New AI Architecture Cuts Model Size through Expert Sharing

Abstract Sparse Mixture of Experts (SMoE) architectures reduce computational cost by activating only a subset of experts per token, yet they often retain large memory footprints and exhibit significant redundancy, both within and across layers. We propose GMoE, a sparse MoE architecture designed to explicitly address these inefficiencies. Instead of maintaining separate expert sets for each layer, GMoE uses Global Experts shared across all layers and adds a Local Expert per layer for layer-specific adaptation. This architecture reuses Global Experts across layers, thereby mitigating inter-layer redundancy while substantially reducing model parameters. In addition, we introduce a Global Router with a GRU-based recurrent component shared across layers and layer-specific routing heads that propagate routing logits across layers. This routing mechanism couples routing decisions across layers, progressively refines routing paths, and helps mitigate intra-layer redundancy. Across diverse language modeling corpora and downstream benchmarks, GMoE remains competitive while using substantially fewer parameters. Routing path analyzes and an ablation study show that GMoE reduces cross-layer routing concentration and increases path diversity, with the Global Experts, the Local Expert, and the Global Router all contributing to the gains. Mixture-of-Experts (MoE) models improve AI efficiency by activating only a small number of specialized components, known as experts , for each input. But as these models grow deeper, similar experts are often repeated across layers, increasing model size and memory without adding much new capability. A research team led by Professor Taehwan Kim of the Graduate School of Artificial Intelligence at UNIST has developed GMoE (Global Mixture of Experts), a new architecture that allows layers to share experts instead. The approach substantially reduces model parameters while maintaining comparable language performance. GMoE replaces separate expert sets at each layer with Global Experts shared throughout the model, while retaining one Local Expert per layer for specialized functions. This allows useful expertise to be reused rather than repeatedly stored across the network. The team also improved how the model selects experts. Conventional MoE models make these decisions independently at each layer, which can repeatedly direct inputs through the same expert combinations. GMoE carries information from earlier selections forward, allowing subsequent layers to make more varied choices. Tests across seven language tasks, including reading comprehension, grammatical judgment, and commonsense reasoning, showed comparable performance with substantially fewer parameters. In a medium-scale model, GMoE reduced the parameter count by about 63%, from 549 million to 204 million, while average accuracy remained nearly unchanged at 39.51%, compared with 39.55% for the conventional model. GMoE also produced 81,561 distinct expert paths, more than three times as many as the conventional approach. The share of inputs following the most common path fell to 11.15%, compared with 25.65% to 45.55% in the baseline models. “Conventional MoE models can repeat similar functions across layers while relying heavily on a limited number of expert combinations,” said Professor Kim. “GMoE addresses both problems by sharing experts across layers and connecting how those experts are selected.” He added, “The results show that language models can maintain competitive performance with far fewer parameters while using a broader range of expert combinations.” Geonwoo Hong, a master's student in the Graduate School of Artificial Intelligence at UNIST, served as first author. The study was accepted to the Annual Meeting of the Association for Computational Linguistics (ACL 2026), a leading international conference in natural language processing. The research code is publicly available through the team's project repository. Journal Reference Geonwoo Hong and Taehwan Kim, "GMoE: Global Mixture of Experts with Logit Propagation," ACL '26, (2026).

