AI Personalization Adapts Answers to What Users Know

Abstract Large language models (LLMs) are increasingly used by end users, yet existing personalization methods relying on static profiles or text-only signals fail to capture query-specific expertise variation. We present ExPerT, a query-wise personalization framework that adapts LLM responses to users' query domain expertise by combining semantic and behavioral cues. ExPerT consists of two key components: (i) a semantic-behavioral expertise inference module that jointly interprets query text and keystroke dynamics via in-context LLM prompting, and (ii) an expertise-conditioned response generation that adapts the level of detail, terminology, and conceptual complexity. Our user study with 40 participants and 1270 queries demonstrated that ExPerT reduced expertise inference error by 65.7% compared to the strongest baseline (MAE = 0.398 vs. 1.162) and improved response satisfaction by 17.52% (from 3.71 to 4.36) on a 5-point Likert scale. A person's expertise is rarely the same across every subject. Someone deeply familiar with one field may know little about another, and even within the same field, expertise can vary from topic to topic. Yet, AI systems often personalize responses using relatively fixed information, such as user profiles or previous conversations. Led by Professor Taesik Gong of the Department of Computer Science and Engineering at UNIST, researchers developed ExPerT, an AI personalization framework that adapts large language model (LLM) responses to a user's expertise for each question. It combines what a user asks with patterns in how they type, then adjusts the detail, terminology, and complexity of the response. Existing personalization methods can capture general preferences or background but may miss how much someone knows about a particular topic. ExPerT instead assesses expertise for each query, drawing on both the wording of the question and typing patterns such as keystroke timing, speed, and corrections. These signals are used to estimate expertise on a five-level scale, from novice to expert. Typing behavior provided useful clues. Participants with greater expertise tended to type domain-specific terms faster and with more consistent timing. But neither language nor typing behavior was reliable on its own: an expert may ask a simple question, while someone less familiar with a subject may use technical terms fluently. Combining the two helped ExPerT distinguish between such cases. The team evaluated ExPerT with 40 participants across chemistry, computer science, and business, collecting 1,270 queries along with typing data and self-reported expertise. Adding typing behavior to query semantics reduced the average expertise-estimation error from 0.488 to 0.398 levels, an improvement of 18.4%. Compared with the strongest existing baseline, ExPerT reduced the error by 65.7%, from a mean absolute error of 1.162 to 0.398. Expertise-adapted responses also received higher satisfaction ratings, rising from 3.71 to 4.36 on a five-point scale, an increase of 17.52%. “Existing personalization methods often rely on fixed profiles or previous conversations, making it difficult to reflect how much a user knows about the topic at hand,” said Professor Gong. “By adapting to expertise at the level of each question, ExPerT could support AI systems that communicate more effectively with users of different backgrounds and levels of knowledge.” The approach could be particularly useful in areas where expertise varies widely among users, including education, professional customer support, and specialized services in medicine, law, and finance. Yeji Park, a graduate student in UNIST's Graduate School of Artificial Intelligence and first author of the study, presented the work at ACL 2026 in San Diego. The paper was selected for an oral presentation, placing it among the top 4% of submissions, and received a SAC Highlight Award , given to the top 2%. The dataset and source code are publicly available through theExPerT GitHub repository. 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), through programs supporting AI research and talent development. Journal Reference Yeji Park, Jiwon Tark, and Taesik Gong, "ExPerT: Personalizing LLM Responses to Users' Domain Expertise via Query-Wise Semantic and Keystroke Behavioral Cues," ACL '26, (2026).

