversational query reformulation (CQR) has become indispensable for improving retrieval in dialogue-based applications. However, existing approaches typically rely on reference passages for optimization, which are impractical to acquire in real-world scenarios. To address this limitation, we introduce a novel reference-free preference optimization framework DualReform that generates pseudo reference passages from commonly-encountered conversational datasets containing only queries and responses. DualReform attains this goal through two key innovations: (1) response-based inference, where responses serve as proxies to infer pseudo reference passages, and (2) response refinement via the dual-role of CQR, where a CQR model refines responses based on the shared objectives between response refinement and CQR. Despite not relying on reference passages, DualReform achieves 96.9--99.1% of the retrieval accuracy attainable only with reference passages and surpasses the state-of-the-art method by up to 31.6%.

cs.CL↗

arXiv2025

Learning Contextual Retrieval for Robust Conversational Search

Effective conversational search demands a deep understanding of user intent across multiple dialogue turns. Users frequently use abbreviations and shift topics in the middle of conversations, posing challenges for conventional retrievers. While query rewriting techniques improve clarity, they often incur significant computational cost due to additional autoregressive steps. Moreover, although LLM-based retrievers demonstrate strong performance, they are not explicitly optimized to track user intent in multi-turn settings, often failing under topic drift or contextual ambiguity. To address these limitations, we propose ContextualRetriever, a novel LLM-based retriever that directly incorporates conversational context into the retrieval process. Our approach introduces: (1) a context-aware embedding mechanism that highlights the current query within the dialogue history; (2) intent-guided supervision based on high-quality rewritten queries; and (3) a training strategy that preserves the generative capabilities of the base LLM. Extensive evaluations across multiple conversational search benchmarks demonstrate that ContextualRetriever significantly outperforms existing methods while incurring no additional inference overhead.

cs.IR↗

arXiv2025

CIFLEX: Contextual Instruction Flow for Sub-task Execution in Multi-Turn Interactions with a Single On-Device LLM

We present CIFLEX (Contextual Instruction Flow for Sub-task Execution), which is a novel execution system for efficient sub-task handling in multi-turn interactions with a single on-device large language model (LLM). As LLMs become increasingly capable, a single model is expected to handle diverse sub-tasks that more effectively and comprehensively support answering user requests. Naive approach reprocesses the entire conversation context when switching between main and sub-tasks (e.g., query rewriting, summarization), incurring significant computational overhead. CIFLEX mitigates this overhead by reusing the key-value (KV) cache from the main task and injecting only task-specific instructions into isolated side paths. After sub-task execution, the model rolls back to the main path via cached context, thereby avoiding redundant prefill computation. To support sub-task selection, we also develop a hierarchical classification strategy tailored for small-scale models, decomposing multi-choice decisions into binary ones. Experiments show that CIFLEX significantly reduces computational costs without degrading task performance, enabling scalable and efficient multi-task dialogue on-device.

cs.CL↗

arXiv2025

ReFeed: Multi-dimensional Summarization Refinement with Reflective Reasoning on Feedback

Summarization refinement faces challenges when extending to multi-dimension. In this paper, we introduce ReFeed, a powerful summarization refinement pipeline that enhances multiple dimensions through reflective reasoning on feedback. To achieve this, we release SumFeed-CoT, a large-scale Long-CoT-based dataset optimized for training a lightweight model with reflective reasoning. Our experiments reveal how the number of dimensions, feedback exposure, and reasoning policy influence refinement performance, highlighting reflective reasoning and simultaneously addressing multiple feedback is crucial to mitigate trade-off between dimensions. Furthermore, ReFeed is robust to noisy feedback and feedback order. Lastly, our finding emphasizes that creating data with a proper goal and guideline constitutes a fundamental pillar of effective reasoning. The dataset and model will be released.

cs.CL↗

arXiv2025

Chain-of-Rank: Enhancing Large Language Models for Domain-Specific RAG in Edge Device