2026.08.24

New AI Learns How Shopping Preferences Change Across Products

Abstract Cross-domain recommendation (CDR) aims to alleviate data sparsity by transferring knowledge across domains, yet existing methods primarily rely on coarse-grained behavioral signals and often overlook intra-domain heterogeneity in user preferences. We propose Multi-TAP, a multi-criteria target-adaptive persona framework that explicitly captures such heterogeneity through semantic persona modeling. To enable effective transfer, Multi-TAP selectively incorporates source-domain signals conditioned on the target domain, preserving relevance during knowledge transfer. Experiments on real-world datasets demonstrate that Multi-TAP consistently outperforms state-of-the-art CDR methods, highlighting the importance of modeling intra-domain heterogeneity for robust cross-domain recommendation. Our preferences are more fluid than most recommendation systems assume. The qualities that matter when choosing one product may carry far less weight when choosing another, even for the same person. A research team, led by Professor Yeon-Chang Lee of the Graduate School of Artificial Intelligence at UNIST has developed Multi-TAP, an AI framework that recognizes these differences across product categories and uses them to make more personalized recommendations. Recommendation systems often have plenty of information about a user in one area but relatively little in another. Cross-domain recommendation tries to fill that gap by drawing on a user's activity elsewhere. But existing methods can miss an important distinction: preferences that matter in one product category may not carry over neatly to another. Multi-TAP accounts for these differences by building multiple semantic personas from a user's purchase history and product information. The system looks at factors such as price, ratings, review counts, purchasing frequency, and the range of products a person buys, allowing it to distinguish how those priorities change across categories. A large language model (LLM) helps turn these patterns into representations that retain those category-specific differences. Multi-TAP then combines them with purchasing behavior learned by the recommendation model, while determining how much information from one domain is actually useful in another. This allows the system to draw selectively on past behavior rather than treating everything known about a user as equally relevant. The researchers tested Multi-TAP using Amazon purchase data across electronics, home and household products, sporting goods, clothing, and toys. It achieved the best performance in five of six cross-domain recommendation tasks. In its strongest result, the rate at which a product a user actually purchased appeared among the top five recommendations was up to 36.3% higher than with the previous best-performing method. “Consumer preferences can vary depending on the context and the type of product,” said Professor Lee. “Multi-TAP represents these differences through multiple personas and uses information from other domains according to its relevance. This could help recommendation systems better reflect individual preferences, particularly in areas where a user has little purchase history.” Daehee Kang served as first author of the study. Their work was accepted to the 2026 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), held in Jeju Island, South Korea from August 9-13, 2026. The research was supported by the Ministry of Science and ICT (MSIT), the Institute of Information & Communications Technology Planning & Evaluation (IITP), and the National Research Foundation of Korea (NRF). Journal Reference Daehee Kang and Yeon-Chang Lee, “Multi-TAP: Multi-criteria Target Adaptive Persona Modeling for Cross-Domain Recommendation,” ACM KDD' 26 (2026).

2026.08.20

Study Reveals Why Perovskite Performance Changes Across Solar Cell Designs

Abstract A perovskite precursor solution that delivers power conversion efficiencies (PCEs) exceeding 26% in conventional nip solar cells exhibits severe performance losses when directly applied to inverted pin architectures, revealing that high-efficiency compositions are not inherently transferable. Here, we identify a buried-interface crystallization mismatch, arising from the distinct physicochemical natures of inorganic SnO2 electron-transporting layers and organic self-assembled hole-transporting monolayers (SA-HTLs), as the origin of this divergence. The methylammonium chloride (MACl)-associated intermediate phase, MA2Pb3I8·2DMSO, persists and decomposes with strong underlayer dependence, stabilizing beneficially on SnO2 but impeding crystallization on SA-HTLs. To overcome this limitation, we develop a chloride-origin engineering strategy that decouples chloride functionality from volatile organic ammonium species by incorporating low-solubility lead chloride (PbCl2) with strong Pb–Cl coordination. This enables controlled interfacial desolvation and nucleation on SA-HTLs, suppresses buried defects, and establishes buried-interface crystallization control as a design principle for architecture-convergent, high-efficiency perovskite solar cells. Building a high-performance solar cell takes more than getting the materials right. The same ingredients can produce very different results when the layers are arranged differently—and researchers at UNIST have discovered why. Led by Distinguished Professor Sang Il Seok of the School of Energy and Chemical Engineering at UNIST, the team found that the surface beneath the light-absorbing perovskite layer can determine how well its crystals form. By tailoring the material chemistry to that surface, the researchers developed an inverted perovskite solar cell with a power conversion efficiency of 26.3% and improved stability. PSCs can be built in conventional nip or inverted pin configurations. Both have reached high efficiencies, but formulations optimized for one do not necessarily perform as well in the other. Inverted cells are particularly important for perovskite–silicon tandem solar cells, where they are commonly used as the top cell. The researchers found that this performance gap originates at the buried interface, where the perovskite layer forms on the material beneath it. A formulation that achieved efficiencies above 26% in conventional cells lost much of its performance when transferred directly to an inverted design. The difference came down in part to methylammonium chloride (MACl), an additive used to improve perovskite crystal growth. In conventional cells, an intermediate formed by MACl supports crystallization on the underlying inorganic layer. But on the organic layer used in inverted cells, the intermediate persists longer, disrupting crystal growth and leaving behind voids and defects that can lead to electrical losses. Rather than removing chloride, the team changed how it was introduced. The researchers reduced the amount of MACl and added a small amount of lead chloride (PbCl2), whose stronger interaction with chloride helped control where and when crystals began to form. This produced a cleaner buried interface with fewer defects. With the revised formulation, the inverted cells reached a peak efficiency of 26.3% and a fill factor of up to 86.8%, while also showing improved stability. The results demonstrate that controlling how perovskite crystals form at the underlying surface can help high-performance materials work across different solar-cell designs. “High-efficiency formulations developed for conventional PSCs do not necessarily behave the same way in inverted devices,” said Professor Seok. “By identifying the origin of this performance loss, we established a way to control crystal formation at the buried interface. The findings provide a design direction for developing efficient inverted cells for high-performance perovskite–silicon tandem solar cells.” Their findings have been published online in Joule on July 28, 2026. The study was supported by the National Research Foundation of Korea (NRF). Journal Reference Jongbeom Kim, Nahye Shin, Chaehoon Jeon, et al ., “Buried-interface crystallization limits the transferability of high-efficiency perovskite precursor compositions,” Joule ,(2026).

