Designing Smarter Communication for Multi-Agent AI

Abstract Communication is a key component in multi-agent reinforcement learning (MARL) for mitigating partial observability, yet prior approaches often rely on inefficient information exchange or fail to transmit sufficient state information. To address this, we propose LLM-driven Multi-Agent Communication (LMAC), which leverages an LLM's reasoning capability to design a communication protocol that enables all agents to reconstruct the underlying state as accurately and uniformly as possible. LMAC iteratively refines the protocol using an explicit state-awareness criterion, improving state recovery while narrowing differences in agents' knowledge. Experiments on diverse MARL benchmarks show that LMAC improves state reconstruction across agents and yields substantial performance gains over prior communication baselines. Multi-agent AI systems must make decisions based on incomplete information. Researchers at UNIST have developed a new approach that uses a large language model (LLM) to help AI agents to build a more accurate and consistent understanding of their shared environment, improving coordination in complex collaborative tasks. Led by Professor Seungyeol Han of the Graduate School of Artificial Intelligence, the team developed LLM-driven Multi-Agent Communication (LMAC), a framework that designs and refines communication protocols for multi-agent reinforcement learning (MARL). Rather than manually specifying what information agents should exchange, LMAC determines what information is essential for cooperation and how it should be shared so that every agent can reconstruct the underlying state of the environment as accurately as possible. The framework is intended for environments where no individual agent has access to the complete picture, such as drone swarms, collaborative robots, and autonomous vehicles. Users describe the task objective, the role of each agent, and the meaning of available observations in natural language. Based on this description, the LLM generates a communication protocol that specifies what information each agent should transmit and to whom. LMAC further improves the protocol through an interactive feedback process. An evaluation module measures how accurately each agent reconstructs the hidden state after receiving messages from teammates while identifying differences in the agents' understanding of the environment. The LLM uses this feedback to refine the communication protocol by incorporating missing contextual information and eliminating unnecessary information exchange. Rather than increasing the amount of communication, the framework improves the quality and consistency of information shared across the team. For example, in a drone surveillance mission, only a reconnaissance drone may be able to observe a target. Simply reporting that " [T]he target is to my right " provides little value if the rest of the team does not know the scout's position. LMAC automatically recognizes this ambiguity and revises the communication protocol to include additional context, such as the scout's location, movement, teammate identities, or a shared coordinate system, allowing all agents to infer the target's position more accurately. Unlike approaches that repeatedly invoke an LLM during deployment, LMAC uses the model only while designing the communication protocol. During execution, the completed protocol directly converts each agent's observations into messages, enabling efficient real-time communication without the computational overhead of continuous LLM inference. The researchers also developed a learning architecture that helps agents make better use of incoming messages. Rather than treating all transmitted information equally, each agent learns to extract the information most relevant to the task while estimating the reliability of its own understanding of the environment. In a cooperative StarCraft II benchmark, LMAC achieved a 96.2% win rate in a scenario where a single reconnaissance unit guided ten attack units that could not directly observe enemy positions. Under the same learning architecture, existing communication methods achieved win rates of 59.0% and 25.2%, respectively. “Previous multi-agent communication research has focused primarily on how agents exchange messages, but has been limited in determining what information is fundamentally necessary for cooperation,” said Professor Han. "LMAC enables a large language model to reason about the relationship between observations and the underlying environment, automatically design communication protocols, and continuously refine them through data-driven feedback. Because the LLM is used only during protocol design rather than execution, the framework also remains computationally efficient during deployment." He further added that the approach could support collaborative AI systems operating under limited information, including autonomous drone swarms, robot teams, smart factories, and cooperative autonomous driving. Professor Han added that the approach could provide a foundation for collaborative AI systems operating under limited information, including autonomous drone swarms, robot teams, smart factories, and cooperative autonomous driving. The study has been accepted for presentation at the International Conference on Machine Learning (ICML) 2026, which took place in Seoul from July 6 to 11, 2026. The research was supported by the Ministry of Science and ICT (MSIT), the Institute for Information & Communications Technology Planning & Evaluation (IITP), and the National Research Foundation of Korea (NRF). Journal Reference Sangjun Bae, Yisak Park, Sanghyeon Lee, and Seungyul Han, "LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning," ICML '26 , (2026).

