Smarter Design Method Improves Efficiency of In-Memory Computing Chips

Abstract Data-centric applications continue to be limited by the memory wall. Logic-in-Memory (LiM) architectures based on non-volatile memories (NVMs), such as memristors, offer a promising solution by enabling in-memory computation and eliminating costly data movement. While several approaches leveraging memristor-aided logic (MAGIC) operations have been proposed, many fail to fully exploit the parallelism and spatial efficiency of memristor crossbars, especially under physical design constraints. In this paper, we propose a parallelism-driven, area-aware look-up table (LUT)-based mapping framework of arbitrary logic circuits into a memristor crossbar using MAGIC operations. We present a crossbar-aware floorplanning strategy that leverages topological LUT information to enable two-directional parallel computation (ie, HNOR and VNOR), thereby maximizing computational LUTs with minimal execution cycles. We formulate an integer linear programming (ILP) approach to define the placement of each LUT and its fanin locations, improving routability while preserving parallelism. Moreover, we introduce our novel A∗-search-based fan-in routing method with Hanan grid that fully utilizes the aligned fans and intermediate results. Experimental results demonstrate that the proposed framework achieves maximum parallelism and crossbar utilization, thereby reducing cycle counts by 19.6% and ensuring that all benchmarks map successfully under strict area constraints. We also confirm the scalability and practical applicability of the proposed framework for large-scale logic mapping in memristor-based LiM systems. Modern computing is increasingly limited not by processing speed, but by the time and energy required to move data between memory and processors. Researchers at UNIST have developed an automated design framework that addresses this challenge by making logic-in-memory (LiM) chips more efficient. Led by Professor Heechun Park of the Department of Electrical Engineering, the team designed a framework that automatically optimizes the placement of logic circuits and data pathways in memristor-based LiM architectures. By enabling more operations to run simultaneously while making better use of limited chip area, the approach reduced the number of computation cycles by nearly 20% compared with existing state-of-the-art methods. Unlike conventional computer chips, LiM architectures perform computation directly where data are stored, reducing costly data movement between memory and processors. This approach helps overcome the memory wall where data transfer limits overall computing performance. The framework determines where computations are performed within a memristor crossbar array and how intermediate results move to the next operation. Unlike previous approaches, it enables computations to proceed in both the horizontal and vertical directions, allowing more operations to run simultaneously while making better use of the available chip area. The framework also reorganizes intermediate data during computation, freeing unused memory cells and reducing wasted space. It then identifies efficient routes for transferring intermediate results between computational blocks, preserving parallel execution even under tight area constraints. In benchmark evaluations, chips designed using the new framework required 19.6% fewer computation cycles than those designed using existing approaches. The framework also successfully mapped every benchmark circuit, including designs that previous methods could not accommodate under the same area constraints, suggesting it can scale to larger LiM systems. “The performance of LiM computing depends not only on the memory device itself, but also on how computations are organized within the array,” said Professor Park. "By arranging operations to maximize parallel execution, our framework improves efficiency while remaning practical under realistic area constraints." The research was conducted by Ikkyum Kim and Minhong Kim as first and second authors, respectively, with Professor Heechun Park serving as the corresponding author. The findings were published online in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (IEEE TCAD) on June 2, 2026. The research was supported by the National Research Foundation of Korea (NRF) and the Institute of Information & Communications Technology Planning & Evaluation (IITP). The EDA tool was supported by the IC Design Education Center (IDEC). Journal Reference Ikkyum Kim, Minhong Kim, and Heechun Park, “A Parallelism-Driven, Area-Aware Technology Mapping Framework for Memristive Logic-in-Memory,” IEEE TCAD , (2026).