2026.08.19

New Study Reveals How Nearby Fat Cells Help Breast Cancer Resist Treatment

Abstract Adipocytes are essential stromal components of the tumor microenvironment (TME) in breast cancer that play pivotal roles in cancer progression and chemoresistance. In close proximity to tumor cells, they undergo phenotypic reprogramming into cancer-associated adipocytes (CAAs), characterized by multilocular lipid droplets, increased mitochondrial content, and elevated expression of uncoupling protein 1 (UCP1). Although these features superficially resemble those of beige adipocytes, they do not recapitulate classical thermogenic programming, reflecting a unique metabolic adaptation driven by the TME. Here, we identified tumor necrosis factor receptor-associated protein 1 (TRAP1), a mitochondrial paralog of HSP90, as a central regulator of the transition of adipocytes into CAAs. TRAP1 was highly upregulated in CAAs and was required to drive a tumor-associated adipocyte secretory program, including the adipokine complement factor D (CFD). Genetic and pharmacological TRAP1 inhibition destabilized the mitochondrial electron transport chain, reduced cellular respiration, and activated the energy sensor AMPK. This subsequently suppressed mTOR and PPARγ signaling, effectively abrogating adipocyte reprogramming and diminishing pro-tumorigenic adipokine secretion. Crucially, this CAA-secreted CFD promoted cancer cell survival and chemoresistance via C3aR-AKT/ERK signaling, and blocking this TRAP1-mediated crosstalk profoundly sensitized breast tumors to chemotherapy in vivo. Collectively, these findings identify TRAP1 as a master regulator of adipocyte transdifferentiation within the TME, offering a novel strategy to restrict tumor growth and overcome drug resistance in breast cancer. Breast tumors can reshape the fat cells around them, creating an environment that helps cancer cells survive and resist chemotherapy. Researchers at UNIST and the National Cancer Center (NCC) have identified a mitochondrial protein that helps drive this transformation—and shown in mice that blocking it can make tumors more responsive to treatment. Led by Professor Byoung Heon Kang of the Department of Biological Sciences at UNIST and Professor Sun-Young Kong of NCC, the team found that TRAP1 plays a key role in converting normal fat cells into cancer-associated adipocytes (CAAs). These altered cell release factors that support tumor growth and protect cancer cells from chemotherapy. The researchers found that TRAP1 sustains mitochondrial energy production in fat cells, keeping the mTOR–PPARγ pathway active and allowing them to acquire cancer-supporting properties. Blocking TRAP1 disrupts mitochondrial function and activates AMPK, an energy sensor that suppresses this transformation. In tissue samples from 31 patients with breast cancer, TRAP1 expression in fat surrounding tumors was 3.03 times higher than in normal adipose tissue. The team also identified complement factor D (CFD) as an important link between the altered fat cells and cancer cells. CAAs secrete CFD, which activates survival signals in breast cancer cells and makes them less susceptible to chemotherapy. Blocking this TRAP1-driven communication reduced that protective effect. In mice, genetically removing TRAP1 suppressed the transformation of adipocytes and reduced tumor weight by as much as 56% compared with controls. Drug-based inhibition produced similar results: combining the TRAP1 inhibitor gamitrinib with cisplatin reduced tumor weight by 76% compared with untreated mice. Another inhibitor, SB-U015, also enhanced the effects of cisplatin and paclitaxel without evident toxicity in the experimental models. “Our study shows how adipocytes surrounding breast tumors can contribute to chemotherapy resistance,” said Professor Kong. “Further studies are needed to evaluate the clinical potential of TRAP1 inhibitors, not only for patients who respond poorly to existing treatments but also for cancers that develop in close contact with adipose tissue, including ovarian, pancreatic, and prostate cancers.” Professor Kang added, “These findings suggest that the cells surrounding a tumor can be important therapeutic targets alongside the cancer cells themselves. By disrupting the metabolic changes in nearby adipocytes, we may be able to improve the effectiveness of existing anticancer drugs.” Professor Kang has transferred the technology arising from the research to SmartinBio, a UNIST faculty startup, where further development of anticancer drug candidates is underway. The study was co-first authored by Dr. Nam Gu Yoon and Dr. So-Yeon Kim of UNIST and published online in Signal Transduction and Targeted Therapy (STTT) on July 28, 2026. The research was supported by the National Research Foundation of Korea (NRF), the National Cancer Center, and the Korea Drug Development Fund (KDDF). Journal Reference Nam Gu Yoon, So-Yeon Kim, So-Youn Jung, et al ., “The pro-tumorigenic functions of cancer-associated adipocytes are dependent on the mitochondrial chaperone tumor necrosis factor receptor-associated protein 1,” STTT., (2026).

