Se-Young Yun

Namgyu Ho, Laura Schmid, Se-Young Yun

Recent works have shown that chain-of-thought (CoT) prompting can elicit language models to solve complex reasoning tasks, step-by-step. However, prompt-based CoT methods are dependent on very large models such as GPT-3 175B which are prohibitive to deploy at scale. In this paper, we use these large models as reasoning teachers to enable complex reasoning in smaller models and reduce model size requirements by several orders of magnitude. We propose Fine-tune-CoT, a method that generates reasoning samples from very large teacher models to fine-tune smaller models. We evaluate our method on a wide range of public models and complex tasks. We find that Fine-tune-CoT enables substantial reasoning capability in small models, far outperforming prompt-based baselines and even the teacher model in many tasks. Additionally, we extend our method by leveraging the teacher model's ability to generate multiple distinct rationales for each original sample. Enriching the fine-tuning data with such diverse reasoning results in a substantial performance boost across datasets, even for very small models. We conduct ablations and sample studies to understand the emergence of reasoning capabilities of student models. Our code implementation and data are available at https://github.com/itsnamgyu/reasoning-teacher.

Hojoon Lee, Hanseul Cho, Hyunseung Kim et al.

In Reinforcement Learning (RL), enhancing sample efficiency is crucial, particularly in scenarios when data acquisition is costly and risky. In principle, off-policy RL algorithms can improve sample efficiency by allowing multiple updates per environment interaction. However, these multiple updates often lead the model to overfit to earlier interactions, which is referred to as the loss of plasticity. Our study investigates the underlying causes of this phenomenon by dividing plasticity into two aspects. Input plasticity, which denotes the model's adaptability to changing input data, and label plasticity, which denotes the model's adaptability to evolving input-output relationships. Synthetic experiments on the CIFAR-10 dataset reveal that finding smoother minima of loss landscape enhances input plasticity, whereas refined gradient propagation improves label plasticity. Leveraging these findings, we introduce the PLASTIC algorithm, which harmoniously combines techniques to address both concerns. With minimal architectural modifications, PLASTIC achieves competitive performance on benchmarks including Atari-100k and Deepmind Control Suite. This result emphasizes the importance of preserving the model's plasticity to elevate the sample efficiency in RL. The code is available at https://github.com/dojeon-ai/plastic.

Gihun Lee, Sangmook Kim, Joonkee Kim et al.

Cell segmentation is a fundamental task for computational biology analysis. Identifying the cell instances is often the first step in various downstream biomedical studies. However, many cell segmentation algorithms, including the recently emerging deep learning-based methods, still show limited generality under the multi-modality environment. Weakly Supervised Cell Segmentation in Multi-modality High-Resolution Microscopy Images was hosted at NeurIPS 2022 to tackle this problem. We propose MEDIAR, a holistic pipeline for cell instance segmentation under multi-modality in this challenge. MEDIAR harmonizes data-centric and model-centric approaches as the learning and inference strategies, achieving a 0.9067 F1-score at the validation phase while satisfying the time budget. To facilitate subsequent research, we provide the source code and trained model as open-source: https://github.com/Lee-Gihun/MEDIAR

Yongjin Yang, Joonkee Kim, Yujin Kim et al.

With the proliferation of social media, accurate detection of hate speech has become critical to ensure safety online. To combat nuanced forms of hate speech, it is important to identify and thoroughly explain hate speech to help users understand its harmful effects. Recent benchmarks have attempted to tackle this issue by training generative models on free-text annotations of implications in hateful text. However, we find significant reasoning gaps in the existing annotations schemes, which may hinder the supervision of detection models. In this paper, we introduce a hate speech detection framework, HARE, which harnesses the reasoning capabilities of large language models (LLMs) to fill these gaps in explanations of hate speech, thus enabling effective supervision of detection models. Experiments on SBIC and Implicit Hate benchmarks show that our method, using model-generated data, consistently outperforms baselines, using existing free-text human annotations. Analysis demonstrates that our method enhances the explanation quality of trained models and improves generalization to unseen datasets. Our code is available at https://github.com/joonkeekim/hare-hate-speech.git.

Jongwoo Ko, Seungjoon Park, Minchan Jeong et al.

Knowledge distillation (KD) is a highly promising method for mitigating the computational problems of pre-trained language models (PLMs). Among various KD approaches, Intermediate Layer Distillation (ILD) has been a de facto standard KD method with its performance efficacy in the NLP field. In this paper, we find that existing ILD methods are prone to overfitting to training datasets, although these methods transfer more information than the original KD. Next, we present the simple observations to mitigate the overfitting of ILD: distilling only the last Transformer layer and conducting ILD on supplementary tasks. Based on our two findings, we propose a simple yet effective consistency-regularized ILD (CR-ILD), which prevents the student model from overfitting the training dataset. Substantial experiments on distilling BERT on the GLUE benchmark and several synthetic datasets demonstrate that our proposed ILD method outperforms other KD techniques. Our code is available at https://github.com/jongwooko/CR-ILD.

Yujin Kim, Jaehoon Oh, Sungnyun Kim et al.

Most of the recent few-shot learning (FSL) algorithms are based on transfer learning, where a model is pre-trained using a large amount of source data, and the pre-trained model is fine-tuned using a small amount of target data. In transfer learning-based FSL, sophisticated pre-training methods have been widely studied for universal representation. Therefore, it has become more important to utilize the universal representation for downstream tasks, but there are few studies on fine-tuning in FSL. In this paper, we focus on how to transfer pre-trained models to few-shot downstream tasks from the three perspectives: update, data augmentation, and test-time augmentation. First, we compare the two popular update methods, full fine-tuning (i.e., updating the entire network, FT) and linear probing (i.e., updating only a linear classifier, LP). We find that LP is better than FT with extremely few samples, whereas FT outperforms LP as training samples increase. Next, we show that data augmentation cannot guarantee few-shot performance improvement and investigate the effectiveness of data augmentation based on the intensity of augmentation. Finally, we adopt augmentation to both a support set for update (i.e., data augmentation) as well as a query set for prediction (i.e., test-time augmentation), considering support-query distribution shifts, and improve few-shot performance. The code is available at https://github.com/kimyuji/updating_FSL.