2026.07.27

New Surface Design Makes QDs More Reliable for Single-Molecule Imaging

Abstract Quantum dots (QDs) are fluorescent nanoparticles widely used for single-molecule imaging because of their exceptional brightness and photostability. However, the impact of QD surface chemistry on biomolecular interactions has not been systematically investigated. Here, we report that commercial QDs unexpectedly destabilize protein-DNA complexes by inducing protein dissociation from DNA. Using the human nucleotide excision repair protein, xeroderma pigmentosum complementation group A (XPA) as a model system, we demonstrate that antibody-conjugated QDs promote dissociation of XPA from DNA substrates, independently of sizes and surface modification of QDs, antibody types, epitope tags, buffer conditions, or DNA structures. We find that polyethylene glycol (PEG), a common polymer coating on QD surfaces, is the primary factor responsible for this effect. To tackle this problem, we engineered QDs with precisely controlled surface polymer compositions. By systematically changing the ratio of anchoring, hydrophilic, and PEG-based functional groups, we find that reducing PEG density below a critical threshold effectively suppresses protein dissociation while maintaining excellent colloidal stability and brightness. Furthermore, antibodies conjugated via click chemistry between azide groups and DBCO enabled specific labeling of XPA without perturbing the DNA binding activity. Using these optimized QDs, we conducted single-molecule DNA curtain assays to visualize XPA-DNA interactions. QD-labeled XPA exhibits one-dimensional diffusion with frequent pausing on undamaged DNA. DNA curtain assays revealed that XPA preferentially binds DNA bubbles and searches for bubble structures through both one-dimensional diffusion and three-dimensional collision. Quantitative analysis showed that three-dimensional collision is the dominant pathway for bubble recognition. Taken together, our results uncover a previously unrecognized limitation of PEG-coated QDs in single-molecule studies and provide an improved surface-engineering strategy to preserve native protein-DNA interactions. Newly engineered QDs establish robust platforms for accurate single-molecule visualization of biomolecular processes. Quantum dots (QDs) are widely used as fluorescent probes for single-molecule imaging because they are brighter and more photostable than conventional fluorescent dyes. Researchers at UNIST have now discovered that the surface chemistry of many commercially available QDs can unintentionally disrupt the protein–DNA interactions they are intended to observe. The team also established a surface-engineering strategy that minimizes this interference without compromising their imaging performance. Led by Professor Jaeil Lee of the Department of Biological Sciences and Professor Jongnam Park of the Department of Energy and Chemical Engineering, the researchers identified polyethylene glycol (PEG), a polymer widely used to coat commercial QDs, as the source of the interference. When PEG occupied too great a proportion of the particle surface, protein–DNA interactions became destabilized, revealing a previously unrecognized limitation of a widely used imaging probe. The finding arose while the team was investigating xeroderma pigmentosum complementation group A (XPA), a DNA repair protein involved in nucleotide excision repair. As the concentration of commercially available QDs increased, XPA unexpectedly dissociated from DNA. The same behavior persisted regardless of quantum dot size, antibody type, surface modification, buffer conditions, or DNA structure, indicating that the disruption originated from the probe itself rather than the biological system. To address this limitation, the researchers systematically redesigned the quantum dot surface by adjusting the proportions of anchoring, hydrophilic, and PEG-derived functional groups. They found that reducing the PEG-derived surface composition below approximately 7% preserved native protein–DNA interactions while maintaining the brightness and colloidal stability required for single-molecule imaging. Using the redesigned QDs, the researchers visualized how XPA searches for sites of DNA damage without perturbing its natural behavior. Single-molecule DNA curtain experiments showed that XPA locates DNA bubble structures through both one-dimensional diffusion along DNA and three-dimensional collisions from solution. Quantitative analysis revealed that three-dimensional collisions are the dominant pathway for recognizing damaged DNA under physiologically relevant protein concentrations. “Commercial quantum dots have long been regarded as passive fluorescent probes,” said Professor Lee. "Our study shows that their surface chemistry can directly influence the molecular interactions being observed. By identifying the source of this interference, we established a practical strategy for preserving native protein–DNA interactions during single-molecule imaging." Professor Park added, "Careful control of surface polymer composition allowed us to reduce molecular interference without compromising brightness or colloidal stability. We expect these design principles to support the development of more reliable nanoparticle-based probes for biological imaging." The study also included contributions from Youngseo Kim of the Department of Biological Sciences and Hyerim Kim of the Department of Energy and Chemical Engineering at UNIST. The findings were published online in Nano Convergence on June 3, 2026. The research was supported by the National Research Foundation of Korea (NRF), the Institute for Basic Science (IBS), and the Ministry of Trade, Industry and Energy (MOTIE). Journal Reference Youngseo Kim, Hyerim Kim, Munryul Choi, et al ., "Engineering quantum dot surfaces to preserve protein-DNA interactions for single-molecule visualization," Nano Converg ., (2026).