2026.08.06

New Organic Photodetector Improves Polarized Near-Infrared Light Detection

Abstract Near-infrared (NIR) circularly polarized light (CPL) photodetection is of great importance due to its broad application potential in bioimaging, wearable healthcare, optical communication, and advanced optoelectronic systems. In this study, a supramolecular chirality evolution strategy in chiral low-bandgap fused-ring conjugated molecules (LFCs) is presented for high-performance NIR CPL photodetection using Schottky barrier vertical organic field-effect transistors (SB-VOFETs). Halogen substitution combined with thermal annealing drives inversion and amplification of supramolecular chirality in enantiopure LFC thin films. F-substituted LFCs exhibit progressive domain growth and hierarchical ordering with increasing annealing temperature, whereas Cl-substituted LFCs show limited structural evolution above 150°C. These distinct crystallization behaviors directly correlate with chiroptical responses, with F-substituted LFCs achieving a maximum absorption dissymmetry factor (|gabs|) of ∼0.1. When integrated into SB-VOFETs, the optimized chiral films enable highly efficient NIR CPL photodetection, delivering a photocurrent dissymmetry factor (|gph|) of ∼0.1, a specific detectivity of 4.9 × 1011 Jones, an external quantum efficiency exceeding 900%, and a fast response time of ∼600 µs at 850 nm. These metrics represent the highest performance reported for NIR CPL. This study provides design guidelines for advancing high-performance chiral optoelectronic devices through the synergistic integration of atomic substitution, thermal annealing, and device architecture engineering. Near-infrared (NIR) circularly polarized light (CPL) carries information that conventional light cannot, making it valuable for applications ranging from autonomous vehicle sensors and bioimaging to optical security and advanced communications. Detecting this light accurately, however, has remained difficult because the organic semiconductor films used in these devices often lose the molecular order needed to distinguish different polarization states. A research team, led by Professor BongSoo Kim of the Department of Chemistry at UNIST and Professor Joon Hak Oh of Seoul National University has found a way to overcome that limitation. By controlling how chiral organic semiconductor molecules organize within thin films, the researchers significantly improved the films' ability to distinguish between left- and right-handed CPL. When incorporated into a photodetector, the optimized films delivered what the team reports as the highest performance yet achieved for NIR CPL detection. Unlike conventional light detectors, CPL can distinguish between two different polarization states of light. This additional information can reveal not only an object's shape but also properties, such as its surface composition or molecular structure, making the technology valuable for sensing and imaging applications. The researchers achieved the improvement by applying a simple thermal annealing process to thin films made from fluorine-substituted chiral organic semiconductor molecules. Rather than remaining randomly arranged, the molecules reorganized into larger, more ordered crystal domains. This more ordered molecular arrangement strengthened the film's interaction with CPL. As a result, the film's ability to selectively absorb CPL increased more than threefold after annealing at 250°C, with its absorption dissymmetry factor rising from approximately 0.03 to 0.1. Structural analysis confirmed that the annealed films adopted a highly ordered molecular arrangement closely resembling that of single crystals. The researchers also discovered an unexpected effect. As the molecular packing reorganized, the film reversed the polarization it preferentially absorbed, demonstrating that heat treatment could control not only the strength but also the handedness of the material's optical response. By comparison, similar films containing chlorine instead of fluorine showed little structural change above 150°C and exhibited lower polarization selectivity after annealing. To turn these improved optical properties into measurable electrical signals, the team integrated the optimized films into a vertical organic transistor. The device architecture efficiently collected and amplified charges generated by incoming light, allowing the detector to fully capitalize on the improved molecular organization. The resulting photodetector achieved a photocurrent dissymmetry factor of 0.1 at a wavelength of 850 nanometers, demonstrating excellent discrimination between left- and right-handed CPL. It also recorded a specific detectivity of 4.9 × 10¹¹ Jones, an external quantum efficiency of up to 909%, and a response time of less than 600 microseconds—the highest performance reported to date for NIR CPL detectors. “We found that simply controlling how the molecules organize within a thin film can dramatically improve how the materials respond to CPL,” the research team said. “We hope this strategy will help advance next-generation optical sensors and imaging technologies that rely on precise light detection.” Their findings were published online in Advanced Science on June 28, 2026. The research was supported by the National Research Foundation of Korea (NRF) and the Ministry of Science and ICT (MSIT). Journal Reference Jaeyong Ahn, Kwangmin Kim, Sangwook Lee, et al ., “Thermally Driven Supramolecular Chirality Evolution in Low-Bandgap Fused-Ring Conjugated Molecules for High-Performance NIR Circularly Polarized Light Detection,” Adv. Sci., (2026).

