Frontier model evaluations are shifting from foundational capabilities (e.g., instruction following and reasoning) toward compositional, agentic ones, but Korean agentic benchmarks remain scarce. We introduce K-BrowseComp, a web-browsing agent benchmark grounded in Korean contexts, consisting of 400 problems. The 300-problem K-BrowseComp-Verified subset is manually constructed and validated by native Korean speakers. On this subset, frontier LLMs, including GPT-5.5, DeepSeek-V4-Pro, and GLM-5.1, reach only 30.00–45.67%, a substantial drop from BrowseComp, while Korean LLMs released through Korea’s Proprietary AI Foundation Model program achieve at most 10.33%. We further construct a 100-problem Synthetic split using hard few-shot exemplars and failure-mode-targeted generation to exploit the asymmetry between solving and creating web browsing problems. On the adversarially filtered synthetic diagnostic split, the strongest model reaches only 26.00%, and we report this split separately as a targeted stress test. We publicly release our data and code at https://github.com/prometheus-eval/K-BrowseComp.

Math reasoning benchmarks have prolifer- ated, yet most lack a per-item difficulty sig- nal grounded in actual human performance. We introduce KCSAT-ML, a decade (2014– 2025) of Korean College Scholastic Ability Test (KCSAT; Suneung) mathematics: 664 problems with a 339-item core set carrying official per-item error rates from nationwide cohorts of hundreds of thousands of examinees. We pair the benchmark with Difficulty-aligned Reasoning Gain (DRG): a score-orthogonal metric that asks whether a model’s mistakes concentrate on the items humans found hard, or on items humans found easy. Together they expose, across a wide range of VLMs (and LLMs with OCR), three patterns: (i) low- budget accuracy collapses on the high-human- error tail at every model size; (ii) test-time scaling (TTS) raises token use roughly linearly with cohort error rate, while accuracy gains follow a non-monotonic curve; (iii) within a single family, TTS flips between anti-scaling on the hardest items and overthinking on eas- ier ones — two faces of the same alignment failure. With the proposed DRG metric, we find that models with near-identical accuracy can sit at near-opposite values: one model gets wrong what humans also find hard, while another solves the hardest items yet fails on items humans find easy, which is a critical con- trast that aggregate accuracy hides. Our code and dataset builder is fully open-sourced at https://github.com/naver-ai/KCSAT-ML.

Production deployments of Multimodal Large Language Models (MLLMs) increasingly rely on system messages to govern model behavior. Yet existing MLLM benchmarks either evaluate constraints in text only or embed them into the user turn, leaving system-message adherence in multimodal contexts largely unmeasured; they also leave open whether compliance comes at the cost of foundational vision-language capabilities. We introduce **VSysBench**, a simple yet effective MLLM benchmark that organizes system-message constraints into 5 main categories and 22 sub-categories, ranging from textual directives in visual contexts to fully vision-grounded ones, each paired with a misaligned counterpart that stress-tests the instructional hierarchy. **VSysBench** scores each response jointly along two axes, constraint compliance and answer correctness, via the Joint Satisfaction Rate (JSR) and Cross-Constraint Sensitivity (CCS). Across 16 MLLMs, we find that imposing system messages substantially erodes base task accuracy, that compliance collapses under user conflict for open-weight models while remaining stable for top proprietary ones, and that vision-grounded constraints are the hardest category for nearly every model.

Fine-tuning a pretrained LLM into a vision-language model (VLM) can erode the backbone’s text capability, with the damage concentrated on tasks that require following exact output rules, such as instruction following, chain-of-thought reasoning graded on a strictly parsed final answer, and similar evaluations with strict graders. We trace this gap to *attention-sink corruption*: VL fine-tuning perturbs the early sink position that anchors a large fraction of attention probability, and how well the base LLM preserves its sink tracks how much of the affected capability survives adaptation. Building on this view, we introduce *Sink Strength*, a single scalar computed on the base LLM in a few seconds on a single GPU that predicts post-VL degradation without any VL training. It consistently tracks relative degradation across the six VLM–LLM pairs and multiple format-sensitive tasks. Complementing this diagnostic, we find that post-pretraining QK-norm injection fails to reproduce the protection of native QK-norm, while several off-the-shelf weight-merging settings fail to recover the lost capability after VL training. These negative results underscore the value of screening backbones with Sink Strength before VL training and narrow the intervention space toward head-selective training-time protection.

