Jaewoo Kang

Wonjin Yoon, Richard Jackson, Elliot Ford et al. · harvard

In order to assist the drug discovery/development process, pharmaceutical companies often apply biomedical NER and linking techniques over internal and public corpora. Decades of study of the field of BioNLP has produced a plethora of algorithms, systems and datasets. However, our experience has been that no single open source system meets all the requirements of a modern pharmaceutical company. In this work, we describe these requirements according to our experience of the industry, and present Kazu, a highly extensible, scalable open source framework designed to support BioNLP for the pharmaceutical sector. Kazu is a built around a computationally efficient version of the BERN2 NER model (TinyBERN2), and subsequently wraps several other BioNLP technologies into one coherent system. KAZU framework is open-sourced: https://github.com/AstraZeneca/KAZU

Seongjun Yun, Seoyoon Kim, Junhyun Lee et al.

Graph Neural Networks (GNNs) have been widely applied to various fields for learning over graph-structured data. They have shown significant improvements over traditional heuristic methods in various tasks such as node classification and graph classification. However, since GNNs heavily rely on smoothed node features rather than graph structure, they often show poor performance than simple heuristic methods in link prediction where the structural information, e.g., overlapped neighborhoods, degrees, and shortest paths, is crucial. To address this limitation, we propose Neighborhood Overlap-aware Graph Neural Networks (Neo-GNNs) that learn useful structural features from an adjacency matrix and estimate overlapped neighborhoods for link prediction. Our Neo-GNNs generalize neighborhood overlap-based heuristic methods and handle overlapped multi-hop neighborhoods. Our extensive experiments on Open Graph Benchmark datasets (OGB) demonstrate that Neo-GNNs consistently achieve state-of-the-art performance in link prediction. Our code is publicly available at https://github.com/seongjunyun/Neo_GNNs.

Minbyul Jeong, Jaewoo Kang

Named entity recognition (NER) is a fundamental part of extracting information from documents in biomedical applications. A notable advantage of NER is its consistency in extracting biomedical entities in a document context. Although existing document NER models show consistent predictions, they still do not meet our expectations. We investigated whether the adjectives and prepositions within an entity cause a low label consistency, which results in inconsistent predictions. In this paper, we present our method, ConNER, which enhances the label dependency of modifiers (e.g., adjectives and prepositions) to achieve higher label agreement. ConNER refines the draft labels of the modifiers to improve the output representations of biomedical entities. The effectiveness of our method is demonstrated on four popular biomedical NER datasets; in particular, its efficacy is proved on two datasets with 7.5-8.6% absolute improvements in the F1 score. We interpret that our ConNER method is effective on datasets that have intrinsically low label consistency. In the qualitative analysis, we demonstrate how our approach makes the NER model generate consistent predictions. Our code and resources are available at https://github.com/dmis-lab/ConNER/.

Gangwoo Kim, Sungdong Kim, Byeongguk Jeon et al.

Questions in open-domain question answering are often ambiguous, allowing multiple interpretations. One approach to handling them is to identify all possible interpretations of the ambiguous question (AQ) and to generate a long-form answer addressing them all, as suggested by Stelmakh et al., (2022). While it provides a comprehensive response without bothering the user for clarification, considering multiple dimensions of ambiguity and gathering corresponding knowledge remains a challenge. To cope with the challenge, we propose a novel framework, Tree of Clarifications (ToC): It recursively constructs a tree of disambiguations for the AQ -- via few-shot prompting leveraging external knowledge -- and uses it to generate a long-form answer. ToC outperforms existing baselines on ASQA in a few-shot setup across the metrics, while surpassing fully-supervised baselines trained on the whole training set in terms of Disambig-F1 and Disambig-ROUGE. Code is available at https://github.com/gankim/tree-of-clarifications.

Mujeen Sung, Jungsoo Park, Jaewoo Kang et al. · princeton

Recent developments of dense retrieval rely on quality representations of queries and contexts from pre-trained query and context encoders. In this paper, we introduce TOUR (Test-Time Optimization of Query Representations), which further optimizes instance-level query representations guided by signals from test-time retrieval results. We leverage a cross-encoder re-ranker to provide fine-grained pseudo labels over retrieval results and iteratively optimize query representations with gradient descent. Our theoretical analysis reveals that TOUR can be viewed as a generalization of the classical Rocchio algorithm for pseudo relevance feedback, and we present two variants that leverage pseudo-labels as hard binary or soft continuous labels. We first apply TOUR on phrase retrieval with our proposed phrase re-ranker, and also evaluate its effectiveness on passage retrieval with an off-the-shelf re-ranker. TOUR greatly improves end-to-end open-domain question answering accuracy, as well as passage retrieval performance. TOUR also consistently improves direct re-ranking by up to 2.0% while running 1.3-2.4x faster with an efficient implementation.

Heedou Kim, Mogan Gim, Donghee Choi et al.

The rapid evolution of online scams, driven by transnational networks and mass produced social engineering scenarios, has exposed the speed limitations of conventional detection, necessitating tighter interagency coordination. While LLMs show promise in scam identification, their role in accelerating integrated response frameworks remains underexplored. We propose Counter Scam, a unified LLM based multiagent framework that orchestrates end to end response from initial detection to crime investigation. The framework first proposes safe data guidelines, emphasizing nonpublic scam data and secure dataset construction via scam specific NER. Developed with insights from 37 stakeholders to reduce delays and improve analytical efficiency, the system integrates CSRA for multiagent mitigation, CSRT comprising nine role aligned NLP tasks, and CSRD, a corpus of 185,300 scam cases and 38,587 knowledge entries. Experiments show that fine tuned sLLMs surpass commercial models by more than 10% across all CSRT tasks and achieve a 0.24 F1 improvement in scam specific NER. These results demonstrate the framework's capability to enable rapid and collaborative mitigation of online scams.

Gangwoo Kim, Sungdong Kim, Kang Min Yoo et al.

In this paper, we introduce a novel framework, SIMSEEK, (Simulating information-Seeking conversation from unlabeled documents), and compare its two variants. In our baseline SIMSEEK-SYM, a questioner generates follow-up questions upon the predetermined answer by an answerer. On the contrary, SIMSEEK-ASYM first generates the question and then finds its corresponding answer under the conversational context. Our experiments show that they can synthesize effective training resources for CQA and conversational search tasks. As a result, conversations from SIMSEEK-ASYM not only make more improvements in our experiments but also are favorably reviewed in a human evaluation. We finally release a large-scale resource of synthetic conversations, WIKI-SIMSEEK, containing 2 million CQA pairs built upon Wikipedia documents. With the dataset, our CQA model achieves state-of-the-art performance on a recent CQA benchmark, QuAC.