Jaehoon Oh, Jongwoo Ko, Se-Young Yun

Translation has played a crucial role in improving the performance on multilingual tasks: (1) to generate the target language data from the source language data for training and (2) to generate the source language data from the target language data for inference. However, prior works have not considered the use of both translations simultaneously. This paper shows that combining them can synergize the results on various multilingual sentence classification tasks. We empirically find that translation artifacts stylized by translators are the main factor of the performance gain. Based on this analysis, we adopt two training methods, SupCon and MixUp, considering translation artifacts. Furthermore, we propose a cross-lingual fine-tuning algorithm called MUSC, which uses SupCon and MixUp jointly and improves the performance. Our code is available at https://github.com/jongwooko/MUSC.

Taehyeon Kim, Namgyu Ho, Donggyu Kim et al.

Precipitation forecasting is an important scientific challenge that has wide-reaching impacts on society. Historically, this challenge has been tackled using numerical weather prediction (NWP) models, grounded on physics-based simulations. Recently, many works have proposed an alternative approach, using end-to-end deep learning (DL) models to replace physics-based NWP models. While these DL methods show improved performance and computational efficiency, they exhibit limitations in long-term forecasting and lack the explainability. In this work, we present a hybrid NWP-DL workflow to fill the gap between standalone NWP and DL approaches. Under this workflow, the outputs of NWP models are fed into a deep neural network, which post-processes the data to yield a refined precipitation forecast. The deep model is trained with supervision, using Automatic Weather Station (AWS) observations as ground-truth labels. This can achieve the best of both worlds, and can even benefit from future improvements in NWP technology. To facilitate study in this direction, we present a novel dataset focused on the Korean Peninsula, termed KoMet (Korea Meteorological Dataset), comprised of NWP outputs and AWS observations. For the NWP model, the Global Data Assimilation and Prediction Systems-Korea Integrated Model (GDAPS-KIM) is utilized. We provide analysis on a comprehensive set of baseline methods aimed at addressing the challenges of KoMet, including the sparsity of AWS observations and class imbalance. To lower the barrier to entry and encourage further study, we also provide an extensive open-source Python package for data processing and model development. Our benchmark data and code are available at https://github.com/osilab-kaist/KoMet-Benchmark-Dataset.

Jongwoo Ko, Bongsoo Yi, Se-Young Yun

As deep neural networks can easily overfit noisy labels, robust training in the presence of noisy labels is becoming an important challenge in modern deep learning. While existing methods address this problem in various directions, they still produce unpredictable sub-optimal results since they rely on the posterior information estimated by the feature extractor corrupted by noisy labels. Lipschitz regularization successfully alleviates this problem by training a robust feature extractor, but it requires longer training time and expensive computations. Motivated by this, we propose a simple yet effective method, called ALASCA, which efficiently provides a robust feature extractor under label noise. ALASCA integrates two key ingredients: (1) adaptive label smoothing based on our theoretical analysis that label smoothing implicitly induces Lipschitz regularization, and (2) auxiliary classifiers that enable practical application of intermediate Lipschitz regularization with negligible computations. We conduct wide-ranging experiments for ALASCA and combine our proposed method with previous noise-robust methods on several synthetic and real-world datasets. Experimental results show that our framework consistently improves the robustness of feature extractors and the performance of existing baselines with efficiency. Our code is available at https://github.com/jongwooko/ALASCA.

Yassir Jedra, Junghyun Lee, Alexandre Proutière et al.

We investigate the problems of model estimation and reward-free learning in episodic Block MDPs. In these MDPs, the decision maker has access to rich observations or contexts generated from a small number of latent states. We are first interested in estimating the latent state decoding function (the mapping from the observations to latent states) based on data generated under a fixed behavior policy. We derive an information-theoretical lower bound on the error rate for estimating this function and present an algorithm approaching this fundamental limit. In turn, our algorithm also provides estimates of all the components of the MDP. We then study the problem of learning near-optimal policies in the reward-free framework. Based on our efficient model estimation algorithm, we show that we can infer a policy converging (as the number of collected samples grows large) to the optimal policy at the best possible rate. Interestingly, our analysis provides necessary and sufficient conditions under which exploiting the block structure yields improvements in the sample complexity for identifying near-optimal policies. When these conditions are met, the sample complexity in the minimax reward-free setting is improved by a multiplicative factor $n$, where $n$ is the number of possible contexts.

Sumyeong Ahn, Jongwoo Ko, Se-Young Yun

Class imbalance problems frequently occur in real-world tasks, and conventional deep learning algorithms are well known for performance degradation on imbalanced training datasets. To mitigate this problem, many approaches have aimed to balance among given classes by re-weighting or re-sampling training samples. These re-balancing methods increase the impact of minority classes and reduce the influence of majority classes on the output of models. However, the extracted representations may be of poor quality owing to the limited number of minority samples. To handle this restriction, several methods have been developed that increase the representations of minority samples by leveraging the features of the majority samples. Despite extensive recent studies, no deep analysis has been conducted on determination of classes to be augmented and strength of augmentation has been conducted. In this study, we first investigate the correlation between the degree of augmentation and class-wise performance, and find that the proper degree of augmentation must be allocated for each class to mitigate class imbalance problems. Motivated by this finding, we propose a simple and efficient novel curriculum, which is designed to find the appropriate per-class strength of data augmentation, called CUDA: CUrriculum of Data Augmentation for long-tailed recognition. CUDA can simply be integrated into existing long-tailed recognition methods. We present the results of experiments showing that CUDA effectively achieves better generalization performance compared to the state-of-the-art method on various imbalanced datasets such as CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018.

Soowon Oh, Nam Cao, Yujin Kim et al.

