Hyunsoo Cho

Kang Min Yoo, Junyeob Kim, Hyuhng Joon Kim et al.

Despite recent explosion of interests in in-context learning, the underlying mechanism and the precise impact of the quality of demonstrations remain elusive. Intuitively, ground-truth labels should have as much impact in in-context learning (ICL) as supervised learning, but recent work reported that the input-label correspondence is significantly less important than previously thought. Intrigued by this counter-intuitive observation, we re-examine the importance of ground-truth labels in in-context learning. With the introduction of two novel metrics, namely Label-Correctness Sensitivity and Ground-truth Label Effect Ratio (GLER), we were able to conduct quantifiable analysis on the impact of ground-truth label demonstrations. Through extensive analyses, we find that the correct input-label mappings can have varying impacts on the downstream in-context learning performances, depending on the experimental configuration. Through additional studies, we identify key components, such as the verbosity of prompt templates and the language model size, as the controlling factor to achieve more noise-resilient ICL.

Hyuhng Joon Kim, Hyunsoo Cho, Junyeob Kim et al.

Large-scale pre-trained language models (PLMs) are well-known for being capable of solving a task simply by conditioning a few input-label pairs dubbed demonstrations on a prompt without being explicitly tuned for the desired downstream task. Such a process (i.e., in-context learning), however, naturally leads to high reliance on the demonstrations which are usually selected from external datasets. In this paper, we propose self-generated in-context learning (SG-ICL), which generates demonstrations for in-context learning from PLM itself to minimize the reliance on the external demonstration. We conduct experiments on four different text classification tasks and show SG-ICL significantly outperforms zero-shot learning and is generally worth approximately 0.6 gold training samples. Moreover, our generated demonstrations show more consistent performance with low variance compared to randomly selected demonstrations from the training dataset.

Hyun Lee, Hyemin Jeong, Yejin Kim et al.

Modern multimodal large language models (MLLMs) typically keep the language model fixed and train a visual projector that maps the pixels into a sequence of tokens in its embedding space, so that images can be presented in essentially the same form as text. However, the language model has been optimized to operate on discrete, semantically meaningful tokens, while prevailing visual projectors transform an image into a long stream of continuous and highly correlated embeddings. This causes the visual tokens to behave differently from the word-like units that LLMs are originally trained to understand. We propose a novel Disentangled Visual Tokenization (DiVT) that clusters patch embeddings into coherent semantic units, so each token corresponds to a distinct visual concept instead of a rigid grid cell. DiVT further adapts its token budget to image complexity, providing an explicit accuracy-compute trade-off modifying neither the vision encoder nor the language model. Across diverse multimodal benchmarks, DiVT matches or surpasses baselines with significantly fewer visual tokens, demonstrating robustness under limited token budgets, significantly reducing memory cost and latency while making visual inputs more compatible with LLMs. Our code is available at https://github.com/snuviplab/DiVT.

Hyunsoo Cho, Hyuhng Joon Kim, Junyeob Kim et al.

Through in-context learning (ICL), large-scale language models are effective few-shot learners without additional model fine-tuning. However, the ICL performance does not scale well with the number of available training samples as it is limited by the inherent input length constraint of the underlying language model. Meanwhile, many studies have revealed that language models are also powerful feature extractors, allowing them to be utilized in a black-box manner and enabling the linear probing paradigm, where lightweight discriminators are trained on top of the pre-extracted input representations. This paper proposes prompt-augmented linear probing (PALP), a hybrid of linear probing and ICL, which leverages the best of both worlds. PALP inherits the scalability of linear probing and the capability of enforcing language models to derive more meaningful representations via tailoring input into a more conceivable form. Throughout in-depth investigations on various datasets, we verified that PALP significantly enhances the input representations closing the gap between ICL in the data-hungry scenario and fine-tuning in the data-abundant scenario with little training overhead, potentially making PALP a strong alternative in a black-box scenario.

Geonhee Ahn, Donghyun Lee, Hayoung Doo et al.