Seungheun Baek, Soyon Park, Yan Ting Chok et al.

Predicting cellular responses to various perturbations is a critical focus in drug discovery and personalized therapeutics, with deep learning models playing a significant role in this endeavor. Single-cell datasets contain technical artifacts that may hinder the predictability of such models, which poses quality control issues highly regarded in this area. To address this, we propose CRADLE-VAE, a causal generative framework tailored for single-cell gene perturbation modeling, enhanced with counterfactual reasoning-based artifact disentanglement. Throughout training, CRADLE-VAE models the underlying latent distribution of technical artifacts and perturbation effects present in single-cell datasets. It employs counterfactual reasoning to effectively disentangle such artifacts by modulating the latent basal spaces and learns robust features for generating cellular response data with improved quality. Experimental results demonstrate that this approach improves not only treatment effect estimation performance but also generative quality as well. The CRADLE-VAE codebase is publicly available at https://github.com/dmis-lab/CRADLE-VAE.

Donghee Choi, Mogan Gim, Samy Badreddine et al.

Determining proper quantities for ingredients is an essential part of cooking practice from the perspective of enriching tastiness and promoting healthiness. We introduce KitchenScale, a fine-tuned Pre-trained Language Model (PLM) that predicts a target ingredient's quantity and measurement unit given its recipe context. To effectively train our KitchenScale model, we formulate an ingredient quantity prediction task that consists of three sub-tasks which are ingredient measurement type classification, unit classification, and quantity regression task. Furthermore, we utilized transfer learning of cooking knowledge from recipe texts to PLMs. We adopted the Discrete Latent Exponent (DExp) method to cope with high variance of numerical scales in recipe corpora. Experiments with our newly constructed dataset and recommendation examples demonstrate KitchenScale's understanding of various recipe contexts and generalizability in predicting ingredient quantities. We implemented a web application for KitchenScale to demonstrate its functionality in recommending ingredient quantities expressed in numerals (e.g., 2) with units (e.g., ounce).

Mogan Gim, Donghee Choi, Kana Maruyama et al.

We propose a computational approach for recipe ideation, a downstream task that helps users select and gather ingredients for creating dishes. To perform this task, we developed RecipeMind, a food affinity score prediction model that quantifies the suitability of adding an ingredient to set of other ingredients. We constructed a large-scale dataset containing ingredient co-occurrence based scores to train and evaluate RecipeMind on food affinity score prediction. Deployed in recipe ideation, RecipeMind helps the user expand an initial set of ingredients by suggesting additional ingredients. Experiments and qualitative analysis show RecipeMind's potential in fulfilling its assistive role in cuisine domain.

Chanhwi Kim, Hyunjae Kim, Sihyeon Park et al.

Generative models have become widely used in biomedical entity linking (BioEL) due to their excellent performance and efficient memory usage. However, these models are usually trained only with positive samples, i.e., entities that match the input mention's identifier, and do not explicitly learn from hard negative samples, which are entities that look similar but have different meanings. To address this limitation, we introduce ANGEL (Learning from Negative Samples in Biomedical Generative Entity Linking), the first framework that trains generative BioEL models using negative samples. Specifically, a generative model is initially trained to generate positive entity names from the knowledge base for given input entities. Subsequently, both correct and incorrect outputs are gathered from the model's top-k predictions. Finally, the model is updated to prioritize the correct predictions through preference optimization. Our models outperform the previous best baseline models by up to an average top-1 accuracy of 1.4% on five benchmarks. When incorporating our framework into pre-training, the performance improvement increases further to 1.7%, demonstrating its effectiveness in both the pre-training and fine-tuning stages. The code and model weights are available at https://github.com/dmis-lab/ANGEL.

Hyunjae Kim, Jaehyo Yoo, Seunghyun Yoon et al.

Most weakly supervised named entity recognition (NER) models rely on domain-specific dictionaries provided by experts. This approach is infeasible in many domains where dictionaries do not exist. While a phrase retrieval model was used to construct pseudo-dictionaries with entities retrieved from Wikipedia automatically in a recent study, these dictionaries often have limited coverage because the retriever is likely to retrieve popular entities rather than rare ones. In this study, we present a novel framework, HighGEN, that generates NER datasets with high-coverage pseudo-dictionaries. Specifically, we create entity-rich dictionaries with a novel search method, called phrase embedding search, which encourages the retriever to search a space densely populated with various entities. In addition, we use a new verification process based on the embedding distance between candidate entity mentions and entity types to reduce the false-positive noise in weak labels generated by high-coverage dictionaries. We demonstrate that HighGEN outperforms the previous best model by an average F1 score of 4.7 across five NER benchmark datasets.

Jinyoung Park, Seongjun Yun, Hyeonjin Park et al.

Transformer-based models have recently shown success in representation learning on graph-structured data beyond natural language processing and computer vision. However, the success is limited to small-scale graphs due to the drawbacks of full dot-product attention on graphs such as the quadratic complexity with respect to the number of nodes and message aggregation from enormous irrelevant nodes. To address these issues, we propose Deformable Graph Transformer (DGT) that performs sparse attention via dynamically sampled relevant nodes for efficiently handling large-scale graphs with a linear complexity in the number of nodes. Specifically, our framework first constructs multiple node sequences with various criteria to consider both structural and semantic proximity. Then, combining with our learnable Katz Positional Encodings, the sparse attention is applied to the node sequences for learning node representations with a significantly reduced computational cost. Extensive experiments demonstrate that our DGT achieves state-of-the-art performance on 7 graph benchmark datasets with 2.5 - 449 times less computational cost compared to transformer-based graph models with full attention.

Junhyun Lee, Bumsoo Kim, Minji Jeon et al.

Graph Neural Networks (GNNs) have proven to be effective in processing and learning from graph-structured data. However, previous works mainly focused on understanding single graph inputs while many real-world applications require pair-wise analysis for graph-structured data (e.g., scene graph matching, code searching, and drug-drug interaction prediction). To this end, recent works have shifted their focus to learning the interaction between pairs of graphs. Despite their improved performance, these works were still limited in that the interactions were considered at the node-level, resulting in high computational costs and suboptimal performance. To address this issue, we propose a novel and efficient graph-level approach for extracting interaction representations using co-attention in graph pooling. Our method, Co-Attention Graph Pooling (CAGPool), exhibits competitive performance relative to existing methods in both classification and regression tasks using real-world datasets, while maintaining lower computational complexity.

Hyunjun Lee, Junhyun Lee, Taehwa Choi et al.

