Sungjun Cho

Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min et al.

We show that standard Transformers without graph-specific modifications can lead to promising results in graph learning both in theory and practice. Given a graph, we simply treat all nodes and edges as independent tokens, augment them with token embeddings, and feed them to a Transformer. With an appropriate choice of token embeddings, we prove that this approach is theoretically at least as expressive as an invariant graph network (2-IGN) composed of equivariant linear layers, which is already more expressive than all message-passing Graph Neural Networks (GNN). When trained on a large-scale graph dataset (PCQM4Mv2), our method coined Tokenized Graph Transformer (TokenGT) achieves significantly better results compared to GNN baselines and competitive results compared to Transformer variants with sophisticated graph-specific inductive bias. Our implementation is available at https://github.com/jw9730/tokengt.

Sungjun Cho, Seonwoo Min, Jinwoo Kim et al.

To overcome the quadratic cost of self-attention, recent works have proposed various sparse attention modules, most of which fall under one of two groups: 1) sparse attention under a hand-crafted patterns and 2) full attention followed by a sparse variant of softmax such as $α$-entmax. Unfortunately, the first group lacks adaptability to data while the second still requires quadratic cost in training. In this work, we propose SBM-Transformer, a model that resolves both problems by endowing each attention head with a mixed-membership Stochastic Block Model (SBM). Then, each attention head data-adaptively samples a bipartite graph, the adjacency of which is used as an attention mask for each input. During backpropagation, a straight-through estimator is used to flow gradients beyond the discrete sampling step and adjust the probabilities of sampled edges based on the predictive loss. The forward and backward cost are thus linear to the number of edges, which each attention head can also choose flexibly based on the input. By assessing the distribution of graphs, we theoretically show that SBM-Transformer is a universal approximator for arbitrary sequence-to-sequence functions in expectation. Empirical evaluations under the LRA and GLUE benchmarks demonstrate that our model outperforms previous efficient variants as well as the original Transformer with full attention. Our implementation can be found in https://github.com/sc782/SBM-Transformer .

Jinwoo Kim, Saeyoon Oh, Sungjun Cho et al.

Many problems in computer vision and machine learning can be cast as learning on hypergraphs that represent higher-order relations. Recent approaches for hypergraph learning extend graph neural networks based on message passing, which is simple yet fundamentally limited in modeling long-range dependencies and expressive power. On the other hand, tensor-based equivariant neural networks enjoy maximal expressiveness, but their application has been limited in hypergraphs due to heavy computation and strict assumptions on fixed-order hyperedges. We resolve these problems and present Equivariant Hypergraph Neural Network (EHNN), the first attempt to realize maximally expressive equivariant layers for general hypergraph learning. We also present two practical realizations of our framework based on hypernetworks (EHNN-MLP) and self-attention (EHNN-Transformer), which are easy to implement and theoretically more expressive than most message passing approaches. We demonstrate their capability in a range of hypergraph learning problems, including synthetic k-edge identification, semi-supervised classification, and visual keypoint matching, and report improved performances over strong message passing baselines. Our implementation is available at https://github.com/jw9730/ehnn.

Sungmin Cha, Sungjun Cho, Dasol Hwang et al.

Since the recent advent of regulations for data protection (e.g., the General Data Protection Regulation), there has been increasing demand in deleting information learned from sensitive data in pre-trained models without retraining from scratch. The inherent vulnerability of neural networks towards adversarial attacks and unfairness also calls for a robust method to remove or correct information in an instance-wise fashion, while retaining the predictive performance across remaining data. To this end, we consider instance-wise unlearning, of which the goal is to delete information on a set of instances from a pre-trained model, by either misclassifying each instance away from its original prediction or relabeling the instance to a different label. We also propose two methods that reduce forgetting on the remaining data: 1) utilizing adversarial examples to overcome forgetting at the representation-level and 2) leveraging weight importance metrics to pinpoint network parameters guilty of propagating unwanted information. Both methods only require the pre-trained model and data instances to forget, allowing painless application to real-life settings where the entire training set is unavailable. Through extensive experimentation on various image classification benchmarks, we show that our approach effectively preserves knowledge of remaining data while unlearning given instances in both single-task and continual unlearning scenarios.