Large language models (LLMs) have enabled agentic AI systems for scientific discovery, but most approaches remain limited to textbased reasoning without automated experimental verification. We propose MIND, an LLM-driven framework for automated hypothesis validation in materials research. MIND organizes the scientific discovery process into hypothesis refinement, experimentation, and debate-based validation within a multi-agent pipeline. For experimental verification, the system integrates Machine Learning Interatomic Potentials, particularly SevenNet-Omni, enabling scalable in-silico experiments. We also provide a web-based user interface for automated hypothesis testing. The modular design allows additional experimental modules to be integrated, making the framework adaptable to broader scientific workflows. The code is available at: https://github.com/IMMS-Ewha/MIND, and a demonstration video at: https://youtu.be/lqiFe1OQzN4.

Youna Kim, Hyuhng Joon Kim, Cheonbok Park et al.

When using large language models (LLMs) in knowledge-intensive tasks, such as open-domain question answering, external context can bridge the gap between external knowledge and the LLMs' parametric knowledge. Recent research has been developed to amplify contextual knowledge over the parametric knowledge of LLMs with contrastive decoding approaches. While these approaches could yield truthful responses when relevant context is provided, they are prone to vulnerabilities when faced with noisy contexts. We extend the scope of previous studies to encompass noisy contexts and propose adaptive contrastive decoding (ACD) to leverage contextual influence effectively. ACD demonstrates improvements in open-domain question answering tasks compared to baselines, especially in robustness by remaining undistracted by noisy contexts in retrieval-augmented generation.

Hyunsoo Cho, Choonghyun Park, Jaewook Kang et al.

Out-of-distribution (OOD) detection aims to discern outliers from the intended data distribution, which is crucial to maintaining high reliability and a good user experience. Most recent studies in OOD detection utilize the information from a single representation that resides in the penultimate layer to determine whether the input is anomalous or not. Although such a method is straightforward, the potential of diverse information in the intermediate layers is overlooked. In this paper, we propose a novel framework based on contrastive learning that encourages intermediate features to learn layer-specialized representations and assembles them implicitly into a single representation to absorb rich information in the pre-trained language model. Extensive experiments in various intent classification and OOD datasets demonstrate that our approach is significantly more effective than other works.

Bruce W. Lee, Hyunsoo Cho, Kang Min Yoo

In this work, we (1) introduce Curriculum Instruction Tuning, (2) explore the potential advantages of employing diverse curriculum strategies, and (3) delineate a synthetic instruction-response generation framework that complements our theoretical approach. Distinct from the existing instruction tuning dataset, our generation pipeline is systematically structured to emulate the sequential and orderly characteristic of human learning. Additionally, we describe a methodology for generating instruction-response datasets that extensively span the various stages of human education, from middle school through the graduate level, utilizing educational subject catalogs. Before training, we meticulously organize the instruction data to ensure that questions escalate in difficulty regarding (A) the subject matter and (B) the intricacy of the instructions. The findings of our study reveal that substantial improvements in performance can be achieved through the mere application of curriculum ordering to instruction data (achieving gains of +4.76 on TruthfulQA, +2.98 on MMLU, +2.8 on OpenbookQA, and +1.28 on ARC-hard) compared to random shuffling. This enhancement is achieved without incurring additional computational expenses. Through comprehensive experimentation, we observe that the advantages of our proposed method are consistently evident across nine benchmarks.

Jaewoong Choi, Geonho Hwang, Hyunsoo Cho et al.

The ideally disentangled latent space in GAN involves the global representation of latent space with semantic attribute coordinates. In other words, considering that this disentangled latent space is a vector space, there exists the global semantic basis where each basis component describes one attribute of generated images. In this paper, we propose an unsupervised method for finding this global semantic basis in the intermediate latent space in GANs. This semantic basis represents sample-independent meaningful perturbations that change the same semantic attribute of an image on the entire latent space. The proposed global basis, called Fréchet basis, is derived by introducing Fréchet mean to the local semantic perturbations in a latent space. Fréchet basis is discovered in two stages. First, the global semantic subspace is discovered by the Fréchet mean in the Grassmannian manifold of the local semantic subspaces. Second, Fréchet basis is found by optimizing a basis of the semantic subspace via the Fréchet mean in the Special Orthogonal Group. Experimental results demonstrate that Fréchet basis provides better semantic factorization and robustness compared to the previous methods. Moreover, we suggest the basis refinement scheme for the previous methods. The quantitative experiments show that the refined basis achieves better semantic factorization while constrained on the same semantic subspace given by the previous method.

