Chanwoo Park

Chanwoo Park, Sangdoo Yun, Sanghyuk Chun

We propose the first unified theoretical analysis of mixed sample data augmentation (MSDA), such as Mixup and CutMix. Our theoretical results show that regardless of the choice of the mixing strategy, MSDA behaves as a pixel-level regularization of the underlying training loss and a regularization of the first layer parameters. Similarly, our theoretical results support that the MSDA training strategy can improve adversarial robustness and generalization compared to the vanilla training strategy. Using the theoretical results, we provide a high-level understanding of how different design choices of MSDA work differently. For example, we show that the most popular MSDA methods, Mixup and CutMix, behave differently, e.g., CutMix regularizes the input gradients by pixel distances, while Mixup regularizes the input gradients regardless of pixel distances. Our theoretical results also show that the optimal MSDA strategy depends on tasks, datasets, or model parameters. From these observations, we propose generalized MSDAs, a Hybrid version of Mixup and CutMix (HMix) and Gaussian Mixup (GMix), simple extensions of Mixup and CutMix. Our implementation can leverage the advantages of Mixup and CutMix, while our implementation is very efficient, and the computation cost is almost neglectable as Mixup and CutMix. Our empirical study shows that our HMix and GMix outperform the previous state-of-the-art MSDA methods in CIFAR-100 and ImageNet classification tasks. Source code is available at https://github.com/naver-ai/hmix-gmix

Alexander K. Saeri, Jess Graham, Michael Noetel et al.

Artificial intelligence poses many risks, ranging from familiar present-day harms to unprecedented and potentially catastrophic ones. Effective risk management requires prioritization: we must understand which risks are most severe, who is most vulnerable, and who is most responsible for addressing them. We report results from a three-round Delphi study conducted late 2025 with 272 international AI experts. Experts rated 24 AI risks on harm probability and severity, sector and actor vulnerability, actor responsibility, and overall concern. Experts estimated the five most severe harms in the next 5 years were likely to come from dangerous capabilities, competitive dynamics, weapons & cyberattacks (including CBRNE), power centralization, and false information. In a business-as-usual scenario, experts judged 18 of 24 risks as having a more than 10% probability of catastrophic outcomes (e.g., more than 1 million deaths or more than USD 100B in financial loss) in the next 5 years (2025-2030). In a scenario where pragmatic mitigations are implemented, experts still judged five risks as having a more than 10% probability of catastrophic outcomes: dangerous capabilities, weapons & cyberattacks, environmental harm, inequality & unemployment, and power centralization. All 24 risks were judged as being more than 5% likely to cause catastrophic outcomes. AI users and the general public were judged the most vulnerable to these risks, but experts assigned the highest responsibility for addressing them to general-purpose AI developers and governance actors (including governments, regulators, and standards bodies). Across most risks, experts identified information, finance, and national security as the most vulnerable sectors. These findings can guide AI risk prioritization and clarify expert expectations about who should bear responsibility for mitigation.

Chanwoo Park, Kaiqing Zhang, Asuman Ozdaglar

We study a new class of Markov games, \emph(multi-player) zero-sum Markov Games} with \emph{Networked separable interactions} (zero-sum NMGs), to model the local interaction structure in non-cooperative multi-agent sequential decision-making. We define a zero-sum NMG as a model where {the payoffs of the auxiliary games associated with each state are zero-sum and} have some separable (i.e., polymatrix) structure across the neighbors over some interaction network. We first identify the necessary and sufficient conditions under which an MG can be presented as a zero-sum NMG, and show that the set of Markov coarse correlated equilibrium (CCE) collapses to the set of Markov Nash equilibrium (NE) in these games, in that the product of per-state marginalization of the former for all players yields the latter. Furthermore, we show that finding approximate Markov \emph{stationary} CCE in infinite-horizon discounted zero-sum NMGs is \texttt{PPAD}-hard, unless the underlying network has a ``star topology''. Then, we propose fictitious-play-type dynamics, the classical learning dynamics in normal-form games, for zero-sum NMGs, and establish convergence guarantees to Markov stationary NE under a star-shaped network structure. Finally, in light of the hardness result, we focus on computing a Markov \emph{non-stationary} NE and provide finite-iteration guarantees for a series of value-iteration-based algorithms. We also provide numerical experiments to corroborate our theoretical results.