Block-diffusion drafters have recently emerged as a powerful alternative for speculative decoding by predicting multiple future-token distributions in a single parallel step. However, since these parallel predictions are sampled from position-wise marginals rather than fully conditioned sequences, committing to a single greedy path often fails to capture the target model's preferred trajectory. To address this, we propose BASTION, a budget-aware speculative decoding framework with tree-based diffusion drafting. Unlike existing methods that rely on static tree topologies, BASTION dynamically constructs query-dependent trees by balancing draft quality against hardware constraints. Our framework integrates three synergistic components: (1) an acceptance surrogate that estimates expected accepted length via path confidence, (2) an online latency estimator that calibrates a hardware-aware roofline model, and (3) an adaptive best-first expansion that grows the tree until marginal gains no longer justify incremental verification costs. BASTION is training-free, preserves the target model's distribution, and requires no per-setting tuning. Across diverse benchmarks and GPU architectures, BASTION achieves up to a 6.61x speedup over standard autoregressive decoding, outperforming state-of-the-art block-diffusion baselines by 39%.

SangMook Kim, Sangmin Bae, Hwanjun Song et al.

Although federated learning has made awe-inspiring advances, most studies have assumed that the client's data are fully labeled. However, in a real-world scenario, every client may have a significant amount of unlabeled instances. Among the various approaches to utilizing unlabeled data, a federated active learning framework has emerged as a promising solution. In the decentralized setting, there are two types of available query selector models, namely 'global' and 'local-only' models, but little literature discusses their performance dominance and its causes. In this work, we first demonstrate that the superiority of two selector models depends on the global and local inter-class diversity. Furthermore, we observe that the global and local-only models are the keys to resolving the imbalance of each side. Based on our findings, we propose LoGo, a FAL sampling strategy robust to varying local heterogeneity levels and global imbalance ratio, that integrates both models by two steps of active selection scheme. LoGo consistently outperforms six active learning strategies in the total number of 38 experimental settings.

Sangmin Bae, Jongwoo Ko, Hwanjun Song et al.

To tackle the high inference latency exhibited by autoregressive language models, previous studies have proposed an early-exiting framework that allocates adaptive computation paths for each token based on the complexity of generating the subsequent token. However, we observed several shortcomings, including performance degradation caused by a state copying mechanism or numerous exit paths, and sensitivity to exit confidence thresholds. Consequently, we propose a Fast and Robust Early-Exiting (FREE) framework, which incorporates a shallow-deep module and a synchronized parallel decoding. Our framework enables faster inference by synchronizing the decoding process of the current token with previously stacked early-exited tokens. Furthermore, as parallel decoding allows us to observe predictions from both shallow and deep models, we present a novel adaptive threshold estimator that exploits a Beta mixture model to determine suitable confidence thresholds. We empirically demonstrated the superiority of our proposed framework on extensive generation tasks.

Yujin Kim, Jaehong Yoon, Seonghyeon Ye et al.

The dynamic nature of knowledge in an ever-changing world presents challenges for language models trained on static data; the model in the real world often requires not only acquiring new knowledge but also overwriting outdated information into updated ones. To study the ability of language models for these time-dependent dynamics in human language, we introduce a novel task, EvolvingQA, a temporally evolving question-answering benchmark designed for training and evaluating LMs on an evolving Wikipedia database. The construction of EvolvingQA is automated with our pipeline using large language models. We uncover that existing continual learning baselines suffer from updating and removing outdated knowledge. Our analysis suggests that models fail to rectify knowledge due to small weight gradients. In addition, we elucidate that language models particularly struggle to reflect the change of numerical or temporal information. Our work aims to model the dynamic nature of real-world information, suggesting faithful evaluations of the evolution-adaptability of language models.

Sumyeong Ahn, Seongyoon Kim, Se-young Yun

The performance of deep neural networks is strongly influenced by the training dataset setup. In particular, when attributes having a strong correlation with the target attribute are present, the trained model can provide unintended prejudgments and show significant inference errors (i.e., the dataset bias problem). Various methods have been proposed to mitigate dataset bias, and their emphasis is on weakly correlated samples, called bias-conflicting samples. These methods are based on explicit bias labels involving human or empirical correlation metrics (e.g., training loss). However, such metrics require human costs or have insufficient theoretical explanation. In this study, we propose a debiasing algorithm, called PGD (Per-sample Gradient-based Debiasing), that comprises three steps: (1) training a model on uniform batch sampling, (2) setting the importance of each sample in proportion to the norm of the sample gradient, and (3) training the model using importance-batch sampling, whose probability is obtained in step (2). Compared with existing baselines for various synthetic and real-world datasets, the proposed method showed state-of-the-art accuracy for a the classification task. Furthermore, we describe theoretical understandings about how PGD can mitigate dataset bias.

Gihun Lee, Minchan Jeong, Sangmook Kim et al.

Federated Learning (FL) aggregates locally trained models from individual clients to construct a global model. While FL enables learning a model with data privacy, it often suffers from significant performance degradation when clients have heterogeneous data distributions. This data heterogeneity causes the model to forget the global knowledge acquired from previously sampled clients after being trained on local datasets. Although the introduction of proximal objectives in local updates helps to preserve global knowledge, it can also hinder local learning by interfering with local objectives. To address this problem, we propose a novel method, Federated Stabilized Orthogonal Learning (FedSOL), which adopts an orthogonal learning strategy to balance the two conflicting objectives. FedSOL is designed to identify gradients of local objectives that are inherently orthogonal to directions affecting the proximal objective. Specifically, FedSOL targets parameter regions where learning on the local objective is minimally influenced by proximal weight perturbations. Our experiments demonstrate that FedSOL consistently achieves state-of-the-art performance across various scenarios.

Jaehoon Oh, Sungnyun Kim, Namgyu Ho et al.

Cross-domain few-shot learning (CD-FSL), where there are few target samples under extreme differences between source and target domains, has recently attracted huge attention. Recent studies on CD-FSL generally focus on transfer learning based approaches, where a neural network is pre-trained on popular labeled source domain datasets and then transferred to target domain data. Although the labeled datasets may provide suitable initial parameters for the target data, the domain difference between the source and target might hinder fine-tuning on the target domain. This paper proposes a simple yet powerful method that re-randomizes the parameters fitted on the source domain before adapting to the target data. The re-randomization resets source-specific parameters of the source pre-trained model and thus facilitates fine-tuning on the target domain, improving few-shot performance.

Taehyeon Kim, Shinhwan Kang, Hyeonjeong Shin et al.