Time-to-event analysis, also known as survival analysis, aims to predict the time of occurrence of an event, given a set of features. One of the major challenges in this area is dealing with censored data, which can make learning algorithms more complex. Traditional methods such as Cox's proportional hazards model and the accelerated failure time (AFT) model have been popular in this field, but they often require assumptions such as proportional hazards and linearity. In particular, the AFT models often require pre-specified parametric distributional assumptions. To improve predictive performance and alleviate strict assumptions, there have been many deep learning approaches for hazard-based models in recent years. However, representation learning for AFT has not been widely explored in the neural network literature, despite its simplicity and interpretability in comparison to hazard-focused methods. In this work, we introduce the Deep AFT Rank-regression model for Time-to-event prediction (DART). This model uses an objective function based on Gehan's rank statistic, which is efficient and reliable for representation learning. On top of eliminating the requirement to establish a baseline event time distribution, DART retains the advantages of directly predicting event time in standard AFT models. The proposed method is a semiparametric approach to AFT modeling that does not impose any distributional assumptions on the survival time distribution. This also eliminates the need for additional hyperparameters or complex model architectures, unlike existing neural network-based AFT models. Through quantitative analysis on various benchmark datasets, we have shown that DART has significant potential for modeling high-throughput censored time-to-event data.

Jueon Park, Wonjune Jang, Jiwoo Lee et al.

Large Language Models (LLMs) and Vision Language Models (VLMs) have recently shown promising capabilities in various scientific domain. In particular, these advances have opened new opportunities in drug discovery, where the ability to understand and modify molecular structures is critical for optimizing drug properties such as efficacy and toxicity. However, existing models and benchmarks often overlook toxicity-related challenges, focusing primarily on general property optimization without adequately addressing safety concerns. In addition, even existing toxicity repair benchmarks suffer from limited data diversity, low structural validity of generated molecules, and heavy reliance on proxy models for toxicity assessment. To address these limitations, we propose MolDeTox, a novel benchmark for molecular detoxification, designed to enable fine-grained and reliable evaluation of toxicity-aware molecular optimization across stepwise tasks. We evaluate a wide range of general-purpose LLMs and VLMs under diverse settings, and demonstrate that understanding and generating molecules at the fragment-level improves structural validity and enhances the quality of generated molecules. Moreover, through detailed task-level performance analysis, MolDeTox provides an interpretable benchmark that enables a deeper understanding of the detoxification process. Our dataset is available at : https://huggingface.co/datasets/MolDeTox/MolDeTox

Seongyun Lee, Hyunjae Kim, Jaewoo Kang

Question answering (QA) models often rely on large-scale training datasets, which necessitates the development of a data generation framework to reduce the cost of manual annotations. Although several recent studies have aimed to generate synthetic questions with single-span answers, no study has been conducted on the creation of list questions with multiple, non-contiguous spans as answers. To address this gap, we propose LIQUID, an automated framework for generating list QA datasets from unlabeled corpora. We first convert a passage from Wikipedia or PubMed into a summary and extract named entities from the summarized text as candidate answers. This allows us to select answers that are semantically correlated in context and is, therefore, suitable for constructing list questions. We then create questions using an off-the-shelf question generator with the extracted entities and original passage. Finally, iterative filtering and answer expansion are performed to ensure the accuracy and completeness of the answers. Using our synthetic data, we significantly improve the performance of the previous best list QA models by exact-match F1 scores of 5.0 on MultiSpanQA, 1.9 on Quoref, and 2.8 averaged across three BioASQ benchmarks.

Chanwoong Yoon, Taewhoo Lee, Hyeon Hwang et al.

Retrieval-augmented generation supports language models to strengthen their factual groundings by providing external contexts. However, language models often face challenges when given extensive information, diminishing their effectiveness in solving questions. Context compression tackles this issue by filtering out irrelevant information, but current methods still struggle in realistic scenarios where crucial information cannot be captured with a single-step approach. To overcome this limitation, we introduce CompAct, a novel framework that employs an active strategy to condense extensive documents without losing key information. Our experiments demonstrate that CompAct brings significant improvements in both performance and compression rate on multi-hop question-answering benchmarks. CompAct flexibly operates as a cost-efficient plug-in module with various off-the-shelf retrievers or readers, achieving exceptionally high compression rates (47x).

Gangwoo Kim, Hajung Kim, Lei Ji et al.

In this paper, we introduce CheXOFA, a new pre-trained vision-language model (VLM) for the chest X-ray domain. Our model is initially pre-trained on various multimodal datasets within the general domain before being transferred to the chest X-ray domain. Following a prominent VLM, we unify various domain-specific tasks into a simple sequence-to-sequence schema. It enables the model to effectively learn the required knowledge and skills from limited resources in the domain. Demonstrating superior performance on the benchmark datasets provided by the BioNLP shared task, our model benefits from its training across multiple tasks and domains. With subtle techniques including ensemble and factual calibration, our system achieves first place on the RadSum23 leaderboard for the hidden test set.

Hyunjae Kim, Hyeon Hwang, Jiwoo Lee et al.

While recent advancements in commercial large language models (LM) have shown promising results in medical tasks, their closed-source nature poses significant privacy and security concerns, hindering their widespread use in the medical field. Despite efforts to create open-source models, their limited parameters often result in insufficient multi-step reasoning capabilities required for solving complex medical problems. To address this, we introduce Meerkat, a new family of medical AI systems ranging from 7 to 70 billion parameters. The models were trained using our new synthetic dataset consisting of high-quality chain-of-thought reasoning paths sourced from 18 medical textbooks, along with diverse instruction-following datasets. Our systems achieved remarkable accuracy across six medical benchmarks, surpassing the previous best models such as MediTron and BioMistral, and GPT-3.5 by a large margin. Notably, Meerkat-7B surpassed the passing threshold of the United States Medical Licensing Examination (USMLE) for the first time for a 7B-parameter model, while Meerkat-70B outperformed GPT-4 by an average of 1.3%. Additionally, Meerkat-70B correctly diagnosed 21 out of 38 complex clinical cases, outperforming humans' 13.8 and closely matching GPT-4's 21.8. Our systems offered more detailed free-form responses to clinical queries compared to existing small models, approaching the performance level of large commercial models. This significantly narrows the performance gap with large LMs, showcasing its effectiveness in addressing complex medical challenges.

Jiwoong Sohn, Yein Park, Chanwoong Yoon et al.

