arXiv subjects

Sungrae Park

Publications and source records attributed to Sungrae Park.

At least 19 recordsLinked to original sources

arXiv2026

Solar Open 2 Technical Report

We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.

cs.CL↗

arXiv2026

Bounded Hyperbolic Tangent: A Stable and Efficient Alternative to Pre-Layer Normalization in Large Language Models

Pre-Layer Normalization (Pre-LN) is the de facto choice for large language models (LLMs) and is crucial for stable pretraining and effective transfer learning. However, Pre-LN incurs repeated statistical-computation overhead and remains vulnerable to the curse of depth, where hidden-state magnitudes and variances grow as the number of layers increases, destabilizing training. Efficiency-oriented normalization-free methods such as Dynamic Tanh (DyT) improve throughput but remain fragile at depth. To jointly address stability and efficiency, we propose Bounded Hyperbolic Tanh (BHyT), a drop-in replacement for Pre-LN. BHyT combines a tanh nonlinearity with explicit, data-driven input bounding to keep activations within a non-saturating range. It prevents depth-wise growth in activation magnitude and variance and provides a theoretical stability guarantee. For efficiency, BHyT computes exact statistics once per block and replaces a second normalization with a lightweight variance approximation. Empirically, BHyT demonstrates improved stability and efficiency during pretraining, achieving an average of 1.6\% faster training and an average of 1.77\% higher token generation throughput compared to RMSNorm, while maintaining strong pretraining-only and post-SFT performance across language understanding and reasoning benchmarks\footnote{Code is available at: https://github.com/MLAI-Yonsei/BHyT}.

cs.CL↗

arXiv2026

Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) plays a pivotal role in improving the reasoning ability of large language models. However, widely used PPO surrogate objectives are fundamentally local, as they rely on a local approximation of the exact policy gradient objective. While this approximation improves stability by reducing the variance induced by importance sampling, it also introduces structural bias into the surrogate objective, which must be controlled through trust region mechanisms. In this work, we introduce the $N$-step forward trace, which augments the PPO surrogate objective using the cumulative likelihood ratio of the next $N-1$ tokens. Building on this idea, we propose $N$-Step Forward-Trace Policy Optimization (NFPO), a practical RLVR algorithm that integrates the $N$-step forward trace into the masked policy gradient framework. NFPO provides a continuous bridge between the PPO surrogate objective and the exact policy gradient objective, offering a principled mechanism for controlling the bias-variance trade-off. Our theoretical analysis shows that, with an appropriate choice of $N$, the proposed objective yields a tighter policy-improvement bound than the standard PPO surrogate. Experiments on comprehensive reasoning benchmarks demonstrate that NFPO consistently improves performance, supporting our theoretical findings.

cs.LG↗

arXiv2026

MIDUS: Memory-Infused Depth Up-Scaling

Expanding pre-trained language models offers a practical way to increase capacity without training larger models from scratch. Depth Up-Scaling (DUS) does so by duplicating Transformer blocks and inserting them into a pre-trained backbone. This process also duplicates FFN-heavy blocks, increasing parameter and compute cost while adding capacity through a block-level dense residual branch. Yet prior work suggests that added capacity need not remain tied to dense FFN branches, while attention heads often play heterogeneous roles, motivating more efficient head-level residual corrections. We propose Memory-Infused Depth Up-Scaling (MIDUS), which replaces the duplicated FFN branches with memory layers and turns added depth into lightweight retrieval-based residual capacity. We introduce a Head-wise Memory Layer (HML), which combines multi-head product-key memory with Head-wise Implicit Value Expansion (HIVE). HML assigns each head a distinct key space, while HIVE realizes head-specific values from a shared latent bank through compact projections. Alongside empirical improvements in performance and efficiency, our head-importance and fixed-retrieval structural analyses characterize HML with HIVE as a structurally distinct, head-conditioned alternative to FFN-based residual expansion.

cs.LG↗

arXiv2026

User-Oriented Multi-Turn Dialogue Generation with Tool Use at scale

The recent paradigm shift toward large reasoning models (LRMs) as autonomous agents has intensified the demand for sophisticated, multi-turn tool-use capabilities. Yet, existing datasets and data-generation approaches are limited by static, predefined toolsets that cannot scale to the complexity of open-ended human-agent collaboration. To address this, we initially developed a framework for automated task-oriented multi-turn dialogue generation at scale, utilizing an LRM-based simulator to dynamically generate high-value, domain-specific tools to solve specified tasks. However, we observe that a purely task-oriented design often results in "solely task-solving" trajectories, where the agent completes the objective with minimal interaction, failing to generate the high turn-count conversations seen in realistic scenarios. To bridge this gap, we shift toward a user-oriented simulation paradigm. By decoupling task generation from a dedicated user simulator that mimics human behavioral rules - such as incremental request-making and turn-by-turn feedback - we facilitate more authentic, extended multi-turn dialogues that reflect the iterative nature of real-world problem solving. Our generation pipeline operates as a versatile, plug-and-play module capable of initiating generation from any state, ensuring high scalability in producing extended tool-use data. Furthermore, by facilitating multiple task completions within a single trajectory, it yields a high-density dataset that reflects the multifaceted demands of real-world human-agent interaction.