The Weather4Cast competition (hosted by NeurIPS 2022) required competitors to predict super-resolution rain movies in various regions of Europe when low-resolution satellite contexts covering wider regions are given. In this paper, we show that a general baseline 3D U-Net can be significantly improved with region-conditioned layers as well as orthogonality regularizations on 1x1x1 convolutional layers. Additionally, we facilitate the generalization with a bag of training strategies: mixup data augmentation, self-distillation, and feature-wise linear modulation (FiLM). Presented modifications outperform the baseline algorithms (3D U-Net) by up to 19.54% with less than 1% additional parameters, which won the 4th place in the core test leaderboard.

Mingyu Kim, Kyeongryeol Go, Se-Young Yun

Neural processes (NPs) aim to stochastically complete unseen data points based on a given context dataset. NPs essentially leverage a given dataset as a context representation to derive a suitable identifier for a novel task. To improve the prediction accuracy, many variants of NPs have investigated context embedding approaches that generally design novel network architectures and aggregation functions satisfying permutation invariant. In this work, we propose a stochastic attention mechanism for NPs to capture appropriate context information. From the perspective of information theory, we demonstrate that the proposed method encourages context embedding to be differentiated from a target dataset, allowing NPs to consider features in a target dataset and context embedding independently. We observe that the proposed method can appropriately capture context embedding even under noisy data sets and restricted task distributions, where typical NPs suffer from a lack of context embeddings. We empirically show that our approach substantially outperforms conventional NPs in various domains through 1D regression, predator-prey model, and image completion. Moreover, the proposed method is also validated by MovieLens-10k dataset, a real-world problem.

Jaehyun Kwak, Nam Cao, Boryeong Cho et al.

Adversarial attacks against Large Vision-Language Models (LVLMs) are crucial for exposing safety vulnerabilities in modern multimodal systems. Recent attacks based on input transformations, such as random cropping, suggest that spatially localized perturbations can be more effective than global image manipulation. However, randomly cropping the entire image is inherently stochastic and fails to use the limited per-pixel perturbation budget efficiently. We make two key observations: (i) regional attention scores are positively correlated with adversarial loss sensitivity, and (ii) attacking high-attention regions induces a structured redistribution of attention toward subsequent salient regions. Based on these findings, we propose Stage-wise Attention-Guided Attack (SAGA), an attention-guided framework that progressively concentrates perturbations on high-attention regions. SAGA enables more efficient use of constrained perturbation budgets, producing highly imperceptible adversarial examples while consistently achieving state-of-the-art attack success rates across ten LVLMs. The source code is available at https://github.com/jackwaky/SAGA.

Sungnyun Kim, Sangmin Bae, Se-Young Yun

Deep learning in general domains has constantly been extended to domain-specific tasks requiring the recognition of fine-grained characteristics. However, real-world applications for fine-grained tasks suffer from two challenges: a high reliance on expert knowledge for annotation and necessity of a versatile model for various downstream tasks in a specific domain (e.g., prediction of categories, bounding boxes, or pixel-wise annotations). Fortunately, the recent self-supervised learning (SSL) is a promising approach to pretrain a model without annotations, serving as an effective initialization for any downstream tasks. Since SSL does not rely on the presence of annotation, in general, it utilizes the large-scale unlabeled dataset, referred to as an open-set. In this sense, we introduce a novel Open-Set Self-Supervised Learning problem under the assumption that a large-scale unlabeled open-set is available, as well as the fine-grained target dataset, during a pretraining phase. In our problem setup, it is crucial to consider the distribution mismatch between the open-set and target dataset. Hence, we propose SimCore algorithm to sample a coreset, the subset of an open-set that has a minimum distance to the target dataset in the latent space. We demonstrate that SimCore significantly improves representation learning performance through extensive experimental settings, including eleven fine-grained datasets and seven open-sets in various downstream tasks.

Kaito Ariu, Alexandre Proutiere, Se-Young Yun

In this paper, we investigate the problem of recovering hidden communities in the Labeled Stochastic Block Model (LSBM) with a finite number of clusters whose sizes grow linearly with the total number of nodes. We derive the necessary and sufficient conditions under which the expected number of misclassified nodes is less than $ s $, for any number $ s = o(n) $. To achieve this, we propose IAC (Instance-Adaptive Clustering), the first algorithm whose performance matches the instance-specific lower bounds both in expectation and with high probability. IAC is a novel two-phase algorithm that consists of a one-shot spectral clustering step followed by iterative likelihood-based cluster assignment improvements. This approach is based on the instance-specific lower bound and notably does not require any knowledge of the model parameters, including the number of clusters. By performing the spectral clustering only once, IAC maintains an overall computational complexity of $ \mathcal{O}(n\, \text{polylog}(n)) $, making it scalable and practical for large-scale problems.

Taehyeon Kim, Se-Young Yun

Efficient deployment of deep neural networks across many devices and resource constraints, particularly on edge devices, is one of the most challenging problems in the presence of data-privacy preservation issues. Conventional approaches have evolved to either improve a single global model while keeping each local heterogeneous training data decentralized (i.e. data heterogeneity; Federated Learning (FL)) or to train an overarching network that supports diverse architectural settings to address heterogeneous systems equipped with different computational capabilities (i.e. system heterogeneity; Neural Architecture Search). However, few studies have considered both directions simultaneously. This paper proposes the federation of supernet training (FedSup) framework to consider both scenarios simultaneously, i.e., where clients send and receive a supernet that contains all possible architectures sampled from itself. The approach is inspired by observing that averaging parameters during model aggregation for FL is similar to weight-sharing in supernet training. Thus, the proposed FedSup framework combines a weight-sharing approach widely used for training single shot models with FL averaging (FedAvg). Furthermore, we develop an efficient algorithm (E-FedSup) by sending the sub-model to clients on the broadcast stage to reduce communication costs and training overhead, including several strategies to enhance supernet training in the FL environment. We verify the proposed approach with extensive empirical evaluations. The resulting framework also ensures data and model heterogeneity robustness on several standard benchmarks.

Jihwan Oh, Joonkee Kim, Se-Young Yun

Distributional reinforcement learning demonstrates state-of-the-art performance in continuous and discrete control settings with the features of variance and risk, which can be used to explore. However, the exploration method employing the risk property is hard to find, although numerous exploration methods in Distributional RL employ the variance of return distribution per action. In this paper, we present risk scheduling approaches that explore risk levels and optimistic behaviors from a risk perspective. We demonstrate the performance enhancement of the DMIX algorithm using risk scheduling in a multi-agent setting with comprehensive experiments.