Large language models (LLM) hold significant potential for applications in biomedicine, but they struggle with hallucinations and outdated knowledge. While retrieval-augmented generation (RAG) is generally employed to address these issues, it also has its own set of challenges: (1) LLMs are vulnerable to irrelevant or incorrect context, (2) medical queries are often not well-targeted for helpful information, and (3) retrievers are prone to bias toward the specific source corpus they were trained on. In this study, we present RAG$^2$ (RAtionale-Guided RAG), a new framework for enhancing the reliability of RAG in biomedical contexts. RAG$^2$ incorporates three key innovations: a small filtering model trained on perplexity-based labels of rationales, which selectively augments informative snippets of documents while filtering out distractors; LLM-generated rationales as queries to improve the utility of retrieved snippets; a structure designed to retrieve snippets evenly from a comprehensive set of four biomedical corpora, effectively mitigating retriever bias. Our experiments demonstrate that RAG$^2$ improves the state-of-the-art LLMs of varying sizes, with improvements of up to 6.1\%, and it outperforms the previous best medical RAG model by up to 5.6\% across three medical question-answering benchmarks. Our code is available at https://github.com/dmis-lab/RAG2.

Heedou Kim, Dain Kim, Jiwoo Lee et al.

Crime situations are race against time. An AI-assisted criminal investigation system, providing prompt but precise legal counsel is in need for police officers. We introduce LAPIS (Language Model Augmented Police Investigation System), an automated system that assists police officers to perform rational and legal investigative actions. We constructed a finetuning dataset and retrieval knowledgebase specialized in crime investigation legal reasoning task. We extended the dataset's quality by incorporating manual curation efforts done by a group of domain experts. We then finetuned the pretrained weights of a smaller Korean language model to the newly constructed dataset and integrated it with the crime investigation knowledgebase retrieval approach. Experimental results show LAPIS' potential in providing reliable legal guidance for police officers, even better than the proprietary GPT-4 model. Qualitative analysis on the rationales generated by LAPIS demonstrate the model's reasoning ability to leverage the premises and derive legally correct conclusions.

Sumin Seo, JaeWoong Shin, Jaewoo Kang et al.

Deep learning has shown great potential in assisting radiologists in reading chest X-ray (CXR) images, but its need for expensive annotations for improving performance prevents widespread clinical application. Visual language pre-training (VLP) can alleviate the burden and cost of annotation by leveraging routinely generated reports for radiographs, which exist in large quantities as well as in paired form (image-text pairs). Additionally, extensions to localization-aware VLPs are being proposed to address the needs for accurate localization of abnormalities for computer-aided diagnosis (CAD) in CXR. However, we find that the formulation proposed by locality-aware VLP literature actually leads to a loss in spatial relationships required for downstream localization tasks. Therefore, we propose Empowering Locality of VLP with Intra-modal Similarity, ELVIS, a VLP aware of intra-modal locality, to better preserve the locality within radiographs or reports, which enhances the ability to comprehend location references in text reports. Our locality-aware VLP method significantly outperforms state-of-the art baselines in multiple segmentation tasks and the MS-CXR phrase grounding task. Qualitatively, we show that ELVIS focuses well on regions of interest described in the report text compared to prior approaches, allowing for enhanced interpretability.

Taewhoo Lee, Chanwoong Yoon, Kyochul Jang et al.

Recent advancements in large language models (LLM) capable of processing extremely long texts highlight the need for a dedicated evaluation benchmark to assess their long-context capabilities. However, existing methods, like the needle-in-a-haystack test, do not effectively assess whether these models fully utilize contextual information, raising concerns about the reliability of current evaluation techniques. To thoroughly examine the effectiveness of existing benchmarks, we introduce a new metric called information coverage (IC), which quantifies the proportion of the input context necessary for answering queries. Our findings indicate that current benchmarks exhibit low IC; although the input context may be extensive, the actual usable context is often limited. To address this, we present ETHIC, a novel benchmark designed to assess LLMs' ability to leverage the entire context. Our benchmark comprises 1,986 test instances spanning four long-context tasks with high IC scores in the domains of books, debates, medicine, and law. Our evaluations reveal significant performance drops in contemporary LLMs, highlighting a critical challenge in managing long contexts. Our benchmark is available at https://github.com/dmis-lab/ETHIC.

Donghee Choi, Jinkyu Kim, Mogan Gim et al.

Utilizing market forecasts is pivotal in optimizing portfolio selection strategies. We introduce DeepClair, a novel framework for portfolio selection. DeepClair leverages a transformer-based time-series forecasting model to predict market trends, facilitating more informed and adaptable portfolio decisions. To integrate the forecasting model into a deep reinforcement learning-driven portfolio selection framework, we introduced a two-step strategy: first, pre-training the time-series model on market data, followed by fine-tuning the portfolio selection architecture using this model. Additionally, we investigated the optimization technique, Low-Rank Adaptation (LoRA), to enhance the pre-trained forecasting model for fine-tuning in investment scenarios. This work bridges market forecasting and portfolio selection, facilitating the advancement of investment strategies.

Jungwoo Park, Young Jin Ahn, Kee-Eung Kim et al.

Understanding the internal computations of large language models (LLMs) is crucial for aligning them with human values and preventing undesirable behaviors like toxic content generation. However, mechanistic interpretability is hindered by polysemanticity -- where individual neurons respond to multiple, unrelated concepts. While Sparse Autoencoders (SAEs) have attempted to disentangle these features through sparse dictionary learning, they have compromised LLM performance due to reliance on post-hoc reconstruction loss. To address this issue, we introduce Mixture of Monosemantic Experts for Transformers (Monet) architecture, which incorporates sparse dictionary learning directly into end-to-end Mixture-of-Experts pretraining. Our novel expert decomposition method enables scaling the expert count to 262,144 per layer while total parameters scale proportionally to the square root of the number of experts. Our analyses demonstrate mutual exclusivity of knowledge across experts and showcase the parametric knowledge encapsulated within individual experts. Moreover, Monet allows knowledge manipulation over domains, languages, and toxicity mitigation without degrading general performance. Our pursuit of transparent LLMs highlights the potential of scaling expert counts to enhance mechanistic interpretability and directly resect the internal knowledge to fundamentally adjust model behavior. The source code and pretrained checkpoints are available at https://github.com/dmis-lab/Monet.

Yein Park, Chanwoong Yoon, Jungwoo Park et al.

Large language models (LLMs) have brought significant changes to many aspects of our lives. However, assessing and ensuring their chronological knowledge remains challenging. Existing approaches fall short in addressing the temporal adaptability of knowledge, often relying on a fixed time-point view. To overcome this, we introduce ChroKnowBench, a benchmark dataset designed to evaluate chronologically accumulated knowledge across three key aspects: multiple domains, time dependency, temporal state. Our benchmark distinguishes between knowledge that evolves (e.g., personal history, scientific discoveries, amended laws) and knowledge that remain constant (e.g., mathematical truths, commonsense facts). Building on this benchmark, we present ChroKnowledge (Chronological Categorization of Knowledge), a novel sampling-based framework for evaluating LLMs' non-parametric chronological knowledge. Our evaluation led to the following observations: (1) The ability of eliciting temporal knowledge varies depending on the data format that model was trained on. (2) LLMs partially recall knowledge or show a cut-off at temporal boundaries rather than recalling all aspects of knowledge correctly. Thus, we apply our ChroKnowPrompt, an in-depth prompting to elicit chronological knowledge by traversing step-by-step through the surrounding time spans. We observe that it successfully recalls objects across both open-source and proprietary LLMs, demonstrating versatility, though it faces challenges with dynamic datasets and unstructured formats.