2026.08.05

New Benchmark Reveals Limits in How AI Understands Hands

Abstract Understanding the fine-grained articulation of human hands is critical in high-stakes settings such as robot-assisted surgery, chip manufacturing, and AR/VR-based human-AI interaction. Despite achieving near-human performance on general vision-language benchmarks, current vision-language models (VLMs) struggle with fine-grained spatial reasoning, especially in interpreting complex and articulated hand poses. We introduce HandVQA, a large-scale diagnostic benchmark designed to evaluate VLMs' understanding of detailed hand anatomy through visual question answering. Built upon high-quality 3D hand datasets (FreiHAND, InterHand2.6M, FPHA), our benchmark includes over 1.6M controlled multiple-choice questions that probe spatial relationships between hand joints, such as angles, distances, and relative positions. We evaluate several state-of-the-art VLMs (LLaVA, DeepSeek and Qwen-VL) in both base and fine-tuned settings, using lightweight fine-tuning via LoRA. Our findings reveal systematic limitations in current models, including hallucinated finger parts, incorrect geometric interpretations, and poor generalization. HandVQA not only exposes these critical reasoning gaps but provides a validated path to improvement. We demonstrate that the 3D-grounded spatial knowledge learned from our benchmark transfers in a zero-shot setting, significantly improving accuracy of model on novel downstream tasks like hand gesture recognition (+10.33%) and hand-object interaction (+2.63%). Understanding a human hand gesture requires more than recognizing fingers. An AI model must also make sense of how those fingers bend, how joints related to one another, and how the entire hand changes with viewpoint. Current vision-language models remain surprisingly weak at this kind of finger-grained reasoning. A research team, led by Professor Seungryul Baek of the Graduate School of Artificial Intelligence at UNIST has developed HandVQA, a new benchmark that tests this ability in detail. Drawing on 3D hand data, the benchmark contains more than 1.6 million questions about joint angles, distances, and relative positions. It also gives researchers a way to improve the skills it is designed to measure. Despite strong performance on general image-and-language tasks, vision-language models can struggle when spatial differences become subtle. Existing benchmarks rarely examine hand anatomy at the level of individual joints, making these weaknesses difficult to measure. HandVQA fills that gap with questions generated from hand images and precise 3D joint coordinates. For each image, the benchmark asks 25 questions about properties such as finger flexion, the distance between joints, and whether one joint is above, below, in front of, or behind another. The results showed just how much current models miss. Without specialized training, several leading vision-language models performed near chance on some spatial questions and had particular difficulty judging distances between joints. They also made geometric errors and, in some cases, referred to finger parts that did not exist in the image. But the same benchmark that exposed these weaknesses also helped correct them. After LLaVA was fine-tuned with HandVQA, its accuracy on questions about joint distances increased from 16.20% to 90.79%. The benefits also carried over to tasks outside the benchmark. Qwen-VL, after learning from HandVQA, improved on two tasks it had not been directly trained for: hand gesture recognition by 10.33 percentage points and hand-object interaction by 2.63 percentage points. This transfer suggests that the model learned a broader understanding of hand geometry rather than simply becoming better at answering HandVQA questions. “HandVQA not only helped the model answer questions it had previously struggled with, but also improved its ability to handle new tasks,” said MD Khalequzzaman Chowdhury Sayem, first author of the study. “The spatial knowledge learned from the benchmark transferred to other hand-related tasks without requiring additional task-specific training.” Professor Baek added, “Accurately understanding hand pose is important in applications where even small errors can matter, from robotic manipulation and AR/VR interfaces to assistive technologies. HandVQA helps identify where current models fall short and provides a way to improve those capabilities.” Their findings have been accepted to the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026, one of the leading international conferences in computer vision. The study was supported by the National Research Foundation of Korea (NRF) through the Mid-Career Researcher Program and the Basic Science Research Program, along with programs administered by the Institute for Information communication Technology Planning and Evaluation (IITP), including the AI Star Fellowship, AI Graduate School, and the LG AI STAR Talent Development Program for Leading Large-Scale Generative AI Models in the Physical AI Domain programs. Journal Reference MD Khalequzzaman Chowdhury Sayem, Mubarrat Tajoar Chowdhury, et al. , “HandVQA: Diagnosing and Improving Fine-Grained Spatial Reasoning about Hands in Vision-Language Models,” CVPR '26, (2026).