Taehyeon Kim, Joonkee Kim, Gihun Lee et al.

While instruction-tuned language models have demonstrated impressive zero-shot generalization, these models often struggle to generate accurate responses when faced with instructions that fall outside their training set. This paper presents Instructive Decoding (ID), a simple yet effective approach that augments the efficacy of instruction-tuned models. Specifically, ID adjusts the logits for next-token prediction in a contrastive manner, utilizing predictions generated from a manipulated version of the original instruction, referred to as a noisy instruction. This noisy instruction aims to elicit responses that could diverge from the intended instruction yet remain plausible. We conduct experiments across a spectrum of such noisy instructions, ranging from those that insert semantic noise via random words to others like 'opposite' that elicit the deviated responses. Our approach achieves considerable performance gains across various instruction-tuned models and tasks without necessitating any additional parameter updates. Notably, utilizing 'opposite' as the noisy instruction in ID, which exhibits the maximum divergence from the original instruction, consistently produces the most significant performance gains across multiple models and tasks.

SangMook Kim, Wonyoung Shin, Soohyuk Jang et al.

Robustness is becoming another important challenge of federated learning in that the data collection process in each client is naturally accompanied by noisy labels. However, it is far more complex and challenging owing to varying levels of data heterogeneity and noise over clients, which exacerbates the client-to-client performance discrepancy. In this work, we propose a robust federated learning method called FedRN, which exploits k-reliable neighbors with high data expertise or similarity. Our method helps mitigate the gap between low- and high-performance clients by training only with a selected set of clean examples, identified by their ensembled mixture models. We demonstrate the superiority of FedRN via extensive evaluations on three real-world or synthetic benchmark datasets. Compared with existing robust training methods, the results show that FedRN significantly improves the test accuracy in the presence of noisy labels.

Taehyeon Kim, Heesoo Myeong, Se-Young Yun

Knowledge Distillation (KD) has recently emerged as a popular method for compressing neural networks. In recent studies, generalized distillation methods that find parameters and architectures of student models at the same time have been proposed. Still, this search method requires a lot of computation to search for architectures and has the disadvantage of considering only convolutional blocks in their search space. This paper introduces a new algorithm, coined as Trust Region Aware architecture search to Distill knowledge Effectively (TRADE), that rapidly finds effective student architectures from several state-of-the-art architectures using trust region Bayesian optimization approach. Experimental results show our proposed TRADE algorithm consistently outperforms both the conventional NAS approach and pre-defined architecture under KD training.

Sumyeong Ahn, Sihyeon Kim, Jongwoo Ko et al.

The presence of noisy labels in a training dataset can significantly impact the performance of machine learning models. To tackle this issue, researchers have explored methods for Learning with Noisy Labels to identify clean samples and reduce the influence of noisy labels. However, constraining the influence of a certain portion of the training dataset can result in a reduction in overall generalization performance. To alleviate this, recent studies have considered the careful utilization of noisy labels by leveraging huge computational resources. Therefore, the increasing training cost necessitates a reevaluation of efficiency. In other areas of research, there has been a focus on developing fine-tuning techniques for large pre-trained models that aim to achieve both high generalization performance and efficiency. However, these methods have mainly concentrated on clean datasets, and there has been limited exploration of the noisy label scenario. In this research, our aim is to find an appropriate way to fine-tune pre-trained models for noisy labeled datasets. To achieve this goal, we investigate the characteristics of pre-trained models when they encounter noisy datasets. Through empirical analysis, we introduce a novel algorithm called TURN, which robustly and efficiently transfers the prior knowledge of pre-trained models. The algorithm consists of two main steps: (1) independently tuning the linear classifier to protect the feature extractor from being distorted by noisy labels, and (2) reducing the noisy label ratio and fine-tuning the entire model based on the noise-reduced dataset to adapt it to the target dataset. The proposed algorithm has been extensively tested and demonstrates efficient yet improved denoising performance on various benchmarks compared to previous methods.

Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim et al.

Chain-of-thought (CoT) reasoning has enabled large language models (LLMs) to utilize additional computation through intermediate tokens to solve complex tasks. However, we posit that typical reasoning traces contain many redundant tokens, incurring extraneous inference costs. Upon examination of the output distribution of current LLMs, we find evidence on their latent ability to reason more concisely, relative to their default behavior. To elicit this capability, we propose simple fine-tuning methods which leverage self-generated concise reasoning paths obtained by best-of-N sampling and few-shot conditioning, in task-specific settings. Our combined method achieves a 30% reduction in output tokens on average, across five model families on GSM8K and MATH, while maintaining average accuracy. By exploiting the fundamental stochasticity and in-context learning capabilities of LLMs, our self-training approach robustly elicits concise reasoning on a wide range of models, including those with extensive post-training. Code is available at https://github.com/TergelMunkhbat/concise-reasoning

Jihwan Oh, Joonkee Kim, Minchan Jeong et al.

The multi-agent setting is intricate and unpredictable since the behaviors of multiple agents influence one another. To address this environmental uncertainty, distributional reinforcement learning algorithms that incorporate uncertainty via distributional output have been integrated with multi-agent reinforcement learning (MARL) methods, achieving state-of-the-art performance. However, distributional MARL algorithms still rely on the traditional $ε$-greedy, which does not take cooperative strategy into account. In this paper, we present a risk-based exploration that leads to collaboratively optimistic behavior by shifting the sampling region of distribution. Initially, we take expectations from the upper quantiles of state-action values for exploration, which are optimistic actions, and gradually shift the sampling region of quantiles to the full distribution for exploitation. By ensuring that each agent is exposed to the same level of risk, we can force them to take cooperatively optimistic actions. Our method shows remarkable performance in multi-agent settings requiring cooperative exploration based on quantile regression appropriately controlling the level of risk.

Stephen Cha, Taehyeon Kim, Hayeon Lee et al.