Seungju Han, Konwoo Kim, Chanwoo Park et al. · stanford

Synthetic data augmentation helps language models learn new knowledge in data-constrained domains. However, naively scaling existing synthetic data methods by training on more synthetic tokens or using stronger generators yields diminishing returns below the performance of RAG. To break the RAG ceiling, we introduce Synthetic Mixed Training, which combines synthetic QAs and synthetic documents. This leverages their complementary training signals, and enables log-linear improvements as both synthetic data volume and generator strength increase. This allows the model to outperform RAG by a 2.6\% relative gain on QuaLITY, a long-document reading comprehension benchmark. In addition, we introduce Focal Rewriting, a simple technique for synthetic document generation that explicitly conditions document generation on specific questions, improving the diversity of synthetic documents and yielding a steeper log-linear scaling curve. On QuaLITY, our final recipe trains a Llama 8B model that outperforms RAG by 4.4\% relatively. Across models and benchmarks (QuaLITY, LongHealth, FinanceBench), our training enables models to beat RAG in five of six settings, outperforms by 2.6\%, and achieves a 9.1\% gain when combined with RAG.

Gagik Magakyan, Amirhossein Reisizadeh, Chanwoo Park et al.

Adaptability has been regarded as a central feature in the foundation models, enabling them to effectively acclimate to unseen downstream tasks. Parameter-efficient fine-tuning methods such as celebrated LoRA facilitate efficient adaptation of large foundation models using labeled, high-quality and generally scarce task data. To mitigate data scarcity in fine-tuning of foundation models, we propose to leverage task similarity across multiple downstream users. Intuitively, users with similar tasks must be able to assist each other in boosting the effective fine-tuning data size. We propose Collaborative Low-Rank Adaptation, or CoLoRA, which exploits task similarity to collaboratively and efficiently fine-tune personalized foundation models. The main idea in CoLoRA is to train one shared adapter capturing underlying task similarities across all tasks, and personalized adapters tailored to user-specific tasks. We theoretically study CoLoRA on heterogeneous linear regression and provide provable guarantees for ground truth recovery. We also conduct several natural language experiments with varying task similarity, which further demonstrate that when trained together with similar tasks, individual performances are significantly boosted.

Chanwoo Park, Ziyang Chen, Asuman Ozdaglar et al.

Large language models (LLMs) are increasingly deployed as "agents" for decision-making (DM) in interactive and dynamic environments. Yet, since they were not originally designed for DM, recent studies show that LLMs can struggle even in basic online DM problems, failing to achieve low regret or an effective exploration-exploitation tradeoff. To address this, we introduce Iterative Regret-Minimization Fine-Tuning (Iterative RMFT), a post-training procedure that repeatedly distills low-regret decision trajectories back into the base model. At each iteration, the model rolls out multiple decision trajectories, selects the k-lowest regret ones, and fine-tunes itself on them. Unlike prior methods that (a) distill action sequences from known DM algorithms or (b) rely on manually crafted chain-of-thought templates, our approach leverages the regret metric to elicit the model's own DM ability and reasoning rationales. This reliance on model-generated reasoning avoids rigid output engineering and provides more flexible, natural-language training signals. Empirical results show that Iterative RMFT improves LLMs' DM performance across diverse models - from Transformers with numerical input/output, to open-weight LLMs, and advanced closed-weight models like GPT-4o mini. Its flexibility in output and reasoning formats enables generalization across tasks with varying horizons, action spaces, reward processes, and natural-language contexts. Finally, we provide theoretical insight showing that a single-layer Transformer under this paradigm can act as a no-regret learner in a simplified setting. Overall, Iterative RMFT offers a principled and general post-training framework for enhancing LLMs' decision-making capabilities.