Sungmin Cha, Sungjun Cho, Dasol Hwang et al.

Large Language Models (LLMs) have demonstrated strong reasoning and memorization capabilities via pretraining on massive textual corpora. However, this poses risk of privacy and copyright violations, highlighting the need for efficient machine unlearning methods that remove sensitive data without retraining from scratch. While Gradient Ascent (GA) is commonly used to unlearn by reducing the likelihood of generating unwanted content, it leads to unstable optimization and catastrophic forgetting of retrained knowledge. We find that combining GA with low-rank adaptation results in poor trade-offs between computational cost and generative performance. To address these challenges, we propose Low-rank Knowledge Unlearning (LoKU), a novel framework that enables robust and efficient unlearning for LLMs. First, we introduce Inverted Hinge Loss, which suppresses unwanted tokens while maintaining fluency by boosting the probability of the next most likely token. Second, we develop a data-adaptive initialization for LoRA adapters via low-rank approximation weighted with relative Fisher information, thereby focusing updates on parameters critical for removing targeted knowledge. Experiments on the Training Data Extraction Challenge dataset using GPT-Neo models as well as on the TOFU benchmark with Phi-1.5B and Llama2-7B models demonstrate that our approach effectively removes sensitive information while maintaining reasoning and generative capabilities with minimal impact. Our implementation can be found in https://github.com/csm9493/efficient-llm-unlearning.

Sungjun Cho, Dae-Woong Jeong, Sung Moon Ko et al.

Pretraining molecular representations from large unlabeled data is essential for molecular property prediction due to the high cost of obtaining ground-truth labels. While there exist various 2D graph-based molecular pretraining approaches, these methods struggle to show statistically significant gains in predictive performance. Recent work have thus instead proposed 3D conformer-based pretraining under the task of denoising, which led to promising results. During downstream finetuning, however, models trained with 3D conformers require accurate atom-coordinates of previously unseen molecules, which are computationally expensive to acquire at scale. In light of this limitation, we propose D&D, a self-supervised molecular representation learning framework that pretrains a 2D graph encoder by distilling representations from a 3D denoiser. With denoising followed by cross-modal knowledge distillation, our approach enjoys use of knowledge obtained from denoising as well as painless application to downstream tasks with no access to accurate conformers. Experiments on real-world molecular property prediction datasets show that the graph encoder trained via D&D can infer 3D information based on the 2D graph and shows superior performance and label-efficiency against other baselines.

Seungyeon Rhyu, Kichang Yang, Sungjun Cho et al.

Music generation introduces challenging complexities to large language models. Symbolic structures of music often include vertical harmonization as well as horizontal counterpoint, urging various adaptations and enhancements for large-scale Transformers. However, existing works share three major drawbacks: 1) their tokenization requires domain-specific annotations, such as bars and beats, that are typically missing in raw MIDI data; 2) the pure impact of enhancing token embedding methods is hardly examined without domain-specific annotations; and 3) existing works to overcome the aforementioned drawbacks, such as MuseNet, lack reproducibility. To tackle such limitations, we develop a MIDI-based music generation framework inspired by MuseNet, empirically studying two structural embeddings that do not rely on domain-specific annotations. We provide various metrics and insights that can guide suitable encoding to deploy. We also verify that multiple embedding configurations can selectively boost certain musical aspects. By providing open-source implementations via HuggingFace, our findings shed light on leveraging large language models toward practical and reproducible music generation.

SungJun Cho, Chetan Gohil, Rukuang Huang et al.