2026.09.09

Perovskite Solar Cells Hold Up Under Extreme Temperature Swings

Summary The advent of self-assembled monolayer (SAM)-based hole-selective layers (HSLs) has revolutionized inverted perovskite solar cells (i-PSCs) by enabling record efficiencies, but they are constrained by incomplete surface coverage and interfacial heterogeneity. We devise topology-governed SAMs featuring isomeric linker patterns (LY-ortho [o], LY-meta [m], and LY-para [p]) and uncover the principle that conformational ordering dictates surface coverage and interfacial interaction coupling with perovskite. The topology-optimized LY-m forms a homogeneously dense, stable HSL with superior energy alignment and interfacial passivation. Devices incorporating LY-m achieve a record-high open-circuit voltage of 1.231 V, accompanied by an efficiency of 26.39% (certified 26.37%), while maintaining robust operational stability and scalability. Beyond laboratory metrics, LY-m-based devices withstand −100°C to 100°C thermal cycling with only ∼9% loss, far exceeding the operational limits of state-of-the-art commercial SAMs, while maintaining over 80% of their initial performance at 85°C (T80 ≅ 2,000 h). This work resolves the long-standing role of linker topology and paves the way for extreme-environment-resilient and scalable perovskite photovoltaics. Temperature swings are a major challenge for perovskite solar cells (PSCs), particularly in environments such as space, where materials may repeatedly move between extreme cold and heat. Led by Professors Changduk Yang and Seungjae Shin of the Department of Energy and Chemical Engineering and Professor Myung Hoon Song of the Department of Materials Science and Engineering at UNIST, the team developed perovskite solar cells that withstand such extremes by stabilizing a critical molecular interface. The cells retained about 91% of their initial efficiency after cycling between −100°C and +100°C, with a certified power-conversion efficiency of 26.37%. The key to this advancement is a self-assembled monolayer (SAM), a molecule-thin layer that helps move electrical charges between the electrode and the perovskite absorber. Under severe thermal stress, SAM molecules can shift or cluster, disrupting the interface and degrading performance. The team addressed this problem by changing where the molecular components of the SAM connect. They synthesized three versions with different linker positions and found that LY-m adopted a tilted orientation, allowing its molecules to pack densely and uniformly on the electrode. This arrangement also suppresses the molecular reorganization that can occur under large temperature swings. In thermal-cycling tests, LY-m cells were held alternately at −100°C and +100°C for 30 minutes at each temperature over 16 cycles. They remained operational with only about a 9% loss in efficiency, while cells using a commercial SAM failed under the same conditions. LY-m cells also reached a peak efficiency of 26.39%, with a certified efficiency of 26.37% and an open-circuit voltage of 1.231 V. At 85°C, they maintained more than 80% of their initial performance for approximately 2,000 hours. The approach also performed well at larger scales: a 1 cm² cell achieved 23.25% efficiency, while a 3 × 1 cm² mini-module reached 21.20%. “By changing only the position of the molecular connection, we were able to control how the molecules pack on the electrode,” said Professor Yang. “The resulting interface combines high efficiency with thermal stability, offering a promising route toward perovskite photovoltaics for environments with extreme temperature variations.” The study has been participated by Seunglok Lee and Jeewon Park of the Department of Energy and Chemical Engineering and Jingyu Yang of the Department of Materials Science and Engineering at UNIST, as co-first authors of the study. Published in Joule on August 3, the study has been supported by the Nano & Material Technology Development Program through the National Research Foundation (NRF) of Korea and the InnoCORE program of the Ministry of Science and ICT (MSIT). Journal Reference Seunglok Lee, Jinkyu Yang, Jeewon Park, et al., "Topology-governed self-assembled monolayers enable perovskite solar cells stable at −100°C to +100°C," Joule , (2026).

2026.09.08

AI Model Maps Seismic Responses Across Nuclear Power Plants

A research team, led by Professor Young-Joo Lee of the Department of Civil, Urban, and Environmental Engineering at UNIST and Senior Research Scientist Jaebeom Lee of the Korea Research Institute of Standards and Science (KRISS) has developed a deep-learning model for post-earthquake assessment of nuclear power plants. The model predicts seismic responses across the structure and estimates which locations may require priority inspection. Post-earthquake assessment can be challenging when sensors are installed at only a limited number of locations. Expanding sensor coverage throughout a nuclear power plant is difficult because of cabling, regulatory, and maintenance constraints. Numerical simulations can fill these gaps, but repeated finite element analyzes become computationally expensive when uncertainties in structural properties are taken into account. The researchers developed a deep-learning surrogate model that combines seismic ground motion with structural parameters to predict acceleration responses at 139 locations. This allows responses to be estimated at locations without physical sensors while reducing the computational cost of conventional analysis. The model also accounts for uncertainty in structural properties. Rather than providing only a binary assessment, it calculates the probability that the response at each location will exceed a predefined equipment-level threshold. These probabilities can help engineers prioritize locations for further inspection. Experimental validation showed an average maximum mean absolute percentage error of 1.37%, with R² values above 0.98. The model generates full acceleration time histories across multiple locations, allowing probabilistic responses to be assessed without repeatedly running computationally intensive finite element simulations. The team also developed a method for configuring the AI architecture according to a structure's natural frequencies. The resulting model achieved accuracy comparable to a deep-learning model with more than 200 times as many parameters, while requiring substantially fewer computational resources. The researchers expect the approach could also be adapted to other facilities that require rapid post-earthquake assessment, including semiconductor plants, data centers, and industrial facilities. “For safety applications, accurate prediction alone is not enough; it is also important to understand the uncertainty in those predictions,” the researchers said. “By quantifying that uncertainty, the framework can provide additional information for experts deciding where further inspection is needed.” The research was conducted jointly by the teams of Senior Research Scientist Jaebeom Lee at KRISS, with Jingoo Lee as first author. The work was supported by the National Research Foundation of Korea and KRISS. The findings have been reported in three papers, including a study published in Reliability Engineering & System Safety. The study was conducted jointly by the UNIST and KRISS research teams, with Jingoo Lee as first author. The work was supported by the National Research Foundation of Korea (NRF) and KRISS. The findings have been reported in three papers, including a study published in Reliability Engineering & System Safety (IF: 13.7). Journal Reference Jingoo Lee, Seungjun Lee, Young-Joo Lee, and Jaebeom Lee, “Predicting seismic floor response for nuclear power plant structures with time-series uncertainty propagation using attention-enhanced multimodal deep learning,” Reliab. Eng. Syst. Saf., (2026).