cs.CL↗

arXiv2026

Solar Open Technical Report

We introduce Solar Open, a 102B-parameter bilingual Mixture-of-Experts language model for underserved languages. Solar Open demonstrates a systematic methodology for building competitive LLMs by addressing three interconnected challenges. First, to train effectively despite data scarcity for underserved languages, we synthesize 4.5T tokens of high-quality, domain-specific, and RL-oriented data. Second, we coordinate this data through a progressive curriculum jointly optimizing composition, quality thresholds, and domain coverage across 20 trillion tokens. Third, to enable reasoning capabilities through scalable RL, we apply our proposed framework SnapPO for efficient optimization. Across benchmarks in English and Korean, Solar Open achieves competitive performance, demonstrating the effectiveness of this methodology for underserved language AI development.

cs.CL↗

arXiv2025

ZERA: Zero-init Instruction Evolving Refinement Agent -- From Zero Instructions to Structured Prompts via Principle-based Optimization

Automatic Prompt Optimization (APO) improves large language model (LLM) performance by refining prompts for specific tasks. However, prior APO methods typically focus only on user prompts, rely on unstructured feedback, and require large sample sizes and long iteration cycles-making them costly and brittle. We propose ZERA (Zero-init Instruction Evolving Refinement Agent), a novel framework that jointly optimizes both system and user prompts through principled, low-overhead refinement. ZERA scores prompts using eight generalizable criteria with automatically inferred weights, and revises prompts based on these structured critiques. This enables fast convergence to high-quality prompts using minimal examples and short iteration cycles. We evaluate ZERA across five LLMs and nine diverse datasets spanning reasoning, summarization, and code generation tasks. Experimental results demonstrate consistent improvements over strong baselines. Further ablation studies highlight the contribution of each component to more effective prompt construction. Our implementation including all prompts is publicly available at https://github.com/younatics/zera-agent.

cs.CL↗

arXiv2025

KIEval: Evaluation Metric for Document Key Information Extraction

Document Key Information Extraction (KIE) is a technology that transforms valuable information in document images into structured data, and it has become an essential function in industrial settings. However, current evaluation metrics of this technology do not accurately reflect the critical attributes of its industrial applications. In this paper, we present KIEval, a novel application-centric evaluation metric for Document KIE models. Unlike prior metrics, KIEval assesses Document KIE models not just on the extraction of individual information (entity) but also of the structured information (grouping). Evaluation of structured information provides assessment of Document KIE models that are more reflective of extracting grouped information from documents in industrial settings. Designed with industrial application in mind, we believe that KIEval can become a standard evaluation metric for developing or applying Document KIE models in practice. The code will be publicly available.

cs.CL↗

arXiv2023

SRFormer: Text Detection Transformer with Incorporated Segmentation and Regression

Existing techniques for text detection can be broadly classified into two primary groups: segmentation-based and regression-based methods. Segmentation models offer enhanced robustness to font variations but require intricate post-processing, leading to high computational overhead. Regression-based methods undertake instance-aware prediction but face limitations in robustness and data efficiency due to their reliance on high-level representations. In our academic pursuit, we propose SRFormer, a unified DETR-based model with amalgamated Segmentation and Regression, aiming at the synergistic harnessing of the inherent robustness in segmentation representations, along with the straightforward post-processing of instance-level regression. Our empirical analysis indicates that favorable segmentation predictions can be obtained at the initial decoder layers. In light of this, we constrain the incorporation of segmentation branches to the first few decoder layers and employ progressive regression refinement in subsequent layers, achieving performance gains while minimizing computational load from the mask.Furthermore, we propose a Mask-informed Query Enhancement module. We take the segmentation result as a natural soft-ROI to pool and extract robust pixel representations, which are then employed to enhance and diversify instance queries. Extensive experimentation across multiple benchmarks has yielded compelling findings, highlighting our method's exceptional robustness, superior training and data efficiency, as well as its state-of-the-art performance. Our code is available at https://github.com/retsuh-bqw/SRFormer-Text-Det.