Recent success in natural language processing has motivated growing interest in large-scale foundation models for neuroimaging data. Such models often require discretization of continuous neural time series data, a process referred to as 'tokenization'. However, the impact of different tokenization strategies for neural data is currently poorly understood. In this work, we present a systematic evaluation of sample-level tokenization strategies for transformer-based large neuroimaging models (LNMs) applied to magnetoencephalography (MEG) data. We compare learnable and non-learnable tokenizers by examining their signal reconstruction fidelity and their impact on subsequent foundation modeling performance (token prediction, biological plausibility of generated data, preservation of subject-specific information, and performance on downstream tasks). For the learnable tokenizer, we introduce a novel approach based on an autoencoder. Experiments were conducted on three publicly available MEG datasets spanning different acquisition sites, scanners, and experimental paradigms. Our results show that both learnable and non-learnable discretization schemes achieve high reconstruction accuracy and broadly comparable performance across most evaluation criteria, suggesting that simple fixed sample-level tokenization strategies can be used in the development of neural foundation models. The code is available at https://github.com/OHBA-analysis/Cho2026_Tokenizer.

Sungjun Cho, Seunghyuk Cho, Sungwoo Park et al.

Real-world graphs naturally exhibit hierarchical or cyclical structures that are unfit for the typical Euclidean space. While there exist graph neural networks that leverage hyperbolic or spherical spaces to learn representations that embed such structures more accurately, these methods are confined under the message-passing paradigm, making the models vulnerable against side-effects such as oversmoothing and oversquashing. More recent work have proposed global attention-based graph Transformers that can easily model long-range interactions, but their extensions towards non-Euclidean geometry are yet unexplored. To bridge this gap, we propose Fully Product-Stereographic Transformer, a generalization of Transformers towards operating entirely on the product of constant curvature spaces. When combined with tokenized graph Transformers, our model can learn the curvature appropriate for the input graph in an end-to-end fashion, without the need of additional tuning on different curvature initializations. We also provide a kernelized approach to non-Euclidean attention, which enables our model to run in time and memory cost linear to the number of nodes and edges while respecting the underlying geometry. Experiments on graph reconstruction and node classification demonstrate the benefits of generalizing Transformers to the non-Euclidean domain.

Sung Moon Ko, Sungjun Cho, Dae-Woong Jeong et al.

Graph pooling is a crucial operation for encoding hierarchical structures within graphs. Most existing graph pooling approaches formulate the problem as a node clustering task which effectively captures the graph topology. Conventional methods ask users to specify an appropriate number of clusters as a hyperparameter, then assume that all input graphs share the same number of clusters. In inductive settings where the number of clusters can vary, however, the model should be able to represent this variation in its pooling layers in order to learn suitable clusters. Thus we propose GMPool, a novel differentiable graph pooling architecture that automatically determines the appropriate number of clusters based on the input data. The main intuition involves a grouping matrix defined as a quadratic form of the pooling operator, which induces use of binary classification probabilities of pairwise combinations of nodes. GMPool obtains the pooling operator by first computing the grouping matrix, then decomposing it. Extensive evaluations on molecular property prediction tasks demonstrate that our method outperforms conventional methods.

Nicholas Roberts, Sungjun Cho, Zhiqi Gao et al.

Modern LLMs scale at test-time, e.g. via repeated sampling, where inference cost grows with model size and the number of samples. This creates a trade-off that pretraining scaling laws, such as Chinchilla, do not address. We present Train-to-Test ($T^2$) scaling laws that jointly optimize model size, training tokens, and number of inference samples under fixed end-to-end budgets. $T^2$ modernizes pretraining scaling laws with pass@$k$ modeling used for test-time scaling, then jointly optimizes pretraining and test-time decisions. Forecasts from $T^2$ are robust over distinct modeling approaches: measuring joint scaling effect on the task loss and modeling impact on task accuracy. Across eight downstream tasks, we find that when accounting for inference cost, optimal pretraining decisions shift radically into the overtraining regime, well-outside of the range of standard pretraining scaling suites. We validate our results by pretraining heavily overtrained models in the optimal region that $T^2$ scaling forecasts, confirming their substantially stronger performance compared to pretraining scaling alone. Finally, as frontier LLMs are post-trained, we show that our findings survive the post-training stage, making $T^2$ scaling meaningful in modern deployments.