Temporal grounding—returning the interval [t_s, t_e] for a natural-language query over a video—is the language interface to long-form video, yet has been mainly studied on non-hour-scale videos; the dynamics of hour-scale natural-language grounding remain underexplored. We take the position that at hour-scale, the binding constraint is search, not recognition: Video-LLMs are bottlenecked not by localizing a nearby event, but by searching for the relevant region of a long video. To test this, we release ExtremeWhenBench (2,273 queries over 194 videos, mean 75.7 min, max 9 hr). It combines three properties that no prior benchmark has together: hour-scale mean video length, open-form explicit natural-language queries, and open access that works with standard Video-LLM toolkits. Every open Video-LLM collapses while a frame-level retrieval baseline outperforms them; a failure taxonomy attributes 88% of failures to search; and a retrieve-then-ground hybrid recovers 6.7× over the monolithic Video-LLM—mirroring retrieve-then-read in open-domain QA.

Building an omni-modal foundation model means evaluating it across text, image, video, and audio. Excellent evaluation toolkits exist for each modality, but their inference engines, prompt conventions, and metric implementations are mutually incompatible, so practitioners end up maintaining separate environments for every toolchain and still struggle to compare results across them. OmniEvaluator grew out of this need in our own model development: rather than reimplementing benchmarks, it connects existing inference engines and curated evaluation libraries at a higher level, exposing four inference backends, four evaluation frameworks, and over a thousand benchmarks through a single interface. Every run is recorded as an artifact capturing the full configuration for exact reproduction, and results flow into a shared dashboard for cross-model comparison. A federated mode shares GPU inference servers across concurrent evaluations, and a built-in verifier, small enough to run on CPU, keeps its score stable across engines and prompts where rule-based scoring fluctuates under configuration mismatch, matching cost-efficient commercial LLM judges without their recurring API cost. The system, demo video, and dashboard are publicly available. (https://github.com/naver-ai/omni-evaluator, https://www.youtube.com/watch?v=4Z5VZZWyXqY, https://omni-evaluator.info)

Instruction tuning is a widely used approach to improve the instruction-following ability of large language models (LLMs). Instruction-tuning datasets typically include a mixture of context-augmented and context-free examples, yet prior work has largely combined these data types without examining their distinct effects. In this paper, we investigate how training LLMs with or without context affects model behavior and downstream performance. First, in the text domain, we show that LLMs trained with context attend more strongly to the provided knowledge, achieving better grounding. We also observe that context-augmented training shifts how LLMs use knowledge: models store and leverage less on parametric knowledge and instead depend more on the provided context. Second, we observe that using LLM trained with context-augmented data as the backbone for vision-language models reduces hallucination and improves grounding in the visual domain. Finally, we explore practical strategies for real-world deployments where context availability varies. We show that maintaining separate context-augmented and context-free models and routing inputs between them yields more robust overall performance than training a single mixed model, as it better preserves their complementary strengths.

This paper presents the Korean National Educational Test Benchmark (KoNET), a new benchmark designed to evaluate Multimodal Generative AI Systems using Korean national educational tests. KoNET comprises four exams: the Korean Elementary General Educational Development Test (KoEGED), Middle (KoMGED), High (KoHGED), and College Scholastic Ability Test (KoCSAT). These exams are renowned for their rigorous standards and diverse questions, facilitating a comprehensive analysis of AI performance across different educational levels. By focusing on Korean, KoNET provides insights into model performance in less-explored languages. We assess a range of models—open-source, open-access, and closed APIs—by examining difficulties, subject diversity, and human error rates. The code and dataset builder will be made fully open-source.

This paper introduces MMRefine, a MultiModal Refinement benchmark designed to evaluate the error refinement capabilities of Multimodal Large Language Models (MLLMs). As the emphasis shifts toward enhancing reasoning during inference, MMRefine provides a framework that evaluates MLLMs’ abilities to detect and correct errors across six distinct scenarios beyond just comparing final accuracy before and after refinement. Furthermore, the benchmark analyzes the refinement performance by categorizing errors into six error types.Experiments with various open and closed MLLMs reveal bottlenecks and factors impeding refinement performance, highlighting areas for improvement in effective reasoning enhancement. Our code and dataset are publicly available at https://github.com/naver-ai/MMRefine.