cs.CV↗

arXiv2022

Multi-modal Text Recognition Networks: Interactive Enhancements between Visual and Semantic Features

Linguistic knowledge has brought great benefits to scene text recognition by providing semantics to refine character sequences. However, since linguistic knowledge has been applied individually on the output sequence, previous methods have not fully utilized the semantics to understand visual clues for text recognition. This paper introduces a novel method, called Multi-modAl Text Recognition Network (MATRN), that enables interactions between visual and semantic features for better recognition performances. Specifically, MATRN identifies visual and semantic feature pairs and encodes spatial information into semantic features. Based on the spatial encoding, visual and semantic features are enhanced by referring to related features in the other modality. Furthermore, MATRN stimulates combining semantic features into visual features by hiding visual clues related to the character in the training phase. Our experiments demonstrate that MATRN achieves state-of-the-art performances on seven benchmarks with large margins, while naive combinations of two modalities show less-effective improvements. Further ablative studies prove the effectiveness of our proposed components. Our implementation is available at https://github.com/wp03052/MATRN.

cs.CV↗

arXiv2022

Domain Generalization by Mutual-Information Regularization with Pre-trained Models

Domain generalization (DG) aims to learn a generalized model to an unseen target domain using only limited source domains. Previous attempts to DG fail to learn domain-invariant representations only from the source domains due to the significant domain shifts between training and test domains. Instead, we re-formulate the DG objective using mutual information with the oracle model, a model generalized to any possible domain. We derive a tractable variational lower bound via approximating the oracle model by a pre-trained model, called Mutual Information Regularization with Oracle (MIRO). Our extensive experiments show that MIRO significantly improves the out-of-distribution performance. Furthermore, our scaling experiments show that the larger the scale of the pre-trained model, the greater the performance improvement of MIRO. Source code is available at https://github.com/kakaobrain/miro.

cs.LG↗

arXiv2022

RewriteNet: Reliable Scene Text Editing with Implicit Decomposition of Text Contents and Styles

Scene text editing (STE), which converts a text in a scene image into the desired text while preserving an original style, is a challenging task due to a complex intervention between text and style. In this paper, we propose a novel STE model, referred to as RewriteNet, that decomposes text images into content and style features and re-writes a text in the original image. Specifically, RewriteNet implicitly distinguishes the content from the style by introducing scene text recognition. Additionally, independent of the exact supervisions with synthetic examples, we propose a self-supervised training scheme for unlabeled real-world images, which bridges the domain gap between synthetic and real data. Our experiments present that RewriteNet achieves better generation performances than other comparisons. Further analysis proves the feature decomposition of RewriteNet and demonstrates the reliability and robustness through diverse experiments. Our implementation is publicly available at \url{https://github.com/clovaai/rewritenet}

cs.CV↗

arXiv2022

BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents

Key information extraction (KIE) from document images requires understanding the contextual and spatial semantics of texts in two-dimensional (2D) space. Many recent studies try to solve the task by developing pre-trained language models focusing on combining visual features from document images with texts and their layout. On the other hand, this paper tackles the problem by going back to the basic: effective combination of text and layout. Specifically, we propose a pre-trained language model, named BROS (BERT Relying On Spatiality), that encodes relative positions of texts in 2D space and learns from unlabeled documents with area-masking strategy. With this optimized training scheme for understanding texts in 2D space, BROS shows comparable or better performance compared to previous methods on four KIE benchmarks (FUNSD, SROIE*, CORD, and SciTSR) without relying on visual features. This paper also reveals two real-world challenges in KIE tasks-(1) minimizing the error from incorrect text ordering and (2) efficient learning from fewer downstream examples-and demonstrates the superiority of BROS over previous methods. Code is available at https://github.com/clovaai/bros.

cs.CL↗

arXiv2021

SWAD: Domain Generalization by Seeking Flat Minima

Domain generalization (DG) methods aim to achieve generalizability to an unseen target domain by using only training data from the source domains. Although a variety of DG methods have been proposed, a recent study shows that under a fair evaluation protocol, called DomainBed, the simple empirical risk minimization (ERM) approach works comparable to or even outperforms previous methods. Unfortunately, simply solving ERM on a complex, non-convex loss function can easily lead to sub-optimal generalizability by seeking sharp minima. In this paper, we theoretically show that finding flat minima results in a smaller domain generalization gap. We also propose a simple yet effective method, named Stochastic Weight Averaging Densely (SWAD), to find flat minima. SWAD finds flatter minima and suffers less from overfitting than does the vanilla SWA by a dense and overfit-aware stochastic weight sampling strategy. SWAD shows state-of-the-art performances on five DG benchmarks, namely PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet, with consistent and large margins of +1.6% averagely on out-of-domain accuracy. We also compare SWAD with conventional generalization methods, such as data augmentation and consistency regularization methods, to verify that the remarkable performance improvements are originated from by seeking flat minima, not from better in-domain generalizability. Last but not least, SWAD is readily adaptable to existing DG methods without modification; the combination of SWAD and an existing DG method further improves DG performances. Source code is available at https://github.com/khanrc/swad.