Retrieval-augmented generation (RAG) with large language models (LLMs) is especially valuable in specialized domains, where precision is critical. To more specialize the LLMs into a target domain, domain-specific RAG has recently been developed by allowing the LLM to access the target domain early via finetuning. The domain-specific RAG makes more sense in resource-constrained environments like edge devices, as they should perform a specific task (e.g. personalization) reliably using only small-scale LLMs. While the domain-specific RAG is well-aligned with edge devices in this respect, it often relies on widely-used reasoning techniques like chain-of-thought (CoT). The reasoning step is useful to understand the given external knowledge, and yet it is computationally expensive and difficult for small-scale LLMs to learn it. Tackling this, we propose the Chain of Rank (CoR) which shifts the focus from intricate lengthy reasoning to simple ranking of the reliability of input external documents. Then, CoR reduces computational complexity while maintaining high accuracy, making it particularly suited for resource-constrained environments. We attain the state-of-the-art (SOTA) results in benchmarks, and analyze its efficacy.

cs.CL↗

arXiv2024

One Size Fits All for Semantic Shifts: Adaptive Prompt Tuning for Continual Learning

In real-world continual learning (CL) scenarios, tasks often exhibit intricate and unpredictable semantic shifts, posing challenges for fixed prompt management strategies which are tailored to only handle semantic shifts of uniform degree (i.e., uniformly mild or uniformly abrupt). To address this limitation, we propose an adaptive prompting approach that effectively accommodates semantic shifts of varying degree where mild and abrupt shifts are mixed. AdaPromptCL employs the assign-and-refine semantic grouping mechanism that dynamically manages prompt groups in accordance with the semantic similarity between tasks, enhancing the quality of grouping through continuous refinement. Our experiment results demonstrate that AdaPromptCL outperforms existing prompting methods by up to 21.3%, especially in the benchmark datasets with diverse semantic shifts between tasks.

cs.LG↗

arXiv2024

Crayon: Customized On-Device LLM via Instant Adapter Blending and Edge-Server Hybrid Inference

The customization of large language models (LLMs) for user-specified tasks gets important. However, maintaining all the customized LLMs on cloud servers incurs substantial memory and computational overheads, and uploading user data can also lead to privacy concerns. On-device LLMs can offer a promising solution by mitigating these issues. Yet, the performance of on-device LLMs is inherently constrained by the limitations of small-scaled models. To overcome these restrictions, we first propose Crayon, a novel approach for on-device LLM customization. Crayon begins by constructing a pool of diverse base adapters, and then we instantly blend them into a customized adapter without extra training. In addition, we develop a device-server hybrid inference strategy, which deftly allocates more demanding queries or non-customized tasks to a larger, more capable LLM on a server. This ensures optimal performance without sacrificing the benefits of on-device customization. We carefully craft a novel benchmark from multiple question-answer datasets, and show the efficacy of our method in the LLM customization.

cs.CL↗

arXiv2024

Active Prompt Learning in Vision Language Models

Pre-trained Vision Language Models (VLMs) have demonstrated notable progress in various zero-shot tasks, such as classification and retrieval. Despite their performance, because improving performance on new tasks requires task-specific knowledge, their adaptation is essential. While labels are needed for the adaptation, acquiring them is typically expensive. To overcome this challenge, active learning, a method of achieving a high performance by obtaining labels for a small number of samples from experts, has been studied. Active learning primarily focuses on selecting unlabeled samples for labeling and leveraging them to train models. In this study, we pose the question, "how can the pre-trained VLMs be adapted under the active learning framework?" In response to this inquiry, we observe that (1) simply applying a conventional active learning framework to pre-trained VLMs even may degrade performance compared to random selection because of the class imbalance in labeling candidates, and (2) the knowledge of VLMs can provide hints for achieving the balance before labeling. Based on these observations, we devise a novel active learning framework for VLMs, denoted as PCB. To assess the effectiveness of our approach, we conduct experiments on seven different real-world datasets, and the results demonstrate that PCB surpasses conventional active learning and random sampling methods. Code will be available in https://github.com/kaist-dmlab/pcb .