Junhyun Lee, Veronika Thost, Bumsoo Kim et al.

Message Passing Neural Networks (MPNNs) hold a key position in machine learning on graphs, but they struggle with unintended behaviors, such as over-smoothing and over-squashing, due to irregular data structures. The observation and formulation of these limitations have become foundational in constructing more informative graph representations. In this paper, we delve into the limitations of MPNNs, focusing on aspects that have previously been overlooked. Our observations reveal that even within a single layer, the information specific to an individual node can become significantly diluted. To delve into this phenomenon in depth, we present the concept of Over-dilution and formulate it with two dilution factors: intra-node dilution for attribute-level and inter-node dilution for node-level representations. We also introduce a transformer-based solution that alleviates over-dilution and complements existing node embedding methods like MPNNs. Our findings provide new insights and contribute to the development of informative representations. The implementation and supplementary materials are publicly available at https://github.com/LeeJunHyun/NATR.

Jaehoon Yun, Jiwoong Sohn, Jungwoo Park et al.

Large language models have shown promise in clinical decision making, but current approaches struggle to localize and correct errors at specific steps of the reasoning process. This limitation is critical in medicine, where identifying and addressing reasoning errors is essential for accurate diagnosis and effective patient care. We introduce Med-PRM, a process reward modeling framework that leverages retrieval-augmented generation to verify each reasoning step against established medical knowledge bases. By verifying intermediate reasoning steps with evidence retrieved from clinical guidelines and literature, our model can precisely assess the reasoning quality in a fine-grained manner. Evaluations on five medical QA benchmarks and two open-ended diagnostic tasks demonstrate that Med-PRM achieves state-of-the-art performance, with improving the performance of base models by up to 13.50% using Med-PRM. Moreover, we demonstrate the generality of Med-PRM by integrating it in a plug-and-play fashion with strong policy models such as Meerkat, achieving over 80\% accuracy on MedQA for the first time using small-scale models of 8 billion parameters. Our code and data are available at: https://med-prm.github.io/

Jueon Park, Yein Park, Minju Song et al.

Drug toxicity remains a major challenge in pharmaceutical development. Recent machine learning models have improved in silico toxicity prediction, but their reliance on annotated data and lack of interpretability limit their applicability. This limits their ability to capture organ-specific toxicities driven by complex biological mechanisms. Large language models (LLMs) offer a promising alternative through step-by-step reasoning and integration of textual data, yet prior approaches lack biological context and transparent rationale. To address this issue, we propose CoTox, a novel framework that integrates LLM with chain-of-thought (CoT) reasoning for multi-toxicity prediction. CoTox combines chemical structure data, biological pathways, and gene ontology (GO) terms to generate interpretable toxicity predictions through step-by-step reasoning. Using GPT-4o, we show that CoTox outperforms both traditional machine learning and deep learning model. We further examine its performance across various LLMs to identify where CoTox is most effective. Additionally, we find that representing chemical structures with IUPAC names, which are easier for LLMs to understand than SMILES, enhances the model's reasoning ability and improves predictive performance. To demonstrate its practical utility in drug development, we simulate the treatment of relevant cell types with drug and incorporated the resulting biological context into the CoTox framework. This approach allow CoTox to generate toxicity predictions aligned with physiological responses, as shown in case study. This result highlights the potential of LLM-based frameworks to improve interpretability and support early-stage drug safety assessment. The code and prompt used in this work are available at https://github.com/dmis-lab/CoTox.

Jungwoo Park, Taewhoo Lee, Chanwoong Yoon et al.

Extreme activation outliers in Large Language Models (LLMs) critically degrade quantization performance, hindering efficient on-device deployment. While channel-wise operations and adaptive gradient scaling are recognized causes, practical mitigation remains challenging. We introduce Outlier-Safe Pre-Training (OSP), a practical guideline that proactively prevents outlier formation rather than relying on post-hoc mitigation. OSP combines three key innovations: (1) the Muon optimizer, eliminating privileged bases while maintaining training efficiency; (2) Single-Scale RMSNorm, preventing channel-wise amplification; and (3) a learnable embedding projection, redistributing activation magnitudes originating from embedding matrices. We validate OSP by training a 1.4B-parameter model on 1 trillion tokens, which is the first production-scale LLM trained without such outliers. Under aggressive 4-bit quantization, our OSP model achieves a 35.7 average score across 10 benchmarks (compared to 26.5 for an Adam-trained model), with only a 2% training overhead. Remarkably, OSP models exhibit near-zero excess kurtosis (0.04) compared to extreme values (1818.56) in standard models, fundamentally altering LLM quantization behavior. Our work demonstrates that outliers are not inherent to LLMs but are consequences of training strategies, paving the way for more efficient LLM deployment. The source code and pretrained checkpoints are available at https://github.com/dmis-lab/Outlier-Safe-Pre-Training.

Hyeon Hwang, Yewon Cho, Chanwoong Yoon et al.

Step-by-step reasoning has become a standard approach for large language models (LLMs) to tackle complex tasks. While this paradigm has proven effective, it raises a fundamental question: How can we verify that an LLM's reasoning is accurately grounded in knowledge? To address this question, we introduce a novel evaluation suite that systematically assesses the knowledge grounding of intermediate reasoning. Our framework comprises three key components. (1) Principal Knowledge Collection, a large-scale repository of atomic knowledge essential for reasoning. Based on the collection, we propose (2) knowledge-grounded evaluation metrics designed to measure how well models recall and apply prerequisite knowledge in reasoning. These metrics are computed by our (3) evaluator LLM, a lightweight model optimized for cost-effective and reliable metric computation. Our evaluation suite demonstrates remarkable effectiveness in identifying missing or misapplied knowledge elements, providing crucial insights for uncovering fundamental reasoning deficiencies in LLMs. Beyond evaluation, we demonstrate how these metrics can be integrated into preference optimization, showcasing further applications of knowledge-grounded evaluation.

Bumsoo Kim, Taeho Choi, Jaewoo Kang et al.