cs.LG↗

arXiv2021

SynthTIGER: Synthetic Text Image GEneratoR Towards Better Text Recognition Models

For successful scene text recognition (STR) models, synthetic text image generators have alleviated the lack of annotated text images from the real world. Specifically, they generate multiple text images with diverse backgrounds, font styles, and text shapes and enable STR models to learn visual patterns that might not be accessible from manually annotated data. In this paper, we introduce a new synthetic text image generator, SynthTIGER, by analyzing techniques used for text image synthesis and integrating effective ones under a single algorithm. Moreover, we propose two techniques that alleviate the long-tail problem in length and character distributions of training data. In our experiments, SynthTIGER achieves better STR performance than the combination of synthetic datasets, MJSynth (MJ) and SynthText (ST). Our ablation study demonstrates the benefits of using sub-components of SynthTIGER and the guideline on generating synthetic text images for STR models. Our implementation is publicly available at https://github.com/clovaai/synthtiger.

cs.CV↗

arXiv2021

Show, Attend and Distill:Knowledge Distillation via Attention-based Feature Matching

Knowledge distillation extracts general knowledge from a pre-trained teacher network and provides guidance to a target student network. Most studies manually tie intermediate features of the teacher and student, and transfer knowledge through pre-defined links. However, manual selection often constructs ineffective links that limit the improvement from the distillation. There has been an attempt to address the problem, but it is still challenging to identify effective links under practical scenarios. In this paper, we introduce an effective and efficient feature distillation method utilizing all the feature levels of the teacher without manually selecting the links. Specifically, our method utilizes an attention-based meta-network that learns relative similarities between features, and applies identified similarities to control distillation intensities of all possible pairs. As a result, our method determines competent links more efficiently than the previous approach and provides better performance on model compression and transfer learning tasks. Further qualitative analyses and ablative studies describe how our method contributes to better distillation. The implementation code is available at github.com/clovaai/attention-feature-distillation.

cs.LG↗

arXiv2020

Character Region Attention For Text Spotting

A scene text spotter is composed of text detection and recognition modules. Many studies have been conducted to unify these modules into an end-to-end trainable model to achieve better performance. A typical architecture places detection and recognition modules into separate branches, and a RoI pooling is commonly used to let the branches share a visual feature. However, there still exists a chance of establishing a more complimentary connection between the modules when adopting recognizer that uses attention-based decoder and detector that represents spatial information of the character regions. This is possible since the two modules share a common sub-task which is to find the location of the character regions. Based on the insight, we construct a tightly coupled single pipeline model. This architecture is formed by utilizing detection outputs in the recognizer and propagating the recognition loss through the detection stage. The use of character score map helps the recognizer attend better to the character center points, and the recognition loss propagation to the detector module enhances the localization of the character regions. Also, a strengthened sharing stage allows feature rectification and boundary localization of arbitrary-shaped text regions. Extensive experiments demonstrate state-of-the-art performance in publicly available straight and curved benchmark dataset.

cs.CV↗

arXiv2020

CLEval: Character-Level Evaluation for Text Detection and Recognition Tasks

Despite the recent success of text detection and recognition methods, existing evaluation metrics fail to provide a fair and reliable comparison among those methods. In addition, there exists no end-to-end evaluation metric that takes characteristics of OCR tasks into account. Previous end-to-end metric contains cascaded errors from the binary scoring process applied in both detection and recognition tasks. Ignoring partially correct results raises a gap between quantitative and qualitative analysis, and prevents fine-grained assessment. Based on the fact that character is a key element of text, we hereby propose a Character-Level Evaluation metric (CLEval). In CLEval, the \textit{instance matching} process handles split and merge detection cases, and the \textit{scoring process} conducts character-level evaluation. By aggregating character-level scores, the CLEval metric provides a fine-grained evaluation of end-to-end results composed of the detection and recognition as well as individual evaluations for each module from the end-performance perspective. We believe that our metrics can play a key role in developing and analyzing state-of-the-art text detection and recognition methods. The evaluation code is publicly available at https://github.com/clovaai/CLEval.

cs.CV↗