Assessing long-form responses generated by Vision-Language Models (VLMs) is challenging. It not only requires checking whether the VLM follows the given instruction but also verifying whether the text output is properly grounded on the given image. Inspired by the recent approach of evaluating LMs with LMs, in this work, we propose to evaluate VLMs with VLMs. For this purpose, we present a new feedback dataset called the Perception Collection, encompassing 15K customized score rubrics that users might care about during assessment. Using the Perception Collection, we train Prometheus-Vision, the first open-source VLM evaluator model that can understand the user-defined score criteria during evaluation. Prometheus-Vision shows the highest Pearson correlation with human evaluators and GPT-4V among open-source models, showing its effectiveness for transparent and accessible evaluation of VLMs. We open-source our code, dataset, and model.

Recent advancements in language and vision assistants have showcased impressive capabilities but suffer from a lack of transparency, limiting broader research and reproducibility. While open-source models handle general image tasks effectively, they face challenges with the high computational demands of complex visually-situated text understanding. Such tasks often require increased token inputs and large vision modules to harness high-resolution information. Striking a balance between model size and data importance remains an open question. This study aims to redefine the design of vision-language models by identifying key components and creating efficient models with constrained inference costs. By strategically formulating datasets, optimizing vision modules, and enhancing supervision techniques, we achieve significant improvements in inference throughput while maintaining high performance. Extensive experiments across models ranging from 160M to 13B parameters offer insights into model optimization.We will fully open-source our codebase, models, and datasets at https://github.com/naver-ai/elva.

Recent advances in Large Language Models (LLMs) have stimulated a surge of research aimed at extending their applications to the visual domain. While these models exhibit promise in generating abstract image captions and facilitating natural conversations, their performance on text-rich images still requires improvement. In this paper, we introduce Contrastive Reading Model (Cream), a novel neural architecture designed to enhance the language-image understanding capability of LLMs by capturing intricate details that are often overlooked in existing methods. Cream combines vision and auxiliary encoders, fortified by a contrastive feature alignment technique, to achieve a more effective comprehension of language information in visually situated contexts within the images. Our approach bridges the gap between vision and language understanding, paving the way for the development of more sophisticated Document Intelligence Assistants. Through rigorous evaluations across diverse visually-situated language understanding tasks that demand reasoning capabilities, we demonstrate the compelling performance of Cream, positioning it as a prominent model in the field of visual document understanding. We provide our codebase and newly-generated datasets at https://github.com/naver-ai/cream.

A real-world information extraction (IE) system for semi-structured document images often involves a long pipeline of multiple modules, whose complexity dramatically increases its development and maintenance cost. One can instead consider an end-to-end model that directly maps the input to the target output and simplify the entire process. However, such generation approach is known to lead to unstable performance if not designed carefully. Here we present our recent effort on transitioning from our existing pipeline-based IE system to an end-to-end system focusing on practical challenges that are associated with replacing and deploying the system in real, large-scale production. By carefully formulating document IE as a sequence generation task, we show that a single end-to-end IE system can be built and still achieve competent performance.

This paper introduces a method that efficiently reduces the computational cost and parameter size of Transformer. The proposed model, refer to as Group-Transformer, splits feature space into multiple groups, factorizes the calculation paths, and reduces computations for the group interaction. Extensive experiments on two benchmark tasks, enwik8 and text8, prove our model’s effectiveness and efficiency in small-scale Transformers. To the best of our knowledge, Group-Transformer is the first attempt to design Transformer with the group strategy, widely used for efficient CNN architectures.

We propose a new type of representation learning method that models words, phrases and sentences seamlessly. Our method does not depend on word segmentation and any human-annotated resources (e.g., word dictionaries), yet it is very effective for noisy corpora written in unsegmented languages such as Chinese and Japanese. The main idea of our method is to ignore word boundaries completely (i.e., segmentation-free), and construct representations for all character n-grams in a raw corpus with embeddings of compositional sub-n-grams. Although the idea is simple, our experiments on various benchmarks and real-world datasets show the efficacy of our proposal.

We propose a new word embedding method called word-like character n-gram embedding, which learns distributed representations of words by embedding word-like character n-grams. Our method is an extension of recently proposed segmentation-free word embedding, which directly embeds frequent character n-grams from a raw corpus. However, its n-gram vocabulary tends to contain too many non-word n-grams. We solved this problem by introducing an idea of expected word frequency. Compared to the previously proposed methods, our method can embed more words, along with the words that are not included in a given basic word dictionary. Since our method does not rely on word segmentation with rich word dictionaries, it is especially effective when the text in the corpus is in unsegmented language and contains many neologisms and informal words (e.g., Chinese SNS dataset). Our experimental results on Sina Weibo (a Chinese microblog service) and Twitter show that the proposed method can embed more words and improve the performance of downstream tasks.