2026.09.07

Reducing Charge Loss in Organic Photoanodes for Solar Water Splitting

Abstract Organic photoelectrochemical (PEC) cells for solar water splitting typically utilize bulk heterojunction (BHJ) structures to circumvent the intrinsically short exciton diffusion lengths of organic semiconductors. However, stochastic donor/acceptor networks in BHJs often result in disordered interfacial contacts, leading to severe charge recombination. Herein, we report a pseudo-bilayer (PS-Bi) organic photoanode featuring a vertically ordered donor–acceptor configuration established via sequential deposition. Our systematic investigation reveals that this PS-Bi structure promotes selective charge transport and suppresses recombination due to its high domain purity, thereby facilitating efficient charge transfer at the interface. Consequently, the optimized PS-Bi photoanodes deliver an enhanced average photocurrent density (Jph) of 1.75 ± 0.03 mA cm−2 at 1.23 V versus the reversible hydrogen electrode for solar water oxidation, while simultaneously exhibiting a significant cathodic onset shift and extended operational stability compared to their BHJ counterparts (Jph = 1.44 ± 0.07 mA cm−2). This study demonstrates that employing the PS-Bi structure is a promising strategy for achieving high-performance and stable organic PEC systems for sustainable solar fuel production. Producing hydrogen from sunlight depends in part on how efficiently a photoelectrode can separate and move electrical charges. In organic photoelectrodes, however, the conventional practice of mixing semiconductor materials can create disordered pathways that allow those charges to recombine, reducing both performance and stability. A research team, led by Professors Han Hee Cho, Moon Kee Choi, and Myung Hoon Song of the Department of Materials Science and Engineering at UNIST, working with researchers at École Polytechnique Fédérale de Lausanne (EPFL), has developed an organic photoanode with a more ordered internal structure. The new design improved photocurrent by about 22% while extending operational stability. The team created what is known as a pseudo-bilayer (PS-Bi) structure by depositing two organic semiconductor materials in sequence. As the second layer is applied, it partially penetrates the first, forming a mixed region at the interface while preserving more distinct layers above and below. Conventional organic photoelectrodes typically use a bulk heterojunction (BHJ), in which donor and acceptor materials are mixed throughout the film. This arrangement helps generate and separate charges, but its irregular internal pathways can also bring electrons and holes back together before they contribute to the chemical reaction. The PS-Bi combines the advantages of both approaches. Charges are generated and separated in the mixed region, while the more ordered layers guide electrons and holes in different directions. This reduces recombination and allows more of the generated charge to drive water oxidation. In tests, the PS-Bi photoanodes achieved an average photocurrent density of 1.75 mA cm⁻², compared with 1.44 mA cm⁻² for conventional BHJ photoanodes, an increase of about 22%. The voltage required to initiate water oxidation also fell from 0.35 V to 0.15 V, allowing the reaction to begin with less external electrical input. The structural change also improved durability. Based on the time required for photocurrent density to fall to 0.5 mA cm⁻², the pseudo-bilayer photoanodes remained operational substantially longer than their BHJ counterparts. The researchers also found that the two structures degrade differently. In BHJ photoanodes, highly reactive holes can damage the organic semiconductor within the intermixed structure. The more ordered pseudo-bilayer reduces damage within the donor–acceptor region, helping the photoanode maintain its performance for longer. “Conventional organic photoelectrodes mix donor and acceptor materials throughout the film, which can lead to charge losses and faster degradation,” said Professor Cho. “By controlling how the two materials are arranged, we were able to improve both charge transport and operational stability.” He further added, “The lower operating voltage could also help in the development of systems that ultimately produce hydrogen using sunlight without an external electrical bias.” The study was published online in Chemical Engineering Journal on July 16, 2026. The research was supported by the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT (MSIT), and the InnoCORE program. Journal Reference Gyu Won Chae, Seong Rae Kang, Jin Su Park, et al. , “A pseudo-bilayer organic photoanode for solar water oxidation,” Chem. Eng. J., (2026).