Yubin Kim, Chanwoo Park, Hyewon Jeong et al. · mit

Foundation models are becoming valuable tools in medicine. Yet despite their promise, the best way to leverage Large Language Models (LLMs) in complex medical tasks remains an open question. We introduce a novel multi-agent framework, named Medical Decision-making Agents (MDAgents) that helps address this gap by automatically assigning a collaboration structure to a team of LLMs. The assigned solo or group collaboration structure is tailored to the medical task at hand, emulating real-world medical decision-making processes adapted to tasks of varying complexities. We evaluate our framework and baseline methods using state-of-the-art LLMs across a suite of real-world medical knowledge and medical diagnosis benchmarks, including a comparison of LLMs' medical complexity classification against human physicians. MDAgents achieved the best performance in seven out of ten benchmarks on tasks requiring an understanding of medical knowledge and multi-modal reasoning, showing a significant improvement of up to 4.2% (p < 0.05) compared to previous methods' best performances. Ablation studies reveal that MDAgents effectively determines medical complexity to optimize for efficiency and accuracy across diverse medical tasks. Notably, the combination of moderator review and external medical knowledge in group collaboration resulted in an average accuracy improvement of 11.8%. Our code can be found at https://github.com/mitmedialab/MDAgents.

Hyunin Lee, Chanwoo Park, David Abel et al.

Black swan events are statistically rare occurrences that carry extremely high risks. A typical view of defining black swan events is heavily assumed to originate from an unpredictable time-varying environments; however, the community lacks a comprehensive definition of black swan events. To this end, this paper challenges that the standard view is incomplete and claims that high-risk, statistically rare events can also occur in unchanging environments due to human misperception of their value and likelihood, which we call as spatial black swan event. We first carefully categorize black swan events, focusing on spatial black swan events, and mathematically formalize the definition of black swan events. We hope these definitions can pave the way for the development of algorithms to prevent such events by rationally correcting human perception.

Yubin Kim, Ken Gu, Chanwoo Park et al.

Agents, language model (LM)-based systems that are capable of reasoning, planning, and acting are becoming the dominant paradigm for real-world AI applications. Despite this widespread adoption, the principles that determine their performance remain underexplored, leaving practitioners to rely on heuristics rather than principled design choices. We address this gap by deriving quantitative scaling principles for agent systems. We evaluate this across four diverse benchmarks: Finance-Agent, BrowseComp-Plus, PlanCraft, and Workbench. Using five canonical architectures (Single, Independent, Centralized, Decentralized, Hybrid) instantiated across three LLM families, we perform a controlled evaluation spanning 180 configurations with standardized tools and token budgets. We derive a predictive model using empirical coordination metrics, including efficiency, overhead, error amplification, and redundancy, that achieves cross-validated R^2=0.513. We identify three dominant effects: (1) a tool-coordination trade-off: under fixed computational budgets, tool-heavy tasks suffer disproportionately from multi-agent overhead. (2) a capability saturation: coordination yields diminishing or negative returns (beta=-0.408, p<0.001) once single-agent baselines exceed ~45%. (3) topology-dependent error amplification: independent agents amplify errors 17.2x through unchecked propagation, while centralized coordination contains this to 4.4x. Centralized coordination improves performance by 80.9% on parallelizable tasks like financial reasoning, while decentralized coordination excels on dynamic web navigation (+9.2% vs. +0.2%). Yet for sequential reasoning tasks, all multi-agent variants degraded performance by 39-70%. The framework predicts the optimal coordination strategy for 87% of held-out configurations, providing a predictive principle of agentic scaling based on measurable task properties.

Zae Myung Kim, Chanwoo Park, Vipul Raheja et al. · deepmind

Reward-based alignment methods for large language models (LLMs) face two key limitations: vulnerability to reward hacking, where models exploit flaws in the reward signal; and reliance on brittle, labor-intensive prompt engineering when LLMs are used as reward models. We introduce Meta Policy Optimization (MPO), a framework that addresses these challenges by integrating a meta-reward model that dynamically refines the reward model's prompt throughout training. In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain high alignment, providing an adaptive reward signal that resists exploitation by the policy. This meta-learning approach promotes a more stable policy optimization, and greatly reduces the need for manual reward prompt design. It yields performance on par with or better than models guided by extensively hand-crafted reward prompts. Furthermore, we show that MPO maintains its effectiveness across diverse tasks, from essay writing to mathematical reasoning, without requiring specialized reward designs. Beyond standard RLAIF, MPO's meta-learning formulation is readily extensible to higher-level alignment frameworks. Overall, this method addresses theoretical and practical challenges in reward-based RL alignment for LLMs, paving the way for more robust and adaptable alignment strategies. The code and data can be accessed at: https://github.com/minnesotanlp/mpo

Sung Jun Cheon, Jaekyung Cho, Seongho Choi et al.