2026.08.04

Solution-Processed Semiconductor Enables High-Performance RF Switches for 6G and Satellite Communications

Abstract Next-generation 6G communication requires radio-frequency components capable of operating above 100 GHz with low loss, high isolation, and zero static power—requirements that challenge complementary metal-oxide-semiconductor (CMOS) and microelectromechanical systems (MEMS) technologies. Here, we report transfer-free, large-area millimeter-wave (mmWave) switches based on solution-processed MoS2. While solution-processed 2D materials are often viewed as inferior to their crystalline counterparts due to high defect densities, we demonstrate that their edge-rich morphology is, in fact, a performance enabler. These edge defects act as intrinsic templates that confine Cu-filament pathways, enabling rapid (76 ns) and low-energy switching (1.57 nJ) with > 2000 cycles and uniform zero static-power operation, as corroborated by Kelvin probe force microscopy, conductive atomic force microscopy and low-temperature studies. The resulting switches achieve an low insertion loss ( < 0.1 dB) and high isolation ( > 35 dB) at 67 GHz. Notably, they exhibit a switching figure-of-merit (RON·COFF) of ~ 0.8 fs (fco ~ 187 THz), surpassing previously reported 2D switches. Importantly, by adopting an inverse-state operational scheme in a SHUNT architecture, we mitigate self-switching and achieve improved power handling (P0.1dB > 10 dBm) and linearity (IIP3 > 42.1 dBm). Finally, we demonstrate the platform's circuit-level viability by integrating the switches into a true-time-delay and hybrid-coupled phase shifters targeting 30 GHz mmWave applications. Researchers at UNIST have developed a high-performance radio-frequency (RF) switch using a solution-processed two-dimensional (2D) semiconductor, demonstrating stable operation at frequencies up to 67 GHz while consuming no static power. The technology could support future 6G wireless networks, satellite communications, radar systems, and other high-frequency communication platforms that demand compact, energy-efficient components. Led by Professor Myungsoo Kim of the Department of Electrical Engineering, the team fabricated the device using a thin film of solution-processed molybdenum disulfide (MoS2). Unlike conventional fabrication methods that require 2D semiconductor films to be grown and transferred onto a substrate, the new approach forms the semiconductor layer directly from a liquid precursor, simplifying manufacturing and making large-area production more practical. RF switches are fundamental components in wireless communication systems. They control the flow of high-frequency signals in smartphones, wireless base stations, satellite communications, radar systems, and autonomous vehicles. As communication technologies move toward the millimeter-wave frequencies required for future 6G and space communications, RF switches must transmit signals with minimal loss, effectively block unwanted leakage, and operate with very low power consumption. The newly developed switch achieved all three. Operating at 67 GHz, it recorded an insertion loss below 0.1 dB and signal isolation greater than 35 dB, allowing nearly all of the incoming signal to pass while suppressing signal leakage to less than 0.03%. The device also requires no standby power. Once switched on or off, it maintains its operating state even after power is removed, reducing unnecessary energy consumption. The researchers attribute the device's durability to the characteristics of the solution-processed MoS2 film. During fabrication, naturally occurring sulfur vacuums help guide the movement of copper ions that create the conductive pathways responsible for switching. Rather than reducing performance, these defects stabilize the switching process. Devices fabricated using solution-processed films continued to operate reliably after more than 2,000 switching cycles, whereas comparable devices fabricated from mechanically exfoliated MoS2 showed a lifetime of only about 100 cycles. To evaluate its practical potential, the team integrated the switch into communication circuits used for high-frequency signal control, including a transmit/receive switch and a beam-steering phase shifter. Both operated successfully, demonstrating that the technology can be incorporated into circuits used in satellite communications, radar, defense systems, and future wireless networks. Beam-steering phase shifters electronically direct radio waves without physically moving the antenna, enabling fast and precise tracking of moving targets such as satellites and drones. “By using solution-processed 2D semiconductor materials, we developed an RF switch that combines low signal loss with excellent isolation in the millimeter-wave band without requiring complex transfer processes,” said Professor Kim. “Because the switch retains its operating state even when power is removed, it consumes no standby power and could contribute to smaller, more energy-efficient communication systems for 6G, satellite communications, radar, and defense applications.” The findings were published online in Nature Communications on June 18, 2026. The research was supported by the National Research Foundation of Korea (NRF) and the Institute of Information & Communications Technology Planning & Evaluation (IITP), including the Space-K BIG Project and the InnoCORE Program of the Ministry of Science and ICT (MSIT). Journal Reference Changwoo Pyo, Sungmoon Park, Seungchan Lee, et al ., "Sub-femtosecond figure-of-merit millimeter-wave switches via solution-processed MoS2 for 6G radio-frequency front-ends," Nat. Commun ., (2026).