2026.08.18

New Framework Reduces Data Needed to Train Re-Identification AI

Abstract Coreset Selection (CS) aims to extract a small yet representative subset from a large dataset, reducing the complexity of model training. Although CS has been primarily investigated for classification tasks, it is still underexplored for object Re-identification (ReID). In this paper, we first formulate Coreset Selection for Object Re-identification (CSOR) as a joint optimization problem to find both the optimal coreset and the optimal class subset. We identify intra-class diversity as a key factor for effective coreset construction for ReID. Based on this insight, we propose a novel two-stage framework, consisting of Diversity-driven Class Pruning (DCP) and Coverage-Prioritized Sampling (CPS), to address the unique challenges of ReID datasets. First, classes with low feature diversity are pruned to allocate the storage budget to the remaining informative classes. Then, samples are greedily selected in an easy-to-hard class order to maximize feature coverage within each class. Extensive experiments on three person ReID datasets and one vehicle ReID dataset demonstrate that our method consistently outperforms existing CS approaches, establishing a new state-of-the-art in CSOR. Video datasets used to train re-identification systems often contain near-duplicate images. The same person or vehicle may appear across many consecutive frames with little change, increasing storage and computing demands without adding equally useful information. Led by Professor Jae-Young Sim in the UNIST Graduate School of Artificial Intelligence, the research team has developed a method that filters out much of this redundancy, while preserving the visual diversity needed for training. The method, called Coreset Selection for Object Re-identification (CSOR), selects a compact set of representative images from a much larger dataset. Object re-identification (ReID) enables AI systems to match the same person or vehicle across different camera views. Unlike image classification, ReID models must distinguish identities they have never encountered during training. This makes conventional coreset methods—which were developed largely for classification—difficult to apply directly. “Re-identification models must distinguish entirely new people or vehicles that were not included in the training data,” said first author Minyoung Oh, a researcher at UNIST. “We began with this distinction and redesigned the selection criteria around the specific demands of ReID.” The researchers formulated CSOR as a joint optimization problem that determines both which identity classes and which images to retain. Their analysis identified diversity within each class—the range of appearances captured for the same person or vehicle—as a key factor in building an effective training subset. CSOR follows a two-stage process. Diversity-driven Class Pruning (DCP) first removes identity classes with little feature variation, preserving more of the storage budget for classes with richer information. Coverage-Prioritized Sampling (CPS) then selects images from the remaining classes, giving priority to those that capture the broader range of features. The team evaluated CSOR on three person and one vehicle ReID datasets. The method consistently outperformed the CS approaches used for comparison. Models trained on approximately half of the original data retained more than 95% of the performance achieved using the full dataset. “ReID datasets are commonly built from video, so duplicate images accumulate quickly as the same person or vehicle appears across consecutive frames,” said Professor Sim. “CSOR removes much of this repetition while retaining the diversity needed for ReID. It could make training and deployment more practical on edge devices with limited storage and computing resources.” Their findings were accepted for presentation at the 2026 International Conference on Machine Learning (ICML), held at COEX in Seoul in July 2026. The study was supported by the National Research Foundation of Korea (NRF) and the Institute for Information & Communications Technology Planning & Evaluation (IITP), with funding from the Ministry of Science and ICT (MSIT). Support was provided through the Mid-Career Researcher Program, the AI Graduate School Program, the AI Star Fellowship Program, and an initiative for training in industrially integrated multimodal generative AI. Journal Reference Minyoung Oh and Jae-Young Sim, "CSOR: Coreset Selection for Object Re-identification via Class Pruning," ICML '26 , (2026).