We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.

Yubin Kim, Chanwoo Park, Taehan Kim et al.

Agent systems often decompose a task across multiple roles, but these roles are typically specified by prompts rather than enforced by access controls. Without enforcement, a team pass rate can mask whether agents actually coordinated or whether one role effectively did another role's work. We present TeamBench, a benchmark with 851 task templates and 931 seeded instances for evaluating agent coordination under operating system-enforced role separation. TeamBench separates specification access, workspace editing, and final certification across Planner, Executor, and Verifier roles, so that no role can read the full requirements, modify the workspace, and certify the final answer. Prompt-only and sandbox-enforced teams reach statistically indistinguishable pass rates, but prompt-only runs produce 3.6 times more cases where the verifier attempts to edit the executor's code. Verifiers approve 49% of submissions that fail the deterministic grader, and removing the verifier improves mean partial score in the ablation. Team value is also conditional. Teams benefit when single agents struggle, but hurt when single agents already perform well. A 40-session human study under the same role separation shows that our benchmark exposes interaction patterns that pass rate misses. Solo participants work through the task directly, human participants paired with agents often collapse into quick approval, and human teams spend more effort coordinating missing information across roles.

Chanwoo Park, Mingyang Liu, Dingwen Kong et al.

Reinforcement learning from human feedback (RLHF) has been an effective technique for aligning AI systems with human values, with remarkable successes in fine-tuning large-language models recently. Most existing RLHF paradigms make the underlying assumption that human preferences are relatively homogeneous, and can be encoded by a single reward model. In this paper, we focus on addressing the issues due to the inherent heterogeneity in human preferences, as well as their potential strategic behavior in providing feedback. Specifically, we propose two frameworks to address heterogeneous human feedback in principled ways: personalization-based one and aggregation-based one. For the former, we propose two approaches based on representation learning and clustering, respectively, for learning multiple reward models that trades off the bias (due to preference heterogeneity) and variance (due to the use of fewer data for learning each model by personalization). We then establish sample complexity guarantees for both approaches. For the latter, we aim to adhere to the single-model framework, as already deployed in the current RLHF paradigm, by carefully aggregating diverse and truthful preferences from humans. We propose two approaches based on reward and preference aggregation, respectively: the former utilizes both utilitarianism and Leximin approaches to aggregate individual reward models, with sample complexity guarantees; the latter directly aggregates the human feedback in the form of probabilistic opinions. Under the probabilistic-opinion-feedback model, we also develop an approach to handle strategic human labelers who may bias and manipulate the aggregated preferences with untruthful feedback. Based on the ideas in mechanism design, our approach ensures truthful preference reporting, with the induced aggregation rule maximizing social welfare functions.

Yubin Kim, Hyewon Jeong, Shan Chen et al.

Hallucinations in foundation models arise from autoregressive training objectives that prioritize token-likelihood optimization over epistemic accuracy, fostering overconfidence and poorly calibrated uncertainty. We define medical hallucination as any model-generated output that is factually incorrect, logically inconsistent, or unsupported by authoritative clinical evidence in ways that could alter clinical decisions. We evaluated 11 foundation models (7 general-purpose, 4 medical-specialized) across seven medical hallucination tasks spanning medical reasoning and biomedical information retrieval. General-purpose models achieved significantly higher proportions of hallucination-free responses than medical-specialized models (median: 76.6% vs 51.3%, difference = 25.2%, 95% CI: 18.7-31.3%, Mann-Whitney U = 27.0, p = 0.012, rank-biserial r = -0.64). Top-performing models such as Gemini-2.5 Pro exceeded 97% accuracy when augmented with chain-of-thought prompting (base: 87.6%), while medical-specialized models like MedGemma ranged from 28.6-61.9% despite explicit training on medical corpora. Chain-of-thought reasoning significantly reduced hallucinations in 86.4% of tested comparisons after FDR correction (q < 0.05), demonstrating that explicit reasoning traces enable self-verification and error detection. Physician audits confirmed that 64-72% of residual hallucinations stemmed from causal or temporal reasoning failures rather than knowledge gaps. A global survey of clinicians (n = 70) validated real-world impact: 91.8% had encountered medical hallucinations, and 84.7% considered them capable of causing patient harm. The underperformance of medical-specialized models despite domain training indicates that safety emerges from sophisticated reasoning capabilities and broad knowledge integration developed during large-scale pre-training, not from narrow optimization.