2026.09.03

New Process Shields 2D Semiconductors from Fabrication Residue

Abstract Achieving the theoretical performance of 2D molybdenum disulfide (MoS2) electronics is currently bottlenecked by interfacial contamination derived from lithographic processing. A critical, yet often overlooked, mechanism is the interaction between photoresist (PR) and plasma, which chemically alters the interface. In this work, we elucidate the atomistic origin of this degradation: theoretical calculations based on molecular dynamics (MD) and density functional theory (DFT) reveal that plasma exposure functionalizes the PR with oxygen, drastically increasing its adsorption energy on the MoS2 surface from −1.15 to −2.36 eV. This doubling of adsorption energy creates thermodynamically stable, hardened residues that resist conventional removal methods. To overcome this fundamental limitation, we introduce a universal prevention-first strategy: Sacrificial Metal Mask Lithography. By utilizing a sacrificial layer to physically isolate the channel, we shield the MoS2 from reactive plasma species, preventing the formation of hardened residue entirely. This strategy results in a ten-fold reduction in contact resistance (RC) from 2.59 × 106 Ω·µm to 2.58 × 105 Ω·µm and an order-of-magnitude enhancement in on-current. Crucially, we demonstrate the universality of this method by successfully applying it to top-gated transistor arrays, electron-beam lithography (EBL), and atomic layer deposition (ALD)-synthesized films. Two-dimensional (2D) semiconductors, such as molybdenum disulfide (MoS2) hold promise for smaller, thinner electronics, but their extreme thinness also makes them particularly vulnerable to contamination during fabrication. Even small amounts of residue can interfere with electrical performance. A research team, led by Professor Myung-Soo Kim of the Department of Electrical Engineering and Professor Byoung Jo Kim of the Graduate School of Semiconductor Materials and Devices Engineering at UNIST, has developed a fabrication method that protects MoS2 from photoresist (PR) residue. The approach uses a thin, removable metal layer to keep the residue from reaching the semiconductor surface. PR is a light-sensitive material used to pattern circuits during semiconductor fabrication. When exposed to plasma, however, it can leave residue on MoS2 that conventional solvents struggle to remove. More aggressive cleaning can damage or even detach the atomically thin material from its substrate. Rather than removing the residue after it forms, the researchers prevent it from reaching the MoS2 in the first place. They place a 5-nanometer-thick metal layer over the semiconductor before patterning. Once fabrication is complete, the metal is dissolved, carrying the accumulated residue with it. Transistors made with the new process had a contact resistance of 2.58 × 10⁵ Ω·μm, about one-tenth that of devices made conventionally. Their on-current, or the current flowing when the transistor is switched on, increased by roughly tenfold. The researchers also found no clear damage to the crystal structure of MoS2 after the metal layer was removed. The team also uncovered why the residue is so difficult to remove. Molecular-level calculations showed that plasma exposure introduces oxygen into the PR, changing its chemistry and causing it to bind much more strongly to MoS2. The calculated adsorption energy shifted from −1.15 to −2.36 eV, roughly doubling the strength of the interaction. “This process can be used not only with conventional semiconductor patterning, but also with electron-beam lithography for finer features,” said Dahyun Kim, first author of the study. “We also found that it works with MoS2 produced by different synthesis methods.” The researchers tested the method with top-gated transistor arrays, electron-beam lithography, and MoS2 films grown by atomic layer deposition, showing that it can be used across different device structures and fabrication methods. “MoS2 is a promising material for making semiconductor devices, such as transistors smaller and thinner,” said Professor Myung-Soo Kim. “By addressing process contamination that has limited device performance, this work could help advance highly integrated logic and memory devices based on 2D semiconductors.” The study was published online in Small on August 11, 2026. The research was supported by the Ministry of Science and ICT (MSIT), National Research Foundation of Korea (NRF), Institute of Information & Communications Technology Planning & Evaluation (IITP), InnoCORE program, Ministry of Trade, Industry and Energy (MOTIE), and Korea Institute for Advancement of Technology (KIAT). Journal Reference Dahyeon Kim, Yanfeng Zhao, Sungyeon Kim, et al ., “Preventing Plasma-Induced Resist Hardening in Molybdenum Disulfide Transistors via Sacrificial Metal Mask Lithography,” Small, (2026).