Recent advances in deep neural networks have achieved significant progress in detecting individual objects from an image. However, object detection is not sufficient to fully understand a visual scene. Towards a deeper visual understanding, the interactions between objects, especially humans and objects are essential. Most prior works have obtained this information with a bottom-up approach, where the objects are first detected and the interactions are predicted sequentially by pairing the objects. This is a major bottleneck in HOI detection inference time. To tackle this problem, we propose UnionDet, a one-stage meta-architecture for HOI detection powered by a novel union-level detector that eliminates this additional inference stage by directly capturing the region of interaction. Our one-stage detector for human-object interaction shows a significant reduction in interaction prediction time 4x~14x while outperforming state-of-the-art methods on two public datasets: V-COCO and HICO-DET.

Hajung Kim, Eunha Lee, Sohyun Chung et al.

Identifying metabolic sites where cytochrome P450 enzymes metabolize small-molecule drugs is essential for drug discovery. Although existing computational approaches have been proposed for site-of-metabolism prediction, they typically ignore cytochrome P450 isoform identity or model isoforms independently, thereby failing to fully capture inherent cross-isoform metabolic patterns. In addition, prior evaluations often rely on top-k metrics, where false positive atoms may be included among the top predictions, underscoring the need for complementary metrics that more directly assess binary atom-level discrimination under severe class imbalance. We propose ATTNSOM, an atom-level site-of-metabolism prediction framework that integrates intrinsic molecular reactivity with cross-isoform relationships. The model combines a shared graph encoder, molecule-conditioned atom representations, and a cross-attention mechanism to capture correlated metabolic patterns across cytochrome P450 isoforms. The model is evaluated on two benchmark datasets annotated with site-of-metabolism labels at atom resolution. Across these benchmarks, the model achieves consistently strong top-k performance across multiple cytochrome P450 isoforms. Relative to ablated variants, the model yields higher Matthews correlation coefficient, indicating improved discrimination of true metabolic sites. These results support the importance of explicitly modeling cross-isoform relationships for site-of-metabolism prediction. The code and datasets are available at https://github.com/dmis-lab/ATTNSOM.

Dain Kim, Jiwoo Lee, Jaehoon Yun et al.

Large Vision-Language Models (LVLMs) hold significant promise for medical applications, yet their deployment is often constrained by insufficient alignment and reliability. While Direct Preference Optimization (DPO) has emerged as a potent framework for refining model responses, its efficacy in high-stakes medical contexts remains underexplored, lacking the rigorous empirical groundwork necessary to guide future methodological advances. To bridge this gap, we present the first comprehensive examination of diverse DPO variants within the medical domain, evaluating nine distinct formulations across two medical LVLMs: LLaVA-Med and HuatuoGPT-Vision. Our results reveal several critical limitations: current DPO approaches often yield inconsistent gains over supervised fine-tuning, with their efficacy varying significantly across different tasks and backbones. Furthermore, they frequently fail to resolve fundamental visual misinterpretation errors. Building on these insights, we present a targeted preference construction strategy as a proof-of-concept that explicitly addresses visual misinterpretation errors frequently observed in existing DPO models. This design yields a 3.6% improvement over the strongest existing DPO baseline on visual question-answering tasks. To support future research, we release our complete framework, including all training data, model checkpoints, and our codebase at https://github.com/dmis-lab/med-vlm-dpo.

Taeyun Roh, Eun-yeong Jo, Wonjune Jang et al.

Scientific figure multiple-choice question answering (MCQA) requires models to reason over diverse visual evidence, ranging from charts and multipanel figures to microscopy and biomedical images. However, this setting suffers from a distinctive bias: answer choices themselves can act as priors, steering multimodal models toward scientifically plausible options even when the figure supports a different answer. We investigate this failure mode through a simple question: what if decoding explicitly discounts what the model would prefer from text alone, so as to favor figure-grounded evidence? To this end, we propose SCICON, a training-free decoding method that scores each candidate by subtracting a text-only option score from its image-conditioned counterpart. Unlike prior contrastive decoding approaches that mitigate hallucinations by contrasting original inputs with distorted images or perturbed instructions, SCICON directly targets the choice-induced prior encoded in candidate text. Across three scientific figure QA benchmarks and three model backbones, SCICON consistently improves accuracy over standard decoding baselines. These results show that decoding against choice-induced priors is an effective and simple way to improve figure-grounded reasoning in scientific MCQA.

Jinkyu Kim, Hyunjung Yi, Mogan Gim et al.

We propose DeepAries , a novel deep reinforcement learning framework for dynamic portfolio management that jointly optimizes the timing and allocation of rebalancing decisions. Unlike prior reinforcement learning methods that employ fixed rebalancing intervals regardless of market conditions, DeepAries adaptively selects optimal rebalancing intervals along with portfolio weights to reduce unnecessary transaction costs and maximize risk-adjusted returns. Our framework integrates a Transformer-based state encoder, which effectively captures complex long-term market dependencies, with Proximal Policy Optimization (PPO) to generate simultaneous discrete (rebalancing intervals) and continuous (asset allocations) actions. Extensive experiments on multiple real-world financial markets demonstrate that DeepAries significantly outperforms traditional fixed-frequency and full-rebalancing strategies in terms of risk-adjusted returns, transaction costs, and drawdowns. Additionally, we provide a live demo of DeepAries at https://deep-aries.github.io/, along with the source code and dataset at https://github.com/dmis-lab/DeepAries, illustrating DeepAries' capability to produce interpretable rebalancing and allocation decisions aligned with shifting market regimes. Overall, DeepAries introduces an innovative paradigm for adaptive and practical portfolio management by integrating both timing and allocation into a unified decision-making process.

Seungheun Baek, Soyon Park, Yan Ting Chok et al.

Motivation: Predicting cellular responses to genetic perturbations is essential for understanding biological systems and developing targeted therapeutic strategies. While variational autoencoders (VAEs) have shown promise in modeling perturbation responses, their limited explainability poses a significant challenge, as the learned features often lack clear biological meaning. Nevertheless, model explainability is one of the most important aspects in the realm of biological AI. One of the most effective ways to achieve explainability is incorporating the concept of gene regulatory networks (GRNs) in designing deep learning models such as VAEs. GRNs elicit the underlying causal relationships between genes and are capable of explaining the transcriptional responses caused by genetic perturbation treatments. Results: We propose GPO-VAE, an explainable VAE enhanced by GRN-aligned Parameter Optimization that explicitly models gene regulatory networks in the latent space. Our key approach is to optimize the learnable parameters related to latent perturbation effects towards GRN-aligned explainability. Experimental results on perturbation prediction show our model achieves state-of-the-art performance in predicting transcriptional responses across multiple benchmark datasets. Furthermore, additional results on evaluating the GRN inference task reveal our model's ability to generate meaningful GRNs compared to other methods. According to qualitative analysis, GPO-VAE posseses the ability to construct biologically explainable GRNs that align with experimentally validated regulatory pathways. GPO-VAE is available at https://github.com/dmis-lab/GPO-VAE

Hoonick Lee, Mogan Gim, Donghyeon Park et al.