Jaewoong Choi, Geonho Hwang, Hyunsoo Cho et al.

The impressive success of style-based GANs (StyleGANs) in high-fidelity image synthesis has motivated research to understand the semantic properties of their latent spaces. In this paper, we approach this problem through a geometric analysis of latent spaces as a manifold. In particular, we propose a local dimension estimation algorithm for arbitrary intermediate layers in a pre-trained GAN model. The estimated local dimension is interpreted as the number of possible semantic variations from this latent variable. Moreover, this intrinsic dimension estimation enables unsupervised evaluation of disentanglement for a latent space. Our proposed metric, called Distortion, measures an inconsistency of intrinsic tangent space on the learned latent space. Distortion is purely geometric and does not require any additional attribute information. Nevertheless, Distortion shows a high correlation with the global-basis-compatibility and supervised disentanglement score. Our work is the first step towards selecting the most disentangled latent space among various latent spaces in a GAN without attribute labels.

Hyunsoo Cho, Choonghyun Park, Junyeop Kim et al.

As the size of the pre-trained language model (PLM) continues to increase, numerous parameter-efficient transfer learning methods have been proposed recently to compensate for the tremendous cost of fine-tuning. Despite the impressive results achieved by large pre-trained language models (PLMs) and various parameter-efficient transfer learning (PETL) methods on sundry benchmarks, it remains unclear if they can handle inputs that have been distributionally shifted effectively. In this study, we systematically explore how the ability to detect out-of-distribution (OOD) changes as the size of the PLM grows or the transfer methods are altered. Specifically, we evaluated various PETL techniques, including fine-tuning, Adapter, LoRA, and prefix-tuning, on three different intention classification tasks, each utilizing various language models with different scales.

Hyuhng Joon Kim, Hyunsoo Cho, Sang-Woo Lee et al.

When deploying machine learning systems to the wild, it is highly desirable for them to effectively leverage prior knowledge to the unfamiliar domain while also firing alarms to anomalous inputs. In order to address these requirements, Universal Domain Adaptation (UniDA) has emerged as a novel research area in computer vision, focusing on achieving both adaptation ability and robustness (i.e., the ability to detect out-of-distribution samples). While UniDA has led significant progress in computer vision, its application on language input still needs to be explored despite its feasibility. In this paper, we propose a comprehensive benchmark for natural language that offers thorough viewpoints of the model's generalizability and robustness. Our benchmark encompasses multiple datasets with varying difficulty levels and characteristics, including temporal shifts and diverse domains. On top of our testbed, we validate existing UniDA methods from computer vision and state-of-the-art domain adaptation techniques from NLP literature, yielding valuable findings: We observe that UniDA methods originally designed for image input can be effectively transferred to the natural language domain while also underscoring the effect of adaptation difficulty in determining the model's performance.

Hyunsoo Cho, Youna Kim, Sang-goo Lee

Utilizing language models (LMs) without internal access is becoming an attractive paradigm in the field of NLP as many cutting-edge LMs are released through APIs and boast a massive scale. The de-facto method in this type of black-box scenario is known as prompting, which has shown progressive performance enhancements in situations where data labels are scarce or unavailable. Despite their efficacy, they still fall short in comparison to fully supervised counterparts and are generally brittle to slight modifications. In this paper, we propose Clustering-enhanced Linear Discriminative Analysis, a novel approach that improves the text classification accuracy with a very weak-supervision signal (i.e., name of the labels). Our framework draws a precise decision boundary without accessing weights or gradients of the LM model or data labels. The core ideas of CELDA are twofold: (1) extracting a refined pseudo-labeled dataset from an unlabeled dataset, and (2) training a lightweight and robust model on the top of LM, which learns an accurate decision boundary from an extracted noisy dataset. Throughout in-depth investigations on various datasets, we demonstrated that CELDA reaches new state-of-the-art in weakly-supervised text classification and narrows the gap with a fully-supervised model. Additionally, our proposed methodology can be applied universally to any LM and has the potential to scale to larger models, making it a more viable option for utilizing large LMs.