2026.09.02

Rethinking What Matters

《Editor's Note: What does it take to make something work better? The answer may be to pare back what is unnecessary, rethink how existing parts are arranged, find another route around a constraint, or widen the frame of the problem itself. This month's UNIST research offers answers in many forms—from leaner AI systems and redesigned material structures to an unexpected route for electron transfer and a broader view of cancer drug resistance. Some of those answers run counter to what we might expect. 》 Less Can Be Enough “How much does AI really need?” AI has spent years scaling up. But recent UNIST research asks a different question—how much can be taken away while keeping what matters? At ACL 2026, researchers introduced GMoE, an architecture that shares experts across layers rather than repeatedly storing similar ones. It cut model parameters by about 63 percent, from 549 million to 204 million, while average accuracy remained virtually unchanged—39.51 percent, compared with 39.55 percent for the baseline. The next target was data. Presented at ICML 2026, CSOR removes redundant training images while preserving meaningful variation within each identity. Using roughly half the original dataset, models retained more than 95 percent of full-data performance in re-identification (ReID) tasks. A third study moved to hardware. An automated design framework for logic-in-memory (LiM) computing optimized the placement and routing of computing elements, cutting computation cycles by 19.6 percent under area constraints. ■ GMoE—Expert Sharing for Smaller AI Models (ACL 2026) ■ CSOR—Less Training Data for ReID AI (ICML 2026 ) ■ LiM—Smarter Chip Design (IEEE TCAD) Arrangement Matters “In materials, where can matter as much as what.” More iron, less corrosion. Iron (Fe) is one of magnesium's most damaging impurities, yet UNIST researchers produced a Mg alloy containing more than six times as much Fe as ultra-high-purity Mg—and it still corroded more slowly. Instead of removing the impurity, the team added scandium (Sc) to form core–shell particles that isolate Fe-rich regions from the surrounding Mg. The corrosion rate fell from 22.2 to 0.38 millimeters per year, nearly a 60-fold reduction. At the molecular scale, thermal annealing changed the packing of an organic semiconductor, more than tripling its absorption dissymmetry and even reversing its preferred handedness of circularly polarized light (CPL). The resulting near-infrared (NIR) photodetector achieved a dissymmetry factor of 0.1 at 850 nanometers. The scale shifts again in perovskite solar cells (PSCs). The same high-efficiency formulation crystallized differently depending on the buried interface beneath it. Tailoring the chloride chemistry to the inverted architecture produced a device with 26.3 percent efficiency and a fill factor (FF) of up to 86.8 percent. ■ Mg Alloy—Shielding Fe to Slow Corrosion (Science Advances) ■ NIR Photodetector—Turning Molecular Packing (Advanced Science) ■ PSCs—Why Architecture Changes Performance (Joule) Another Way Through “How spin opens a new route for electron transfer.” Conventional energetics suggested that electron transfer should be unfavorable. Yet in manganese (Mn)-doped quantum dots (QDs), spin-exchange interactions enabled transfer more than ten times faster than in undoped QDs. The mechanism could help capture energetic hot electrons before their excess energy dissipates—a longstanding challenge in photocatalysis and solar-energy conversion. ■ Hot Electrons—Spin-Enabled Transfer (Nature Communications) Beyond the Tumor “The cells helping cancer resist treatment.” Some of the cells helping breast cancer resist treatment are not part of the tumor at all. UNIST researchers found that nearby fat cells can become cancer-associated adipocytes, which release complement factor D and activate survival signaling in cancer cells. In mouse models, disrupting this TRAP1-driven communication improved the response to chemotherapy, pointing to surrounding adipocytes as potential therapeutic targets alongside the tumor itself. ■ Breast Cancer—How Nearby Fat Cells Drive Drug Resistance (STTT) Explore the latest issue and follow 『The UNIST Brief』 on LinkedIn to stay updated. Subscribe via LinkedIn

2026.09.01