The advent of Large Language Models (LLMs) have shown promise in various creative domains, including culinary arts. However, many LLMs still struggle to deliver the desired level of culinary creativity, especially when tasked with adapting recipes to meet specific cultural requirements. This study focuses on cuisine transfer-applying elements of one cuisine to another-to assess LLMs' culinary creativity. We employ a diverse set of LLMs to generate and evaluate culturally adapted recipes, comparing their evaluations against LLM and human judgments. We introduce the ASH (authenticity, sensitivity, harmony) benchmark to evaluate LLMs' recipe generation abilities in the cuisine transfer task, assessing their cultural accuracy and creativity in the culinary domain. Our findings reveal crucial insights into both generative and evaluative capabilities of LLMs in the culinary domain, highlighting strengths and limitations in understanding and applying cultural nuances in recipe creation. The code and dataset used in this project will be openly available in \url{http://github.com/dmis-lab/CulinaryASH}.

Mogan Gim, Jueon Park, Soyon Park et al.

Molecular core structures and R-groups are essential concepts in drug development. Integration of these concepts with conventional graph pre-training approaches can promote deeper understanding in molecules. We propose MolPLA, a novel pre-training framework that employs masked graph contrastive learning in understanding the underlying decomposable parts inmolecules that implicate their core structure and peripheral R-groups. Furthermore, we formulate an additional framework that grants MolPLA the ability to help chemists find replaceable R-groups in lead optimization scenarios. Experimental results on molecular property prediction show that MolPLA exhibits predictability comparable to current state-of-the-art models. Qualitative analysis implicate that MolPLA is capable of distinguishing core and R-group sub-structures, identifying decomposable regions in molecules and contributing to lead optimization scenarios by rationally suggesting R-group replacements given various query core templates. The code implementation for MolPLA and its pre-trained model checkpoint is available at https://github.com/dmis-lab/MolPLA

Seongjun Yun, Minbyul Jeong, Sungdong Yoo et al.

Graph Neural Networks (GNNs) have been widely applied to various fields due to their powerful representations of graph-structured data. Despite the success of GNNs, most existing GNNs are designed to learn node representations on the fixed and homogeneous graphs. The limitations especially become problematic when learning representations on a misspecified graph or a heterogeneous graph that consists of various types of nodes and edges. To address this limitations, we propose Graph Transformer Networks (GTNs) that are capable of generating new graph structures, which preclude noisy connections and include useful connections (e.g., meta-paths) for tasks, while learning effective node representations on the new graphs in an end-to-end fashion. We further propose enhanced version of GTNs, Fast Graph Transformer Networks (FastGTNs), that improve scalability of graph transformations. Compared to GTNs, FastGTNs are 230x faster and use 100x less memory while allowing the identical graph transformations as GTNs. In addition, we extend graph transformations to the semantic proximity of nodes allowing non-local operations beyond meta-paths. Extensive experiments on both homogeneous graphs and heterogeneous graphs show that GTNs and FastGTNs with non-local operations achieve the state-of-the-art performance for node classification tasks. The code is available: https://github.com/seongjunyun/Graph_Transformer_Networks

Wonjin Yoon, Richard Jackson, Aron Lagerberg et al.

Current studies in extractive question answering (EQA) have modeled the single-span extraction setting, where a single answer span is a label to predict for a given question-passage pair. This setting is natural for general domain EQA as the majority of the questions in the general domain can be answered with a single span. Following general domain EQA models, current biomedical EQA (BioEQA) models utilize the single-span extraction setting with post-processing steps. In this article, we investigate the question distribution across the general and biomedical domains and discover biomedical questions are more likely to require list-type answers (multiple answers) than factoid-type answers (single answer). This necessitates the models capable of producing multiple answers for a question. Based on this preliminary study, we propose a sequence tagging approach for BioEQA, which is a multi-span extraction setting. Our approach directly tackles questions with a variable number of phrases as their answer and can learn to decide the number of answers for a question from training data. Our experimental results on the BioASQ 7b and 8b list-type questions outperformed the best-performing existing models without requiring post-processing steps. Source codes and resources are freely available for download at https://github.com/dmis-lab/SeqTagQA

Minbyul Jeong, Jaewoo Kang

When performing named entity recognition (NER), entity length is variable and dependent on a specific domain or dataset. Pre-trained language models (PLMs) are used to solve NER tasks and tend to be biased toward dataset patterns such as length statistics, surface form, and skewed class distribution. These biases hinder the generalization ability of PLMs, which is necessary to address many unseen mentions in real-world situations. We propose a novel debiasing method RegLER to improve predictions for entities of varying lengths. To close the gap between evaluation and real-world situations, we evaluated PLMs on partitioned benchmark datasets containing unseen mention sets. Here, RegLER shows significant improvement over long-named entities that can predict through debiasing on conjunction or special characters within entities. Furthermore, there is a severe class imbalance in most NER datasets, causing easy-negative examples to dominate during training, such as "The". Our approach alleviates skewed class distribution by reducing the influence of easy-negative examples. Extensive experiments on the biomedical and general domains demonstrated the generalization capabilities of our method. To facilitate reproducibility and future work, we release our code."https://github.com/minstar/RegLER"

Jinhyuk Lee, Wonjin Yoon, Sungdong Kim et al.

Biomedical text mining is becoming increasingly important as the number of biomedical documents rapidly grows. With the progress in natural language processing (NLP), extracting valuable information from biomedical literature has gained popularity among researchers, and deep learning has boosted the development of effective biomedical text mining models. However, directly applying the advancements in NLP to biomedical text mining often yields unsatisfactory results due to a word distribution shift from general domain corpora to biomedical corpora. In this article, we investigate how the recently introduced pre-trained language model BERT can be adapted for biomedical corpora. We introduce BioBERT (Bidirectional Encoder Representations from Transformers for Biomedical Text Mining), which is a domain-specific language representation model pre-trained on large-scale biomedical corpora. With almost the same architecture across tasks, BioBERT largely outperforms BERT and previous state-of-the-art models in a variety of biomedical text mining tasks when pre-trained on biomedical corpora. While BERT obtains performance comparable to that of previous state-of-the-art models, BioBERT significantly outperforms them on the following three representative biomedical text mining tasks: biomedical named entity recognition (0.62% F1 score improvement), biomedical relation extraction (2.80% F1 score improvement) and biomedical question answering (12.24% MRR improvement). Our analysis results show that pre-training BERT on biomedical corpora helps it to understand complex biomedical texts. We make the pre-trained weights of BioBERT freely available at https://github.com/naver/biobert-pretrained, and the source code for fine-tuning BioBERT available at https://github.com/dmis-lab/biobert.