Bruce W. Lee, Yeongheon Lee, Hyunsoo Cho

Large Language Models (LLMs) exhibit non-deterministic behavior, and prompting has emerged as a primary method for steering their outputs toward desired directions. One popular strategy involves assigning a specific "persona" to the model to induce more varied and context-sensitive responses, akin to the diversity found in human perspectives. However, contrary to the expectation that persona-based prompting would yield a wide range of opinions, our experiments demonstrate that LLMs maintain consistent value orientations. In particular, we observe a persistent inertia in their responses, where certain moral and value dimensions, especially harm avoidance and fairness, remain distinctly skewed in one direction despite varied persona settings. To investigate this phenomenon systematically, use role-play at scale, which combines randomized, diverse persona prompts with a macroscopic trend analysis of model outputs. Our findings highlight the strong internal biases and value preferences in LLMs, underscoring the need for careful scrutiny and potential adjustment of these models to ensure balanced and equitable applications.

Hyunsoo Cho

Many recent studies endeavor to improve open-source language models through imitation learning, and re-training on the synthetic instruction data from state-of-the-art proprietary models like ChatGPT and GPT-4. However, the innate nature of synthetic data inherently contains noisy data, giving rise to a substantial presence of low-quality data replete with erroneous responses, and flawed reasoning. Although we intuitively grasp the potential harm of noisy data, we lack a quantitative understanding of its impact. To this end, this paper explores the correlation between the degree of noise and its impact on language models through instruction tuning. We first introduce the Falsity-Controllable (FACO) dataset, which comprises pairs of true answers with corresponding reasoning, as well as false pairs to manually control the falsity ratio of the dataset.Through our extensive experiments, we found multiple intriguing findings of the correlation between the factuality of the dataset and instruction tuning: Specifically, we verified falsity of the instruction is highly relevant to various benchmark scores. Moreover, when LLMs are trained with false instructions, they learn to lie and generate fake unfaithful answers, even though they know the correct answer for the user request. Additionally, we noted that once the language model is trained with a dataset contaminated by noise, restoring its original performance is possible, but it failed to reach full performance.

Choonghyun Park, Hyuhng Joon Kim, Junyeob Kim et al.

AI Generated Text (AIGT) detectors are developed with texts from humans and LLMs of common tasks. Despite the diversity of plausible prompt choices, these datasets are generally constructed with a limited number of prompts. The lack of prompt variation can introduce prompt-specific shortcut features that exist in data collected with the chosen prompt, but do not generalize to others. In this paper, we analyze the impact of such shortcuts in AIGT detection. We propose Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts. FAILOpt effectively drops the detection performance of the target detector, comparable to other attacks based on adversarial in-context examples. We also utilize our method to enhance the robustness of the detector by mitigating the shortcuts. Based on the findings, we further train the classifier with the dataset augmented by FAILOpt prompt. The augmented classifier exhibits improvements across generation models, tasks, and attacks. Our code will be available at https://github.com/zxcvvxcz/FAILOpt.

Doyeon Lee, Eunyi Lyou, Hyunsoo Cho et al.

GRPO-style reinforcement learning (RL)-based LLM fine-tuning algorithms have recently gained popularity. Relying on heuristic trust-region approximations, however, they can lead to brittle optimization behavior, as global importance-ratio clipping and group-wise normalization fail to regulate samples whose importance ratios fall outside the clipping range. We propose Query-Adaptive Trust-Region policy Optimization (QUATRO), which directly enforces trust-region constraints through a principled optimization. This yields a clear and interpretable objective that enables explicit control over policy updates and stable, entropy-controlled optimization, with a stabilizer terms arising intrinsically from the exact trust-region formulation. Empirically verified on diverse mathematical reasoning benchmarks, QUATRO shows stable training under increased policy staleness and aggressive learning rates, maintaining well-controlled entropy throughout training.