This survey focuses on categorizing and evaluating the methods of supernet optimization in the field of Neural Architecture Search (NAS). Supernet optimization involves training a single, over-parameterized network that encompasses the search space of all possible network architectures. The survey analyses supernet optimization methods based on their approaches to spatial and temporal optimization. Spatial optimization relates to optimizing the architecture and parameters of the supernet and its subnets, while temporal optimization deals with improving the efficiency of selecting architectures from the supernet. The benefits, limitations, and potential applications of these methods in various tasks and settings, including transferability, domain generalization, and Transformer models, are also discussed.

Sungnyun Kim, Kangwook Jang, Sangmin Bae et al.

Audio-visual speech recognition (AVSR) aims to transcribe human speech using both audio and video modalities. In practical environments with noise-corrupted audio, the role of video information becomes crucial. However, prior works have primarily focused on enhancing audio features in AVSR, overlooking the importance of video features. In this study, we strengthen the video features by learning three temporal dynamics in video data: context order, playback direction, and the speed of video frames. Cross-modal attention modules are introduced to enrich video features with audio information so that speech variability can be taken into account when training on the video temporal dynamics. Based on our approach, we achieve the state-of-the-art performance on the LRS2 and LRS3 AVSR benchmarks for the noise-dominant settings. Our approach excels in scenarios especially for babble and speech noise, indicating the ability to distinguish the speech signal that should be recognized from lip movements in the video modality. We support the validity of our methodology by offering the ablation experiments for the temporal dynamics losses and the cross-modal attention architecture design.

Sihyeon Kim, Boryeong Cho, Sangmin Bae et al.

Despite the astonishing performance of recent Large Vision-Language Models (LVLMs), these models often generate inaccurate responses. To address this issue, previous studies have focused on mitigating hallucinations by employing contrastive decoding (CD) with augmented images, which amplifies the contrast with the original image. However, these methods have limitations, including reliance on a single augmentation, which is restrictive for certain tasks, as well as the high cost of using external knowledge. In this study, we address these limitations by exploring how to utilize multiple image augmentations. Through extensive experiments, we observed that different augmentations produce varying levels of contrast depending on the task. Based on this observation, we introduce a novel method called VACoDe, Visual Augmented Contrastive Decoding. This method adaptively selects the augmentation with the highest contrast for each task using the proposed softmax distance metric. Our empirical tests show that \alg outperforms previous methods and improves output quality in various vision-language tasks. Additionally, VACoDe can be universally applied across different model types and sizes without additional training or the use of external models and data.

Segyu Lee, Boryeong Cho, Hojung Jung et al.

Unified Multimodal Models (UMMs) offer powerful cross-modality capabilities but introduce new safety risks not observed in single-task models. Despite their emergence, existing safety benchmarks remain fragmented across tasks and modalities, limiting the comprehensive evaluation of complex system-level vulnerabilities. To address this gap, we introduce UniSAFE, the first comprehensive benchmark for system-level safety evaluation of UMMs across 7 I/O modality combinations, spanning conventional tasks and novel multimodal-context image generation settings. UniSAFE is built with a shared-target design that projects common risk scenarios across task-specific I/O configurations, enabling controlled cross-task comparisons of safety failures. Comprising 6,802 curated instances, we use UniSAFE to evaluate 15 state-of-the-art UMMs, both proprietary and open-source. Our results reveal critical vulnerabilities across current UMMs, including elevated safety violations in multi-image composition and multi-turn settings, with image-output tasks consistently more vulnerable than text-output tasks. These findings highlight the need for stronger system-level safety alignment for UMMs. Our code and data are publicly available at https://github.com/segyulee/UniSAFE

Jongwoo Ko, Seungjoon Park, Yujin Kim et al.

Structured pruning methods have proven effective in reducing the model size and accelerating inference speed in various network architectures such as Transformers. Despite the versatility of encoder-decoder models in numerous NLP tasks, the structured pruning methods on such models are relatively less explored compared to encoder-only models. In this study, we investigate the behavior of the structured pruning of the encoder-decoder models in the decoupled pruning perspective of the encoder and decoder component, respectively. Our findings highlight two insights: (1) the number of decoder layers is the dominant factor of inference speed, and (2) low sparsity in the pruned encoder network enhances generation quality. Motivated by these findings, we propose a simple and effective framework, NASH, that narrows the encoder and shortens the decoder networks of encoder-decoder models. Extensive experiments on diverse generation and inference tasks validate the effectiveness of our method in both speedup and output quality.

Felix den Breejen, Sangmin Bae, Stephen Cha et al.

While interests in tabular deep learning has significantly grown, conventional tree-based models still outperform deep learning methods. To narrow this performance gap, we explore the innovative retrieval mechanism, a methodology that allows neural networks to refer to other data points while making predictions. Our experiments reveal that retrieval-based training, especially when fine-tuning the pretrained TabPFN model, notably surpasses existing methods. Moreover, the extensive pretraining plays a crucial role to enhance the performance of the model. These insights imply that blending the retrieval mechanism with pretraining and transfer learning schemes offers considerable potential for advancing the field of tabular deep learning.

Seongha Eom, Namgyu Ho, Jaehoon Oh et al.

Contrastive language-image pre-training (CLIP) has demonstrated remarkable zero-shot classification ability, namely image classification using novel text labels. Existing works have attempted to enhance CLIP by fine-tuning on downstream tasks, but these have inadvertently led to performance degradation on unseen classes, thus harming zero-shot generalization. This paper aims to address this challenge by leveraging readily available image-text pairs from an external dataset for cross-modal guidance during inference. To this end, we propose X-MoRe, a novel inference method comprising two key steps: (1) cross-modal retrieval and (2) modal-confidence-based ensemble. Given a query image, we harness the power of CLIP's cross-modal representations to retrieve relevant textual information from an external image-text pair dataset. Then, we assign higher weights to the more reliable modality between the original query image and retrieved text, contributing to the final prediction. X-MoRe demonstrates robust performance across a diverse set of tasks without the need for additional training, showcasing the effectiveness of utilizing cross-modal features to maximize CLIP's zero-shot ability.