Yonggyu Park, Junhyun Lee, Yookyung Koh et al.

The mood of a text and the intention of the writer can be reflected in the typeface. However, in designing a typeface, it is difficult to keep the style of various characters consistent, especially for languages with lots of morphological variations such as Chinese. In this paper, we propose a Typeface Completion Network (TCN) which takes one character as an input, and automatically completes the entire set of characters in the same style as the input characters. Unlike existing models proposed for image-to-image translation, TCN embeds a character image into two separate vectors representing typeface and content. Combined with a reconstruction loss from the latent space, and with other various losses, TCN overcomes the inherent difficulty in designing a typeface. Also, compared to previous image-to-image translation models, TCN generates high quality character images of the same typeface with a much smaller number of model parameters. We validate our proposed model on the Chinese and English character datasets, which is paired data, and the CelebA dataset, which is unpaired data. In these datasets, TCN outperforms recently proposed state-of-the-art models for image-to-image translation. The source code of our model is available at https://github.com/yongqyu/TCN.

Hajung Kim, Jueon Park, Junseok Choe et al.

Quantitative structure-activity relationship assumes a smooth relationship between molecular structure and biological activity. However, activity cliffs defined as pairs of structurally similar compounds with large potency differences break this continuity. Recent benchmarks targeting activity cliffs have revealed that classical machine learning models with extended connectivity fingerprints outperform graph neural networks. Our analysis shows that graph embeddings fail to adequately separate structurally similar molecules in the embedding space, making it difficult to distinguish between structurally similar but functionally different molecules. Despite this limitation, molecular graph structures are inherently expressive and attractive, as they preserve molecular topology. To preserve the structural representation of molecules as graphs, we propose a new model, GraphCliff, which integrates short- and long-range information through a gating mechanism. Experimental results demonstrate that GraphCliff consistently improves performance on both non-cliff and cliff compounds. Furthermore, layer-wise node embedding analyses reveal reduced over-smoothing and enhanced discriminative power relative to strong baseline graph models.

Jaehyo Yoo, Jaewoo Kang

Many machine translation models are trained on bilingual corpus, which consist of aligned sentence pairs from two different languages with same semantic. However, there is a qualitative discrepancy between train and test set in bilingual corpus. While the most train sentences are created via automatic techniques such as crawling and sentence-alignment methods, the test sentences are annotated with the consideration of fluency by human. We suppose this discrepancy in training corpus will yield performance drop of translation model. In this work, we define \textit{fluency noise} to determine which parts of train sentences cause them to seem unnatural. We show that \textit{fluency noise} can be detected by simple gradient-based method with pre-trained classifier. By removing \textit{fluency noise} in train sentences, our final model outperforms the baseline on WMT-14 DE$\rightarrow$EN and RU$\rightarrow$EN. We also show the compatibility with back-translation augmentation, which has been commonly used to improve the fluency of the translation model. At last, the qualitative analysis of \textit{fluency noise} provides the insight of what points we should focus on.

Taeyun Roh, Wonjune Jang, Junha Jung et al.

Large language model agents heavily rely on external memory to support knowledge reuse and complex reasoning tasks. Yet most memory systems store experiences in a single global retrieval pool which can gradually dilute or corrupt stored knowledge. This problem is especially pronounced for small language models (SLMs), which are highly vulnerable to irrelevant context. We introduce CLAG, a CLustering-based AGentic memory framework where an SLM agent actively organizes memory by clustering. CLAG employs an SLM-driven router to assign incoming memories to semantically coherent clusters and autonomously generates cluster-specific profiles, including topic summaries and descriptive tags, to establish each cluster as a self-contained functional unit. By performing localized evolution within these structured neighborhoods, CLAG effectively reduces cross-topic interference and enhances internal memory density. During retrieval, the framework utilizes a two-stage process that first filters relevant clusters via their profiles, thereby excluding distractors and reducing the search space. Experiments on multiple QA datasets with three SLM backbones show that CLAG consistently improves answer quality and robustness over prior memory systems for agents, remaining lightweight and efficient.

Hyeon Hwang, Jiwoo Lee, Jaewoo Kang

Tool-integrated reasoning (TIR) has emerged as a dominant paradigm for mathematical problem solving in language models, combining natural language (NL) reasoning with code execution. However, this interleaved setup has three key limitations: code often acts as a post-hoc verifier, intermediate NL computations are error-prone, and NL and code play overlapping rather than clearly distinct roles. We propose ThinC (Thinking in Code), a framework in which code itself serves as the reasoner rather than as a tool invoked by NL. A ThinC trajectory begins with a brief NL planning step, after which all reasoning unfolds through code blocks connected only by their execution outputs. We distill 12.2k code-centric trajectories from a teacher model and train ThinC-1.7B and ThinC-4B with supervised fine-tuning followed by reinforcement learning. ThinC-4B consistently outperforms every TIR baseline on five competition-level math benchmarks and even surpasses the much larger Qwen3-235B-A22B-Thinking. Further analysis shows that ThinC reasons through code: 99.2% of its final answers are grounded in interpreter output, and the model recovers reliably from code execution failures without intermediate NL reasoning. Our code and models will be released soon.

Minbyul Jeong, Jiwoong Sohn, Mujeen Sung et al.

Recent proprietary large language models (LLMs), such as GPT-4, have achieved a milestone in tackling diverse challenges in the biomedical domain, ranging from multiple-choice questions to long-form generations. To address challenges that still cannot be handled with the encoded knowledge of LLMs, various retrieval-augmented generation (RAG) methods have been developed by searching documents from the knowledge corpus and appending them unconditionally or selectively to the input of LLMs for generation. However, when applying existing methods to different domain-specific problems, poor generalization becomes apparent, leading to fetching incorrect documents or making inaccurate judgments. In this paper, we introduce Self-BioRAG, a framework reliable for biomedical text that specializes in generating explanations, retrieving domain-specific documents, and self-reflecting generated responses. We utilize 84k filtered biomedical instruction sets to train Self-BioRAG that can assess its generated explanations with customized reflective tokens. Our work proves that domain-specific components, such as a retriever, domain-related document corpus, and instruction sets are necessary for adhering to domain-related instructions. Using three major medical question-answering benchmark datasets, experimental results of Self-BioRAG demonstrate significant performance gains by achieving a 7.2% absolute improvement on average over the state-of-the-art open-foundation model with a parameter size of 7B or less. Overall, we analyze that Self-BioRAG finds the clues in the question, retrieves relevant documents if needed, and understands how to answer with information from retrieved documents and encoded knowledge as a medical expert does. We release our data and code for training our framework components and model weights (7B and 13B) to enhance capabilities in biomedical and clinical domains.