Kang Min Yoo, Jaegeun Han, Sookyo In et al.

We introduce HyperCLOVA X, a family of large language models (LLMs) tailored to the Korean language and culture, along with competitive capabilities in English, math, and coding. HyperCLOVA X was trained on a balanced mix of Korean, English, and code data, followed by instruction-tuning with high-quality human-annotated datasets while abiding by strict safety guidelines reflecting our commitment to responsible AI. The model is evaluated across various benchmarks, including comprehensive reasoning, knowledge, commonsense, factuality, coding, math, chatting, instruction-following, and harmlessness, in both Korean and English. HyperCLOVA X exhibits strong reasoning capabilities in Korean backed by a deep understanding of the language and cultural nuances. Further analysis of the inherent bilingual nature and its extension to multilingualism highlights the model's cross-lingual proficiency and strong generalization ability to untargeted languages, including machine translation between several language pairs and cross-lingual inference tasks. We believe that HyperCLOVA X can provide helpful guidance for regions or countries in developing their sovereign LLMs.

Jeonghyun Kim, SooKyung Kim, Richeng Xuan et al.

The core of knowledge distillation lies in transferring the teacher's rich 'dark knowledge'-subtle probabilistic patterns that reveal how classes are related and the distribution of uncertainties. While this idea is well established, teachers trained with conventional cross-entropy often fail to preserve such signals. Their distributions collapse into sharp, overconfident peaks that appear decisive but are in fact brittle, offering little beyond the hard label or subtly hindering representation-level transfer. This overconfidence is especially problematic in high-cardinality tasks, where the nuances among many plausible classes matter most for guiding a compact student. Moreover, such brittle targets reduce robustness under distribution shift, leaving students vulnerable to miscalibration in real-world conditions. To address this limitation, we revisit distillation from a distributional perspective and propose Calibrated Uncertainty Distillation (CUD), a framework designed to make dark knowledge more faithfully accessible. Instead of uncritically adopting the teacher's overconfidence, CUD encourages teachers to reveal uncertainty where it is informative and guides students to learn from targets that are calibrated rather than sharpened certainty. By directly shaping the teacher's predictive distribution before transfer, our approach balances accuracy and calibration, allowing students to benefit from both confident signals on easy cases and structured uncertainty on hard ones. Across diverse benchmarks, CUD yields students that are not only more accurate, but also more calibrated under shift and more reliable on ambiguous, long-tail inputs.

Lovisa Hagström, Youna Kim, Haeun Yu et al.

Incorporating external knowledge is crucial for knowledge-intensive tasks, such as question answering and fact checking. However, language models (LMs) may ignore relevant information that contradicts outdated parametric memory or be distracted by irrelevant contexts. While many context utilisation manipulation techniques (CMTs) have recently been proposed to alleviate these issues, few have seen systematic comparison. In this paper, we develop CUB (Context Utilisation Benchmark) - the first comprehensive benchmark designed to help practitioners within retrieval-augmented generation (RAG) diagnose CMTs under different context conditions. With this benchmark, we conduct the most extensive evaluation to date of seven state-of-the-art methods, representative of the main categories of CMTs, across three diverse datasets and tasks, applied to nine LMs. Our results reveal that most existing CMTs struggle to handle the full spectrum of context types encountered in real-world retrieval-augmented scenarios. We also find that many CMTs display inflated performance on simple synthesised datasets, compared to more realistic datasets with naturally occurring samples. Our findings expose critical gaps in current CMT evaluation practices and demonstrate the need for holistic testing and the development of CMTs that can robustly handle multiple context types.

Hyemin Boo, Hyoryung Kim, Myungjin Lee et al.