Sumyeong Ahn, Se-Young Yun

Improperly constructed datasets can result in inaccurate inferences. For instance, models trained on biased datasets perform poorly in terms of generalization (i.e., dataset bias). Recent debiasing techniques have successfully achieved generalization performance by underestimating easy-to-learn samples (i.e., bias-aligned samples) and highlighting difficult-to-learn samples (i.e., bias-conflicting samples). However, these techniques may fail owing to noisy labels, because the trained model recognizes noisy labels as difficult-to-learn and thus highlights them. In this study, we find that earlier approaches that used the provided labels to quantify difficulty could be affected by the small proportion of noisy labels. Furthermore, we find that running denoising algorithms before debiasing is ineffective because denoising algorithms reduce the impact of difficult-to-learn samples, including valuable bias-conflicting samples. Therefore, we propose an approach called denoising after entropy-based debiasing, i.e., DENEB, which has three main stages. (1) The prejudice model is trained by emphasizing (bias-aligned, clean) samples, which are selected using a Gaussian Mixture Model. (2) Using the per-sample entropy from the output of the prejudice model, the sampling probability of each sample that is proportional to the entropy is computed. (3) The final model is trained using existing denoising algorithms with the mini-batches constructed by following the computed sampling probability. Compared to existing debiasing and denoising algorithms, our method achieves better debiasing performance on multiple benchmarks.

Mingyu Kim, Jihwan Oh, Yongsik Lee et al.

In this paper, we propose a novel benchmark called the StarCraft Multi-Agent Challenges+, where agents learn to perform multi-stage tasks and to use environmental factors without precise reward functions. The previous challenges (SMAC) recognized as a standard benchmark of Multi-Agent Reinforcement Learning are mainly concerned with ensuring that all agents cooperatively eliminate approaching adversaries only through fine manipulation with obvious reward functions. This challenge, on the other hand, is interested in the exploration capability of MARL algorithms to efficiently learn implicit multi-stage tasks and environmental factors as well as micro-control. This study covers both offensive and defensive scenarios. In the offensive scenarios, agents must learn to first find opponents and then eliminate them. The defensive scenarios require agents to use topographic features. For example, agents need to position themselves behind protective structures to make it harder for enemies to attack. We investigate MARL algorithms under SMAC+ and observe that recent approaches work well in similar settings to the previous challenges, but misbehave in offensive scenarios. Additionally, we observe that an enhanced exploration approach has a positive effect on performance but is not able to completely solve all scenarios. This study proposes new directions for future research.

Seongyoon Kim, Gihun Lee, Jaehoon Oh et al.

Federated Learning (FL) is a collaborative method for training models while preserving data privacy in decentralized settings. However, FL encounters challenges related to data heterogeneity, which can result in performance degradation. In our study, we observe that as data heterogeneity increases, feature representation in the FedAVG model deteriorates more significantly compared to classifier weight. Additionally, we observe that as data heterogeneity increases, the gap between higher feature norms for observed classes, obtained from local models, and feature norms of unobserved classes widens, in contrast to the behavior of classifier weight norms. This widening gap extends to encompass the feature norm disparities between local and the global models. To address these issues, we introduce Federated Averaging with Feature Normalization Update (FedFN), a straightforward learning method. We demonstrate the superior performance of FedFN through extensive experiments, even when applied to pretrained ResNet18. Subsequently, we confirm the applicability of FedFN to foundation models.

Junghyun Lee, Sanghwa Kim, Yassir Jedra et al.

Evaluating large language models increasingly relies on LLM-as-a-judge protocols, but such evaluations remain costly: different judges have different prices and reliabilities, and the difficulty of each prompt-response pair can vary substantially. This raises a basic allocation question: under a fixed budget, how should one distribute evaluation queries across heterogeneous judges and instances to obtain the most accurate score estimates? We formalize this question as *budgeted heteroskedastic multi-judge estimation*. Given $K$ prompt-response pairs, $J$ judges with known costs, and unknown query-judge variances, the goal is to estimate a bounded score vector while minimizing an $\ell_p$-error. Our first contribution is to analyze the inverse-variance weighted estimator (IVWE) and to derive the oracle allocation that minimizes its error rate. Since this allocation depends on the unknown variances, we then address the practical unknown-variance setting by proposing EST-IVWE, an adaptive algorithm that constructs and leverages *optimistically biased* variance estimates to stabilize the empirical allocation. We prove that EST-IVWE matches the oracle IVWE rate up to lower-order terms in the budget. Our second and central theoretical contribution is a matching *local* minimax lower bound, which establishes the instance-optimality of the proposed algorithms. A key technical insight is that Fano-type high-probability arguments are too coarse for this problem: their packing construction loses the local variance structure that governs the optimal allocation. We instead use an Assouad-type in-expectation argument, based on local perturbations, which preserves this structure and yields the sharp allocation-dependent lower bound. Finally, we numerically validate the superiority of our approach over naïve uniform allocation on synthetic and HelpSteer2 datasets.

Junghyun Lee, Se-Young Yun, Kwang-Sung Jun

Logistic bandit is a ubiquitous framework of modeling users' choices, e.g., click vs. no click for advertisement recommender system. We observe that the prior works overlook or neglect dependencies in $S \geq \lVert θ_\star \rVert_2$, where $θ_\star \in \mathbb{R}^d$ is the unknown parameter vector, which is particularly problematic when $S$ is large, e.g., $S \geq d$. In this work, we improve the dependency on $S$ via a novel approach called {\it regret-to-confidence set conversion (R2CS)}, which allows us to construct a convex confidence set based on only the \textit{existence} of an online learning algorithm with a regret guarantee. Using R2CS, we obtain a strict improvement in the regret bound w.r.t. $S$ in logistic bandits while retaining computational feasibility and the dependence on other factors such as $d$ and $T$. We apply our new confidence set to the regret analyses of logistic bandits with a new martingale concentration step that circumvents an additional factor of $S$. We then extend this analysis to multinomial logistic bandits and obtain similar improvements in the regret, showing the efficacy of R2CS. While we applied R2CS to the (multinomial) logistic model, R2CS is a generic approach for developing confidence sets that can be used for various models, which can be of independent interest.