2026.08.03

Beyond the Obvious

《Editor's Note: The biggest scientific advances do not always come from discovering something new. Sometimes, they begin by seeing familiar problems in a new light. This month's stories invite us to look beyond the obvious. Across artificial intelligence, semiconductor materials, nanotechnology, and biomedical science, UNIST researchers examined assumptions, structures, and interactions that can easily go unnoticed—but often determine how a system behaves. Their findings show how a deeper understanding of what is already there can open new directions for discovery.》 Questioning Assumptions “The first answer is not always the right one.” Scientific discovery often begins by asking whether the original question was complete. Several AI studies featured this month did just that, revisiting assumptions about how machines learn, forget, reason, and communicate. In federated learning, researchers found that differences in camera viewpoints—not only privacy constraints or data distribution—can significantly affect how well AI models learn together. Another team showed that machine unlearning involves more than simply removing data: forgetting can unintentionally alter knowledge that should remain. Elsewhere, a new communication framework showed that effective cooperation among AI agents depends less on exchanging more information than on exchanging the right information. Other studies reconsidered how AI systems reason across multiple views of the same scene and explored how quantum computing could reduce the cost of training robust AI models. Different problems, but a common starting point: questioning the assumptions built into them. Together, these studies point to a simple lesson: better answers often begin with better questions. ■ Rethinking Federated Learning: Camera Viewpoints Matter (CVPR 2026) ■ Making AI Forget—Without Losing What Matters (ICML 2026 ) ■ Smarter Communication for Collaborative AI (ICML 2026) ■ Rethinking Multiview Reasoning (ICML 2026) ■ Quantum Algorithm Cuts the Cost of Training Robust AI (ICML 2026) Looking Beneath the Surface “The smallest details often hold the biggest clues.” Across materials science, biology, nanotechnology, and semiconductor engineering, UNIST researchers examined structures and interactions that are easy to miss but can have an outsized influence on how materials, devices, and biological systems behave. In semiconductor research, the spacing between neighboring atoms emerged as a critical factor in controlling material properties and device performance. At the nanoscale, researchers examined how the surface chemistry of quantum dots influences their behavior in imaging. Another study turned to microscopic wrinkles, showing how their geometry can reshape optical properties. The same attention to underlying mechanisms extended to biology, where researchers found that familiar forms of cell death do not always follow the pathways scientists expect. Together, these studies show how seemingly minor details—from atomic spacing and surface chemistry to microscopic geometry and cellular mechanisms—can fundamentally change our understanding of a system. Sometimes, the most consequential discoveries begin with something that was there all along, waiting to be examined more closely. ■ Atomic Precision for Better Semiconductors (Chem. Mater., 2026) ■ A New Perspective on QD Imaging (Nano Converg., 2026) ■ Wrinkles that Reshape Photonics (Adv. Funct. Mater., 2026) ■ When Cell Death Follows Different Rules (Cell Death Differ., 2026) Across very different fields, these studies share a common instinct: question the assumption, examine the detail, and look again. What appears familiar can still hold unanswered questions. Looking beyond the obvious is not simply how these discoveries were made. It is how scientific discovery moves forward. Explore the latest issue and follow 『The UNIST Brief』 on LinkedIn to stay updated. Subscribe via LinkedIn