HyunJae Lee, Gihyeon Lee, Junhwan Kim et al.

Despite the evolution of Convolutional Neural Networks (CNNs), their performance is surprisingly dependent on the choice of hyperparameters. However, it remains challenging to efficiently explore large hyperparameter search space due to the long training times of modern CNNs. Multi-fidelity optimization enables the exploration of more hyperparameter configurations given budget by early termination of unpromising configurations. However, it often results in selecting a sub-optimal configuration as training with the high-performing configuration typically converges slowly in an early phase. In this paper, we propose Multi-fidelity Optimization with a Recurring Learning rate (MORL) which incorporates CNNs' optimization process into multi-fidelity optimization. MORL alleviates the problem of slow-starter and achieves a more precise low-fidelity approximation. Our comprehensive experiments on general image classification, transfer learning, and semi-supervised learning demonstrate the effectiveness of MORL over other multi-fidelity optimization methods such as Successive Halving Algorithm (SHA) and Hyperband. Furthermore, it achieves significant performance improvements over hand-tuned hyperparameter configuration within a practical budget.

Jaehoon Lee, Hankook Lee, Sungik Choi et al.

When solving forecasting problems including multiple time-series features, existing approaches often fall into two extreme categories, depending on whether to utilize inter-feature information: univariate and complete-multivariate models. Unlike univariate cases which ignore the information, complete-multivariate models compute relationships among a complete set of features. However, despite the potential advantage of leveraging the additional information, complete-multivariate models sometimes underperform univariate ones. Therefore, our research aims to explore a middle ground between these two by introducing what we term Partial-Multivariate models where a neural network captures only partial relationships, that is, dependencies within subsets of all features. To this end, we propose PMformer, a Transformer-based partial-multivariate model, with its training algorithm. We demonstrate that PMformer outperforms various univariate and complete-multivariate models, providing a theoretical rationale and empirical analysis for its superiority. Additionally, by proposing an inference technique for PMformer, the forecasting accuracy is further enhanced. Finally, we highlight other advantages of PMformer: efficiency and robustness under missing features.

Minhyuk Seo, Hyunseo Koh, Wonje Jeung et al.

Online continual learning suffers from an underfitted solution due to insufficient training for prompt model update (e.g., single-epoch training). To address the challenge, we propose an efficient online continual learning method using the neural collapse phenomenon. In particular, we induce neural collapse to form a simplex equiangular tight frame (ETF) structure in the representation space so that the continuously learned model with a single epoch can better fit to the streamed data by proposing preparatory data training and residual correction in the representation space. With an extensive set of empirical validations using CIFAR-10/100, TinyImageNet, ImageNet-200, and ImageNet-1K, we show that our proposed method outperforms state-of-the-art methods by a noticeable margin in various online continual learning scenarios such as disjoint and Gaussian scheduled continuous (i.e., boundary-free) data setups.

Sungjun Cho

Hybrid-capture novel view synthesis combines images at substantially different camera distances (e.g., aerial drone and ground-level views). Standard 3D Gaussian Splatting (3DGS), trained for 30K iterations with one rendered view per optimizer step, under-fits the minority regime by 1-3 dB on five hybrid-capture benchmarks. We isolate the lever that closes this gap. Among compute-matched alternatives -- vanilla 60K iterations, magnitude corrections (GradNorm), direction-aware near/far gradient surgery, projective preconditioning, confidence-gated sample-level surgery, and a random two-view-per-step control -- the simplest structural change wins: rendering two views per optimizer step. The pairing rule (geometry-defined near/far, random, or active loss-disparity) does not change PSNR beyond seed variance on any of the five scenes; the structural change of having two views per step does. We propose a variance-decomposition framework that predicts and explains this finding: under bimodal camera regimes, between-regime gradient variance turns out to be small relative to within-regime variance in 3DGS, so structured and random pairings are variance-equivalent in expectation, and the variance halving from two-view accumulation itself is the dominant effect. We verify the framework on five scenes whose camera-altitude bimodality coefficients span [0.55, 1.00], and we report the negative result that direction-aware projection, magnitude correction, confidence gating, and an active loss-disparity pairing all fall within seed variance of random two-view pairing. The two-view structural lever transfers cleanly to the Scaffold-GS and Pixel-GS backbones. We position this work as an honest characterization of which training-side axes do and do not move PSNR for hybrid-capture 3DGS, together with the framework that explains why.