Text-to-image diffusion models generate realistic and coherent images but often fail to follow numerical instructions in text, revealing a gap between language and visual representation. Interestingly, we found that these models are not entirely blind to numbers-they are implicitly aware of their own counting accuracy, as their internal signals shift in consistent ways depending on whether the output meets the specified count. This observation suggests that the model already encodes a latent notion of numerical correctness, which can be harnessed to guide generation more precisely. Building on this intuition, we introduce CountSteer, a training-free method that improves generation of specified object counts by steering the model's cross-attention hidden states during inference. In our experiments, CountSteer improved object-count accuracy by about 4% without compromising visual quality, demonstrating a simple yet effective step toward more controllable and semantically reliable text-to-image generation.

Minsuh Joo, Hyunsoo Cho

Despite the outstanding performance of large language models (LLMs) across various NLP tasks, hallucinations in LLMs--where LLMs generate inaccurate responses--remains as a critical problem as it can be directly connected to a crisis of building safe and reliable LLMs. Uncertainty estimation is primarily used to measure hallucination levels in LLM responses so that correct and incorrect answers can be distinguished clearly. This study proposes an effective uncertainty estimation approach, \textbf{Cl}ust\textbf{e}ring-based sem\textbf{an}tic con\textbf{s}ist\textbf{e}ncy (\textbf{Cleanse}). Cleanse quantifies the uncertainty with the proportion of the intra-cluster consistency in the total consistency between LLM hidden embeddings which contain adequate semantic information of generations, by employing clustering. The effectiveness of Cleanse for detecting hallucination is validated using four off-the-shelf models, LLaMA-7B, LLaMA-13B, LLaMA2-7B and Mistral-7B and two question-answering benchmarks, SQuAD and CoQA.

Youna Kim, Hyuhng Joon Kim, Minjoon Choi et al.

Language models often benefit from external knowledge beyond parametric knowledge. While this combination enhances performance, achieving reliable knowledge utilization remains challenging, as it requires assessing the state of each knowledge source based on the presence of relevant information. Yet, prior work on knowledge integration often overlooks this challenge by assuming ideal conditions and provides limited coverage of knowledge scenarios. To address this gap, we introduce UniKnow, a Unified framework for reliable LM behavior across parametric and external Knowledge. UniKnow enables controlled evaluation across knowledge scenarios such as knowledge conflict, distraction, and absence conditions that are rarely addressed together. Beyond evaluating existing methods under this setting, we extend our work by introducing UniKnow-Aware methods to support comprehensive evaluation. Experiments on UniKnow reveal that existing methods struggle to generalize across a broader range of knowledge configurations and exhibit scenario-specific biases. UniKnow thus provides a foundation for systematically exploring and improving reliability under knowledge scenarios.

Hoin Jung, Hyunsoo Cho, Myungje Choi et al.

When it comes to a personalized item recommendation system, It is essential to extract users' preferences and purchasing patterns. Assuming that users in the real world form a cluster and there is common favoritism in each cluster, in this work, we introduce Co-Clustering Wrapper (CCW). We compute co-clusters of users and items with co-clustering algorithms and add CF subnetworks for each cluster to extract the in-group favoritism. Combining the features from the networks, we obtain rich and unified information about users. We experimented real world datasets considering two aspects: Finding the number of groups divided according to in-group preference, and measuring the quantity of improvement of the performance.

Hyunsoo Cho, Jinseok Seol, Sang-goo Lee

Detecting anomalies is one fundamental aspect of a safety-critical software system, however, it remains a long-standing problem. Numerous branches of works have been proposed to alleviate the complication and have demonstrated their efficiencies. In particular, self-supervised learning based methods are spurring interest due to their capability of learning diverse representations without additional labels. Among self-supervised learning tactics, contrastive learning is one specific framework validating their superiority in various fields, including anomaly detection. However, the primary objective of contrastive learning is to learn task-agnostic features without any labels, which is not entirely suited to discern anomalies. In this paper, we propose a task-specific variant of contrastive learning named masked contrastive learning, which is more befitted for anomaly detection. Moreover, we propose a new inference method dubbed self-ensemble inference that further boosts performance by leveraging the ability learned through auxiliary self-supervision tasks. By combining our models, we can outperform previous state-of-the-art methods by a significant margin on various benchmark datasets.