Chanwoo Park, Xiangyu Liu, Asuman Ozdaglar et al.

Large language models (LLMs) have been increasingly employed for (interactive) decision-making, via the development of LLM-based autonomous agents. Despite their emerging successes, the performance of LLM agents in decision-making has not been fully investigated through quantitative metrics, especially in the multi-agent setting when they interact with each other, a typical scenario in real-world LLM-agent applications. To better understand the limits of LLM agents in these interactive environments, we propose to study their interactions in benchmark decision-making settings in online learning and game theory, through the performance metric of \emph{regret}. We first empirically study the {no-regret} behaviors of LLMs in canonical (non-stationary) online learning problems, as well as the emergence of equilibria when LLM agents interact through playing repeated games. We then provide some theoretical insights into the no-regret behaviors of LLM agents, under certain assumptions on the supervised pre-training and the rationality model of human decision-makers who generate the data. Notably, we also identify (simple) cases where advanced LLMs such as GPT-4 fail to be no-regret. To promote the no-regret behaviors, we propose a novel \emph{unsupervised} training loss of \emph{regret-loss}, which, in contrast to the supervised pre-training loss, does not require the labels of (optimal) actions. We then establish the statistical guarantee of generalization bound for regret-loss minimization, followed by the optimization guarantee that minimizing such a loss may automatically lead to known no-regret learning algorithms. Our further experiments demonstrate the effectiveness of our regret-loss, especially in addressing the above ``regrettable'' cases.

Chanwoo Park, Seungju Han, Xingzhi Guo et al. · stanford

Leveraging multiple large language models (LLMs) to build collaborative multi-agentic workflows has demonstrated significant potential. However, most previous studies focus on prompting the out-of-the-box LLMs, relying on their innate capability for collaboration, which may not improve LLMs' performance as shown recently. In this paper, we introduce a new post-training paradigm MAPoRL (Multi-Agent Post-co-training for collaborative LLMs with Reinforcement Learning), to explicitly elicit the collaborative behaviors and further unleash the power of multi-agentic LLM frameworks. In MAPoRL, multiple LLMs first generate their own responses independently and engage in a multi-turn discussion to collaboratively improve the final answer. In the end, a MAPoRL verifier evaluates both the answer and the discussion, by assigning a score that verifies the correctness of the answer, while adding incentives to encourage corrective and persuasive discussions. The score serves as the co-training reward, and is then maximized through multi-agent RL. Unlike existing LLM post-training paradigms, MAPoRL advocates the co-training of multiple LLMs together using RL for better generalization. Accompanied by analytical insights, our experiments demonstrate that training individual LLMs alone is insufficient to induce effective collaboration. In contrast, multi-agent co-training can boost the collaboration performance across benchmarks, with generalization to unseen domains.

Minyoung Hwang, Luca Weihs, Chanwoo Park et al. · allen-ai

Customizing robotic behaviors to be aligned with diverse human preferences is an underexplored challenge in the field of embodied AI. In this paper, we present Promptable Behaviors, a novel framework that facilitates efficient personalization of robotic agents to diverse human preferences in complex environments. We use multi-objective reinforcement learning to train a single policy adaptable to a broad spectrum of preferences. We introduce three distinct methods to infer human preferences by leveraging different types of interactions: (1) human demonstrations, (2) preference feedback on trajectory comparisons, and (3) language instructions. We evaluate the proposed method in personalized object-goal navigation and flee navigation tasks in ProcTHOR and RoboTHOR, demonstrating the ability to prompt agent behaviors to satisfy human preferences in various scenarios. Project page: https://promptable-behaviors.github.io

Chanwoo Park, Anna Seo Gyeong Choi, Sunghye Cho et al.