Changho Shin, Xinya Yan, Suenggwan Jo et al.

Language models often struggle with temporal misalignment, performance degradation caused by shifts in the temporal distribution of data. Continuously updating models to avoid degradation is expensive. Can models be adapted without updating model weights? We present TARDIS, an unsupervised representation editing method that addresses this challenge. TARDIS extracts steering vectors from unlabeled data and adjusts the model's representations to better align with the target time period's distribution. Our experiments reveal that TARDIS enhances downstream task performance without the need for fine-tuning, can mitigate temporal misalignment even when exact target time period data is unavailable, and remains efficient even when the temporal information of the target data points is unknown at inference time.

Sungjun Cho

Diabetic retinopathy (DR) is a leading cause of preventable blindness, and automated fundus image grading can play an important role in large-scale screening. In this work, we investigate three CLIP-based approaches for five-class DR severity grading: (1) a zero-shot baseline using prompt engineering, (2) a hybrid FCN-CLIP model augmented with CBAM attention, and (3) a ranking-aware prompting model that encodes the ordinal structure of DR progression. We train and evaluate on a combined dataset of APTOS 2019 and Messidor-2 (n=5,406), addressing class imbalance through resampling and class-specific optimal thresholding. Our experiments show that the ranking-aware model achieves the highest overall accuracy (93.42%, AUROC 0.9845) and strong recall on clinically critical severe cases, while the hybrid FCN-CLIP model (92.49%, AUROC 0.99) excels at detecting proliferative DR. Both substantially outperform the zero-shot baseline (55.17%, AUROC 0.75). We analyze the complementary strengths of each approach and discuss their practical implications for screening contexts.

Sungjun Cho, Changho Shin, Suenggwan Jo et al.

Forecasting in the real world requires integrating structured time-series data with unstructured textual information, but existing methods are architecturally limited by fixed input/output horizons and are unable to model or quantify uncertainty. We address this challenge by introducing LLM-integrated Bayesian State space models (LBS), a novel probabilistic framework for multimodal temporal forecasting. At a high level, LBS consists of two components: (1) a state space model (SSM) backbone that captures the temporal dynamics of latent states from which both numerical and textual observations are generated and (2) a pretrained large language model (LLM) that is adapted to encode textual inputs for posterior state estimation and decode textual forecasts consistent with the latent trajectory. This design enables flexible lookback and forecast windows, principled uncertainty quantification, and improved temporal generalization thanks to the well-suited inductive bias of SSMs toward modeling dynamical systems. Experiments on the TextTimeCorpus benchmark demonstrate that LBS improves the previous state-of-the-art by 13.20% while providing human-readable summaries of each forecast. Our work is the first to unify LLMs and SSMs for joint numerical and textual prediction, offering a novel foundation for multimodal temporal reasoning.

Rukuang Huang, Sungjun Cho, Chetan Gohil et al.

Modelling the complex spatiotemporal patterns of large-scale brain dynamics is crucial for neuroscience, but traditional methods fail to capture the rich structure in modalities such as magnetoencephalography (MEG). Recent advances in deep learning have enabled significant progress in other domains, such as language and vision, by using foundation models at scale. Here, we introduce MEG-GPT, a transformer based foundation model that uses time-attention and next time-point prediction. To facilitate this, we also introduce a novel data-driven tokeniser for continuous MEG data, which preserves the high temporal resolution of continuous MEG signals without lossy transformations. We trained MEG-GPT on tokenised brain region time-courses extracted from a large-scale MEG dataset (N=612, eyes-closed rest, Cam-CAN data), and show that the learnt model can generate data with realistic spatio-spectral properties, including transient events and population variability. Critically, it performs well in downstream decoding tasks, improving downstream supervised prediction task, showing improved zero-shot generalisation across sessions (improving accuracy from 0.54 to 0.59) and subjects (improving accuracy from 0.41 to 0.49) compared to a baseline methods. Furthermore, we show the model can be efficiently fine-tuned on a smaller labelled dataset to boost performance in cross-subject decoding scenarios. This work establishes a powerful foundation model for electrophysiological data, paving the way for applications in computational neuroscience and neural decoding.