cs.CV↗

arXiv2023

Adaptive Shortcut Debiasing for Online Continual Learning

We propose a novel framework DropTop that suppresses the shortcut bias in online continual learning (OCL) while being adaptive to the varying degree of the shortcut bias incurred by continuously changing environment. By the observed high-attention property of the shortcut bias, highly-activated features are considered candidates for debiasing. More importantly, resolving the limitation of the online environment where prior knowledge and auxiliary data are not ready, two novel techniques -- feature map fusion and adaptive intensity shifting -- enable us to automatically determine the appropriate level and proportion of the candidate shortcut features to be dropped. Extensive experiments on five benchmark datasets demonstrate that, when combined with various OCL algorithms, DropTop increases the average accuracy by up to 10.4% and decreases the forgetting by up to 63.2%.

cs.LG↗

arXiv2023

Prompt-Guided Transformers for End-to-End Open-Vocabulary Object Detection

Prompt-OVD is an efficient and effective framework for open-vocabulary object detection that utilizes class embeddings from CLIP as prompts, guiding the Transformer decoder to detect objects in both base and novel classes. Additionally, our novel RoI-based masked attention and RoI pruning techniques help leverage the zero-shot classification ability of the Vision Transformer-based CLIP, resulting in improved detection performance at minimal computational cost. Our experiments on the OV-COCO and OVLVIS datasets demonstrate that Prompt-OVD achieves an impressive 21.2 times faster inference speed than the first end-to-end open-vocabulary detection method (OV-DETR), while also achieving higher APs than four two-stage-based methods operating within similar inference time ranges. Code will be made available soon.

cs.CV↗

arXiv2023

Meta-Query-Net: Resolving Purity-Informativeness Dilemma in Open-set Active Learning

Unlabeled data examples awaiting annotations contain open-set noise inevitably. A few active learning studies have attempted to deal with this open-set noise for sample selection by filtering out the noisy examples. However, because focusing on the purity of examples in a query set leads to overlooking the informativeness of the examples, the best balancing of purity and informativeness remains an important question. In this paper, to solve this purity-informativeness dilemma in open-set active learning, we propose a novel Meta-Query-Net,(MQ-Net) that adaptively finds the best balancing between the two factors. Specifically, by leveraging the multi-round property of active learning, we train MQ-Net using a query set without an additional validation set. Furthermore, a clear dominance relationship between unlabeled examples is effectively captured by MQ-Net through a novel skyline regularization. Extensive experiments on multiple open-set active learning scenarios demonstrate that the proposed MQ-Net achieves 20.14% improvement in terms of accuracy, compared with the state-of-the-art methods.

cs.LG↗

arXiv2022

Online Continual Learning on a Contaminated Data Stream with Blurry Task Boundaries

Learning under a continuously changing data distribution with incorrect labels is a desirable real-world problem yet challenging. A large body of continual learning (CL) methods, however, assumes data streams with clean labels, and online learning scenarios under noisy data streams are yet underexplored. We consider a more practical CL task setup of an online learning from blurry data stream with corrupted labels, where existing CL methods struggle. To address the task, we first argue the importance of both diversity and purity of examples in the episodic memory of continual learning models. To balance diversity and purity in the episodic memory, we propose a novel strategy to manage and use the memory by a unified approach of label noise aware diverse sampling and robust learning with semi-supervised learning. Our empirical validations on four real-world or synthetic noise datasets (CIFAR10 and 100, mini-WebVision, and Food-101N) exhibit that our method significantly outperforms prior arts in this realistic and challenging continual learning scenario. Code and data splits are available in https://github.com/clovaai/puridiver.

cs.CV↗

arXiv2021

Rainbow Memory: Continual Learning with a Memory of Diverse Samples

Continual learning is a realistic learning scenario for AI models. Prevalent scenario of continual learning, however, assumes disjoint sets of classes as tasks and is less realistic rather artificial. Instead, we focus on 'blurry' task boundary; where tasks shares classes and is more realistic and practical. To address such task, we argue the importance of diversity of samples in an episodic memory. To enhance the sample diversity in the memory, we propose a novel memory management strategy based on per-sample classification uncertainty and data augmentation, named Rainbow Memory (RM). With extensive empirical validations on MNIST, CIFAR10, CIFAR100, and ImageNet datasets, we show that the proposed method significantly improves the accuracy in blurry continual learning setups, outperforming state of the arts by large margins despite its simplicity. Code and data splits will be available in https://github.com/clovaai/rainbow-memory.

cs.CV↗