Societies worldwide are rapidly entering a super-aged era, making elderly health a pressing concern. The aging population is increasing the burden on national economies and households. Dementia cases are rising significantly with this demographic shift. Recent research using voice-based models and large language models (LLM) offers new possibilities for dementia diagnosis and treatment. Our Chain-of-Thought (CoT) reasoning method combines speech and language models. The process starts with automatic speech recognition to convert speech to text. We add a linear layer to an LLM for Alzheimer's disease (AD) and non-AD classification, using supervised fine-tuning (SFT) with CoT reasoning and cues. This approach showed an 16.7% relative performance improvement compared to methods without CoT prompt reasoning. To the best of our knowledge, our proposed method achieved state-of-the-art performance in CoT approaches.

Kyusu Ahn, Jinpyo Kim, Chanwoo Park et al.

Under-Display Camera (UDC) houses a digital camera lens under a display panel. However, UDC introduces complex degradations such as noise, blur, decrease in transmittance, and flare. Despite the remarkable progress, previous research on UDC mainly focuses on eliminating diffraction in the spatial domain and rarely explores its potential in the frequency domain. It is essential to consider both the spatial and frequency domains effectively. For example, degradations, such as noise and blur, can be addressed by local information (e.g., CNN kernels in the spatial domain). At the same time, tackling flares may require leveraging global information (e.g., the frequency domain). In this paper, we revisit the UDC degradations in the Fourier space and figure out intrinsic frequency priors that imply the presence of the flares. Based on this observation, we propose a novel multi-level DNN architecture called SFIM. It efficiently restores UDC-distorted images by integrating local and global (the collective contribution of all points in the image) information. The architecture exploits CNNs to capture local information and FFT-based models to capture global information. SFIM comprises a spatial domain block (SDB), a Frequency Domain Block (FDB), and an Attention-based Multi-level Integration Block (AMIB). Specifically, SDB focuses more on detailed textures such as noise and blur, FDB emphasizes irregular texture loss in extensive areas such as flare, and AMIB enables effective cross-domain interaction. SFIM's superior performance over state-of-the-art approaches is demonstrated through rigorous quantitative and qualitative assessments across three UDC benchmarks.

Yubin Kim, Chanwoo Park, Hyewon Jeong et al.

Medical Decision-Making (MDM) is a multi-faceted process that requires clinicians to assess complex multi-modal patient data patient, often collaboratively. Large Language Models (LLMs) promise to streamline this process by synthesizing vast medical knowledge and multi-modal health data. However, single-agent are often ill-suited for nuanced medical contexts requiring adaptable, collaborative problem-solving. Our MDAgents addresses this need by dynamically assigning collaboration structures to LLMs based on task complexity, mimicking real-world clinical collaboration and decision-making. This framework improves diagnostic accuracy and supports adaptive responses in complex, real-world medical scenarios, making it a valuable tool for clinicians in various healthcare settings, and at the same time, being more efficient in terms of computing cost than static multi-agent decision making methods.

Yubin Kim, Zhiyuan Hu, Hyewon Jeong et al.

Large Language Models (LLMs) as clinical agents require careful behavioral adaptation. While adept at reactive tasks (e.g., diagnosis reasoning), LLMs often struggle with proactive engagement, like unprompted identification of critical missing information or risks. We introduce BehaviorBench, a comprehensive dataset to evaluate agent behaviors across a clinical assistance spectrum, ranging from reactive query responses to proactive interventions (e.g., clarifying ambiguities, flagging overlooked critical data). Our BehaviorBench experiments reveal LLMs' inconsistent proactivity. To address this, we propose BehaviorSFT, a novel training strategy using behavioral tokens to explicitly condition LLMs for dynamic behavioral selection along this spectrum. BehaviorSFT boosts performance, achieving up to 97.3% overall Macro F1 on BehaviorBench and improving proactive task scores (e.g., from 95.0% to 96.5% for Qwen2.5-7B-Ins). Crucially, blind clinician evaluations confirmed BehaviorSFT-trained agents exhibit more realistic clinical behavior, striking a superior balance between helpful proactivity (e.g., timely, relevant suggestions) and necessary restraint (e.g., avoiding over-intervention) versus standard fine-tuning or explicit instructed agents.