Sungjun Cho, Dasol Hwang, Frederic Sala et al.

Current unlearning metrics for generative models evaluate success based on reference responses or classifier outputs rather than assessing the core objective: whether the unlearned model behaves indistinguishably from a model that never saw the unwanted data. This reference-specific approach creates systematic blind spots, allowing models to appear successful while retaining unwanted knowledge accessible through alternative prompts or attacks. We address these limitations by proposing Functional Alignment for Distributional Equivalence (FADE), a novel metric that measures distributional similarity between unlearned and reference models by comparing bidirectional likelihood assignments over generated samples. Unlike existing approaches that rely on predetermined references, FADE captures functional alignment across the entire output distribution, providing a principled assessment of genuine unlearning. Our experiments on the TOFU benchmark for LLM unlearning and the UnlearnCanvas benchmark for text-to-image diffusion model unlearning reveal that methods achieving near-optimal scores on traditional metrics fail to achieve distributional equivalence, with many becoming more distant from the gold standard than before unlearning. These findings expose fundamental gaps in current evaluation practices and demonstrate that FADE provides a more robust foundation for developing and assessing truly effective unlearning methods.

Sungmin Cha, Sungjun Cho, Dasol Hwang et al.

Batch Normalization (BN) and its variants has been extensively studied for neural nets in various computer vision tasks, but relatively little work has been dedicated to studying the effect of BN in continual learning. To that end, we develop a new update patch for BN, particularly tailored for the exemplar-based class-incremental learning (CIL). The main issue of BN in CIL is the imbalance of training data between current and past tasks in a mini-batch, which makes the empirical mean and variance as well as the learnable affine transformation parameters of BN heavily biased toward the current task -- contributing to the forgetting of past tasks. While one of the recent BN variants has been developed for "online" CIL, in which the training is done with a single epoch, we show that their method does not necessarily bring gains for "offline" CIL, in which a model is trained with multiple epochs on the imbalanced training data. The main reason for the ineffectiveness of their method lies in not fully addressing the data imbalance issue, especially in computing the gradients for learning the affine transformation parameters of BN. Accordingly, our new hyperparameter-free variant, dubbed as Task-Balanced BN (TBBN), is proposed to more correctly resolve the imbalance issue by making a horizontally-concatenated task-balanced batch using both reshape and repeat operations during training. Based on our experiments on class incremental learning of CIFAR-100, ImageNet-100, and five dissimilar task datasets, we demonstrate that our TBBN, which works exactly the same as the vanilla BN in the inference time, is easily applicable to most existing exemplar-based offline CIL algorithms and consistently outperforms other BN variants.

Moontae Lee, Sungjun Cho, Kun Dong et al.

Across many data domains, co-occurrence statistics about the joint appearance of objects are powerfully informative. By transforming unsupervised learning problems into decompositions of co-occurrence statistics, spectral algorithms provide transparent and efficient algorithms for posterior inference such as latent topic analysis and community detection. As object vocabularies grow, however, it becomes rapidly more expensive to store and run inference algorithms on co-occurrence statistics. Rectifying co-occurrence, the key process to uphold model assumptions, becomes increasingly more vital in the presence of rare terms, but current techniques cannot scale to large vocabularies. We propose novel methods that simultaneously compress and rectify co-occurrence statistics, scaling gracefully with the size of vocabulary and the dimension of latent space. We also present new algorithms learning latent variables from the compressed statistics, and verify that our methods perform comparably to previous approaches on both textual and non-textual data.