2026.08.14

Cold-Sensitive Epigenetic Pathway Regulates Heat Production in Fat

Abstract Adipose tissue thermogenesis is a major determinant of energy homeostasis, and its dysregulation contributes to obesity and metabolic disease. Parkin-mediated mitophagy is required for thermogenic adaptation, but the upstream mechanisms linking thermal cues to this pathway remain poorly defined. Here, we identify the ten-eleven translocation (TET) family of DNA dioxygenases as thermosensitive epigenetic regulators of Prkn transcription in adipocytes. Cold exposure coordinately suppressed TET expression and reduced global 5-hydroxymethylcytosine (5hmC) levels in white and brown adipose tissue through β-adrenergic signaling. Adipose-specific TET triple-knockout mice exhibited enhanced white fat begging, brown fat activation, increased energy expenditure, and improved cold tolerance. Transcriptomic network analysis identified Parkin as a key mitophagy node in TET-deficient adipose tissue. Consistent with this, loss of adipose TET reduced Parkin expression, impaired mitophagic flux, and promoted accumulation of metabolically active mitochondria with increased respiratory capacity. Mechanistically, TET proteins occupied the Prkn promoter and maintained a transcriptionally permissive state through catalytic conversion of 5-methylcytosine to 5hmC, whereas TET loss increased promoter methylation and suppressed Prkn expression. Re-expression of wild-type, but not catalytically inactive, Parkin largely normalized mitochondrial content and respiratory activity in TET-deficient adipocytes. Together, these findings define a thermosensitive TET-Parkin epigenetic axis that links environmental cold signals to mitochondrial quality control during adaptive thermogenesis. Cold exposure prompts fat cells to burn more energy and generate heat. Researchers at UNIST have identified a molecular pathway that helps these cells retain more mitochondria, sustaining their heat-producing capacity. Led by Professor Myunggon Ko of the Department of Biological Sciences at UNIST, the team found that TET proteins, a family of enzymes that regulate DNA modifications, control the expression of Parkin, a key protein involved in mitochondrial removal. When cold reduces TET levels in fat cells, Parkin levels also fall, allowing more metabolically active mitochondria to remain. Published in the July 2026 issue of Metabolism , their findings reveal a previously unknown link between environmental temperature and mitochondrial quality control in adipose tissue. White fat primarily stores excess energy, while brown fat burns energy to produce heat. Under cold conditions, some white fat cells can acquire heat-producing properties, becoming beige fat. This transition is accompanied by an increase in mitochondria, which supply the energy needed for heat production. The researchers found that cold suppresses TET expression in both white and brown adipose tissue through β-adrenergic signaling. Levels of 5-hydroxymethylcytosine (5hmC)—a DNA modification associated with TET activity—also decline. TET proteins normally act at the Prkn gene promoter, where they help maintain Parkin expression by converting 5-methylcytosine to 5hmC. When TET levels fall, methylation increases at the promoter and Parkin expression declines. Parkin plays a central role in mitophagy, the process cells use to selectively remove mitochondria. With less Parkin available, mitophagy slows, allowing more mitochondria to accumulate in fat cells. The effect was evident in mice lacking all three TET proteins specifically in adipose tissue. Compared with control mice, they showed increased becoming of white fat, greater brown-fat activation, higher energy expenditure, and better tolerance to cold. Their adipose tissue also contained less Parkin, along with increased mitochondrial DNA and proteins involved in cellular respiration. Cell-based experiments further confirmed Parkin's role in the pathway. Restoring functional Parkin in TET-deficient adipocytes largely returned mitochondrial abundance and oxygen consumption to normal levels, while a catalytically inactive form of Parkin did not. This showed that Parkin is a key link between TET activity and mitochondrial turnover. “Cold-induced increases in mitochondria and thermogenesis in adipose tissue have been well documented, but the upstream mechanism controlling these changes has remained unclear,” said Professor Ko. “Our findings identify TET proteins as part of that regulatory system and reveal how changes in Parkin-mediated mitochondrial turnover contribute to the thermogenic response.” Professor Ko added that understanding this pathway could inform future research into metabolic diseases such as obesity and type 2 diabetes, where increasing energy expenditure in adipose tissue has been explored as a potential therapeutic strategy. The study was supported by the Ministry of Science and ICT (MSIT) and the National Research Foundation of Korea (NRF), among other funding sources. Journal Reference Seongjun Byun, Chan Hyeong Lee, Kyumin Jang, et al., “A thermosensitive TET–Parkin epigenetic axis couples mitochondrial quality control to adaptive thermogenesis in adipose tissue,” Metabolism, (2026).

2026.08.12