Chanwoo Park, Suyoung Park, JiA Kang et al.

We present Ko-MuSR, the first benchmark to comprehensively evaluate multistep, soft reasoning in long Korean narratives while minimizing data contamination. Built following MuSR, Ko-MuSR features fully Korean narratives, reasoning chains, and multiple-choice questions verified by human annotators for logical consistency and answerability. Evaluations of four large language models -- two multilingual and two Korean-specialized -- show that multilingual models outperform Korean-focused ones even in Korean reasoning tasks, indicating cross-lingual generalization of reasoning ability. Carefully designed prompting strategies, which combine few-shot examples, reasoning traces, and task-specific hints, further boost accuracy, approaching human-level performance. Ko-MuSR offers a solid foundation for advancing Korean NLP by enabling systematic evaluation of long-context reasoning and prompting strategies.

Chanwoo Park, Suyoung Park, Yelim Ahn et al.

While traditional line-level filtering techniques, such as line-level deduplication and trailing-punctuation filters, are commonly used, these basic methods can sometimes discard valuable content, negatively affecting downstream performance. In this paper, we introduce two methods-pattern-aware line-level deduplication (PLD) and pattern-aware trailing punctuation filtering (PTF)-by enhancing the conventional filtering techniques. Our approach not only considers line-level signals but also takes into account their sequential distribution across documents, enabling us to retain structurally important content that might otherwise be removed. We evaluate these proposed methods by training small language models (1 B parameters) in both English and Korean. The results demonstrate that our methods consistently improve performance on multiple-choice benchmarks and significantly enhance generative question-answering accuracy on both SQuAD v1 and KorQuAD v1.

Jinpyo Kim, Gyeongje Cho, Chanwoo Park et al.

Since state-of-the-art LLMs often underperform in languages other than English or Chinese, improving the capability of LLMs in new languages has become an essential task. Moreover, LLMs' entire end-to-end training process remains largely unknown to the public due to proprietary reasons, technical complexity, inconsistent documentation, and ethical considerations. The complete picture remains a closely guarded secret within the industry. This paper presents methods to adapt an existing English-based LLM to Korean in a low-budget scenario. We describe the entire end-to-end process: collecting Korean datasets, preprocessing the data, training the model, creating downstream benchmarks, and conducting evaluations. The evaluation results indicate that our method can effectively and cost-efficiently add new language capabilities to existing LLMs. Our new bilingual models, Thunder-LLM and Thunder-LLM-Ins, achieve superior Korean performance compared to state-of-the-art models while utilizing minimal data and computational resources. We share our comprehensive experience and make the code publicly available.

Gyeongje Cho, Yeonkyoun So, Chanwoo Park et al.

This paper introduces Thunder-Tok, a new Korean tokenizer designed to reduce token fertility without compromising model performance. Our approach uses a rule-based pre-tokenization method that aligns with the linguistic structure of the Korean language. We also create a seed vocabulary containing tokens that resemble linguistic units and employ a branching entropy-based selection algorithm. These techniques increase the average token length, thus lowering fertility while preserving linguistic information. Experimental results indicate that Thunder-Tok reduces fertility by approximately 10% (i.e., reduces the number of tokens by 10%, improving the inference speed by 10%) compared to BPE without compromising performance across various downstream tasks. These findings demonstrate that our linguistically informed approach is effective and practical for designing efficient tokenizers for language models.

Yubin Kim, Hyewon Jeong, Chanwoo Park et al.