Junghyun Lee, Se-Young Yun, Kwang-Sung Jun

We present a unified likelihood ratio-based confidence sequence (CS) for any (self-concordant) generalized linear model (GLM) that is guaranteed to be convex and numerically tight. We show that this is on par or improves upon known CSs for various GLMs, including Gaussian, Bernoulli, and Poisson. In particular, for the first time, our CS for Bernoulli has a $\mathrm{poly}(S)$-free radius where $S$ is the norm of the unknown parameter. Our first technical novelty is its derivation, which utilizes a time-uniform PAC-Bayesian bound with a uniform prior/posterior, despite the latter being a rather unpopular choice for deriving CSs. As a direct application of our new CS, we propose a simple and natural optimistic algorithm called OFUGLB, applicable to any generalized linear bandits (GLB; Filippi et al. (2010)). Our analysis shows that the celebrated optimistic approach simultaneously attains state-of-the-art regrets for various self-concordant (not necessarily bounded) GLBs, and even $\mathrm{poly}(S)$-free for bounded GLBs, including logistic bandits. The regret analysis, our second technical novelty, follows from combining our new CS with a new proof technique that completely avoids the previously widely used self-concordant control lemma (Faury et al., 2020, Lemma 9). Numerically, OFUGLB outperforms or is at par with prior algorithms for logistic bandits.

Jung-hun Kim, Se-Young Yun

We study the adversarial bandit problem against arbitrary strategies, where the difficulty is captured by an unknown parameter $S$, which is the number of switches in the best arm in hindsight. To handle this problem, we adopt the master-base framework using the online mirror descent method (OMD). We first provide a master-base algorithm with simple OMD, achieving $\tilde{O}(S^{1/2}K^{1/3}T^{2/3})$, in which $T^{2/3}$ comes from the variance of loss estimators. To mitigate the impact of the variance, we propose using adaptive learning rates for OMD and achieve $\tilde{O}(\min\{\sqrt{SKTρ},S\sqrt{KT}\})$, where $ρ$ is a variance term for loss estimators.

Woojin Chung, Jiwoo Hong, Na Min An et al.

Stable pre-training is essential for achieving better-performing language models. However, tracking pre-training stability by calculating gradient variance at every step is impractical due to the significant computational costs. We explore Token Embedding Variability (TEV) as a simple and efficient proxy for assessing pre-training stability in language models with pre-layer normalization, given that shallower layers are more prone to gradient explosion (section 2.2). Moreover, we propose Multi-head Low-Rank Attention (MLRA) as an architecture to alleviate such instability by limiting the exponential growth of output embedding variance, thereby preventing the gradient explosion (section 3.2). Empirical results on GPT-2 with MLRA demonstrate increased stability and lower perplexity, particularly in deeper models.

Junghyun Lee, Hanseul Cho, Se-Young Yun et al.

Fair Principal Component Analysis (PCA) is a problem setting where we aim to perform PCA while making the resulting representation fair in that the projected distributions, conditional on the sensitive attributes, match one another. However, existing approaches to fair PCA have two main problems: theoretically, there has been no statistical foundation of fair PCA in terms of learnability; practically, limited memory prevents us from using existing approaches, as they explicitly rely on full access to the entire data. On the theoretical side, we rigorously formulate fair PCA using a new notion called \emph{probably approximately fair and optimal} (PAFO) learnability. On the practical side, motivated by recent advances in streaming algorithms for addressing memory limitation, we propose a new setting called \emph{fair streaming PCA} along with a memory-efficient algorithm, fair noisy power method (FNPM). We then provide its {\it statistical} guarantee in terms of PAFO-learnability, which is the first of its kind in fair PCA literature. Lastly, we verify the efficacy and memory efficiency of our algorithm on real-world datasets.

Junghyun Lee, Laura Schmid, Se-Young Yun

Multi-armed bandits are extensively used to model sequential decision-making, making them ubiquitous in many real-life applications such as online recommender systems and wireless networking. We consider a multi-agent setting where each agent solves their own bandit instance endowed with a different set of arms. Their goal is to minimize their group regret while collaborating via some communication protocol over a given network. Previous literature on this problem only considered arm heterogeneity and networked agents separately. In this work, we introduce a setting that encompasses both features. For this novel setting, we first provide a rigorous regret analysis for a standard flooding protocol combined with the classic UCB policy. Then, to mitigate the issue of high communication costs incurred by flooding in complex networks, we propose a new protocol called Flooding with Absorption (FwA). We provide a theoretical analysis of the resulting regret bound and discuss the advantages of using FwA over flooding. Lastly, we experimentally verify on various scenarios, including dynamic networks, that FwA leads to significantly lower communication costs despite minimal regret performance loss compared to other network protocols.

Jaehoon Oh, Se-Young Yun

Few-shot class-incremental learning (FSCIL) has addressed challenging real-world scenarios where unseen novel classes continually arrive with few samples. In these scenarios, it is required to develop a model that recognizes the novel classes without forgetting prior knowledge. In other words, FSCIL aims to maintain the base performance and improve the novel performance simultaneously. However, there is little study to investigate the two performances separately. In this paper, we first decompose the entire model into four types of parameters and demonstrate that the tendency of the two performances varies greatly with the updated parameters when the novel classes appear. Based on the analysis, we propose a simple method for FSCIL, coined as NoNPC, which uses normalized prototype classifiers without further training for incremental novel classes. It is shown that our straightforward method has comparable performance with the sophisticated state-of-the-art algorithms.

Jihwan Oh, Sungnyun Kim, Gahee Kim et al.

Offline multi-agent reinforcement learning (MARL) is increasingly recognized as crucial for effectively deploying RL algorithms in environments where real-time interaction is impractical, risky, or costly. In the offline setting, learning from a static dataset of past interactions allows for the development of robust and safe policies without the need for live data collection, which can be fraught with challenges. Building on this foundational importance, we present EAQ, Episodes Augmentation guided by Q-total loss, a novel approach for offline MARL framework utilizing diffusion models. EAQ integrates the Q-total function directly into the diffusion model as a guidance to maximize the global returns in an episode, eliminating the need for separate training. Our focus primarily lies on cooperative scenarios, where agents are required to act collectively towards achieving a shared goal-essentially, maximizing global returns. Consequently, we demonstrate that our episodes augmentation in a collaborative manner significantly boosts offline MARL algorithm compared to the original dataset, improving the normalized return by +17.3% and +12.9% for medium and poor behavioral policies in SMAC simulator, respectively.