2026.08.01

New Federated Learning Method Advances Privacy-Preserving AI

Abstract Extending person re-identification (ReID) to a federated scenario has recently drawn attention due to privacy concerns of individuals, but existing methods mostly assume sufficient diversity in pose variations even within a decentralized client. We focus on a more realistic federated-by-camera scenario, where each client corresponds to a single camera and thus captures only a sparse set of poses. To enrich variety, we propose Pose-guided Enriched Feature Learning (PEFL) that explicitly augments pose-diverse samples in the federated ReID scenario. Specifically, a Pose-Extraction Module (PEM) disentangles pose-relevant and pose-irrelevant feature components, where Pose-Relationship Knowledge Distillation (PKD) method helps identify the correct pose and Semantic Consistency Maintenance (SCM) method preserves semantics even with pose changes. In addition, a Compatibility Regularization method ensures the PEM to be compatible with the feature space of the global model. By recombining pose-relevant and -irrelevant components across identities via PEM, our PEFL synthesizes pose-swapped features, thereby largely facilitating contrastive learning of ReID models. Extensive experiments on Market1501 and MSMT17 under the federated-by-camera setting demonstrate that PEFL consistently outperforms federated ReID baselines and their conjunctions with the existing feature augmentation methods; thus achieving state-of-the-art federated ReID performance. Better AI often begins with better data. For systems designed to identify the same person across different camera views—whether for missing-person searches or public safety—that data is becoming harder to obtain as privacy concerns restrict the sharing of sensitive video. A research team, led by Professor Jae-Young Sim of the Graduate School of Artificial Intelligence at UNIST has developed a federated learning method that overcomes this challenge without transferring raw camera footage to a central server. Determining whether people captured by different cameras are the same individual—a task known as Person Re-Identification (ReID)—is essential for a wide range of AI vision systems. While federated learning enables models to be trained without sharing sensitive images, distributing training across individual cameras creates an unexpected challenge: each camera observes people from only a limited range of viewpoints. A camera mounted at one location may capture mostly front-facing views, while another records only side or rear profiles. As a result, local AI models learn from incomplete pose information, making it more difficult for the combined federated model to accurately recognize the same person under different viewing conditions. To overcome this limitation, the research team developed Pose-guided Enriched Feature Learning (PEFL). Rather than generating or exchanging additional images, PEFL separates visual features related to a person's identity from those associated with pose, then recombines them to synthesize feature representations from viewpoints that were never observed by an individual camera. The approach expands pose diversity during training while preserving each person's distinctive characteristics. The researchers evaluated PEFL using the widely adopted Market-1501 and MSMT17 benchmarks under a federated-by-camera setting. The method achieved a mean Average Precision (mAP) of 45.9%, improving on the previous state of the art by 4.5 percentage points, while reaching a Rank-1 accuracy of 66.4%. “We found that the limited range of viewpoints captured by individual cameras significantly reduces the effectiveness of federated person re-identification,” said JooHyung Oh, co-first author of the study. “By enriching pose diversity without sharing raw images, our method enables models to learn more robust identity representations.” Professor Sim added, "Privacy and performance are often viewed as competing goals in AI. Our work shows they do not have to be. We hope this approach will contribute to distributed vision systems that require both strong recognition performance and robust privacy protection." The study was co-first authored by JooHyung Oh and Minyoung Oh of UNIST. The research was accepted for presentation at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 , one of the world's leading conferences in computer vision. Journal Reference JooHyung Oh, Minyoung Oh, Sung Whan Yoon, and Jae-Young Sim, "Pose-guided Enriched Feature Learning for Federated-by-camera Person Re-identification," CVPR '26, (2026).

2026.07.31