Large language models (LLMs) deployed as agents introduce significant safety risks in clinical settings due to their potential for error and single points of failure. We introduce Tiered Agentic Oversight (TAO), a hierarchical multi-agent system that enhances AI safety through layered, automated supervision. Inspired by clinical hierarchies (e.g., nurse-physician-specialist) in hospital, TAO routes tasks to specialized agents based on complexity, creating a robust safety framework through automated inter- and intra-tier communication and role-playing. Crucially, this hierarchical structure functions as an effective error-correction mechanism, absorbing up to 24% of individual agent errors before they can compound. Our experiments reveal TAO outperforms single-agent and other multi-agent systems on 4 out of 5 healthcare safety benchmarks, with up to an 8.2% improvement. Ablation studies confirm key design principles of the system: (i) its adaptive architecture is over 3% safer than static, single-tier configurations, and (ii) its lower tiers are indispensable, as their removal causes the most significant degradation in overall safety. Finally, we validated the system's synergy with human doctors in a user study where a physician, acting as the highest tier agent, provided corrective feedback that improved medical triage accuracy from 40% to 60%. Project Page: https://tiered-agentic-oversight.github.io/

Kyusu Ahn, JiSoo Kim, Sangik Lee et al.

Even though an Under-Display Camera (UDC) is an advanced imaging system, the display panel significantly degrades captured images or videos, introducing low transmittance, blur, noise, and flare issues. Tackling such issues is challenging because of the complex degradation of UDCs, including diverse flare patterns. However, no dataset contains videos of real-world UDC degradation. In this paper, we propose a real-world UDC video dataset called UDC-VIT. Unlike existing datasets, UDC-VIT exclusively includes human motions for facial recognition. We propose a video-capturing system to acquire clean and UDC-degraded videos of the same scene simultaneously. Then, we align a pair of captured videos frame by frame, using discrete Fourier transform (DFT). We compare UDC-VIT with six representative UDC still image datasets and two existing UDC video datasets. Using six deep-learning models, we compare UDC-VIT and an existing synthetic UDC video dataset. The results indicate the ineffectiveness of models trained on earlier synthetic UDC video datasets, as they do not reflect the actual characteristics of UDC-degraded videos. We also demonstrate the importance of effective UDC restoration by evaluating face recognition accuracy concerning PSNR, SSIM, and LPIPS scores. UDC-VIT is available at our official GitHub repository.

Jae Myung Kim, Hyungjin Kim, Chanwoo Park et al.

In recent years, deep neural networks (DNN) have become a highly active area of research, and shown remarkable achievements on a variety of computer vision tasks. DNNs, however, are known to often make overconfident yet incorrect predictions on out-of-distribution samples, which can be a major obstacle to real-world deployments because the training dataset is always limited compared to diverse real-world samples. Thus, it is fundamental to provide guarantees of robustness to the distribution shift between training and test time when we construct DNN models in practice. Moreover, in many cases, the deep learning models are deployed as black boxes and the performance has been already optimized for a training dataset, thus changing the black box itself can lead to performance degradation. We here study the robustness to the geometric transformations in a specific condition where the black-box image classifier is given. We propose an additional learner, \emph{REinforcement Spatial Transform learner (REST)}, that transforms the warped input data into samples regarded as in-distribution by the black-box models. Our work aims to improve the robustness by adding a REST module in front of any black boxes and training only the REST module without retraining the original black box model in an end-to-end manner, i.e. we try to convert the real-world data into training distribution which the performance of the black-box model is best suited for. We use a confidence score that is obtained from the black-box model to determine whether the transformed input is drawn from in-distribution. We empirically show that our method has an advantage in generalization to geometric transformations and sample efficiency.

Chanwoo Park, Jae Myung Kim, Seok Hyeon Ha et al.

Deep neural networks(NNs) have achieved impressive performance, often exceed human performance on many computer vision tasks. However, one of the most challenging issues that still remains is that NNs are overconfident in their predictions, which can be very harmful when this arises in safety critical applications. In this paper, we show that predictive uncertainty can be efficiently estimated when we incorporate the concept of gradients uncertainty into posterior sampling. The proposed method is tested on two different datasets, MNIST for in-distribution confusing examples and notMNIST for out-of-distribution data. We show that our method is able to efficiently represent predictive uncertainty on both datasets.