2026.07.23

New Wrinkle Architecture Broadens Optical Responses for Advanced Security Films

Abstract Wrinkle-based diffraction gratings offer a scalable route to pigment-free structural coloration, yet most platforms are limited by globally aligned, linear wrinkles that provide limited spatial programmability and narrow angular selectivity. Here, we report geometrically programmed wrinkle architectonics in a UV-curable chitosan (UVCC)/polydimethylsiloxane (PDMS) bilayer, where lithography-defined, “geometric voids” serve as boundary constraints and stress concentrators to deterministically steer wrinkle trajectories during mechanical deformation. Simple geometric primitives effectively reshape the local stress field through geometrically incompatible confinement, transforming 1D gratings into shape-specific, spatially bifurcated domains characterized by curvilinear orientation and straight-wrinkle fields. We quantify these trajectory fields using the curvature metric and map critical design levers, such as inter-pattern spacing, UVCC thickness, and pattern asymmetry, that modulate curvature magnitude and steering range without altering diffraction periodicity. Circular perimeters sustain extended semi-circular trajectories, whereas polygonal vertices concentrate stress to induce aggressive local curvature. This programmed orientation dramatically amplifies angular dispersion: symmetric circular arrays achieve peak viewing-angle responsivity nearly three times higher than that of unpatterned thin films. A macaw security motif confirms wide-angle azimuthal readability and the formation of device-unique optical fingerprints for anti-counterfeiting. This work transforms a ubiquitous mechanical instability into a geometry-addressable photonic platform for high-level authentication and advanced displays. Hidden optical images that appear only when a transparent film is bent could provide a new generation of anti-counterfeiting technologies. Researchers at UNIST have demonstrated a strategy for programming the microscopic wrinkle patterns responsible for structural color, enabling optical films with enhanced viewing-angle sensitivity and unique optical fingerprints. Led by Professor Taesung Kim of the Department of Mechanical Engineering, the researchers discovered that simple geometric patterns can precisely control how microscopic wrinkles form as the film bends. By reshaping local mechanical stresses, the patterned boundaries guide wrinkle growth into both straight and curved domains while preserving the regular spacing responsible for structural color. This independent control over wrinkle orientation expands the film's optical capabilities. The enhanced wrinkle architecture translated directly into a broader optical response. The engineered films required only seven degrees of viewing-angle change to display the full visible color spectrum, compared with approximately thirty degrees for conventional wrinkle-based films. Distinct color changes also appeared with angular shifts as small as one degree, allowing substantially more optical information to be encoded within the same viewing range. To demonstrate these capabilities, the team fabricated a transparent security film featuring a hidden macaw image that appeared only when the film was bent. Unlike conventional wrinkle-based security films, which are visible only over a narrow range of viewing angles, the new design maintains clear image visibility and vivid structural colors across viewing angles from 0° to 90°. Each film also developed subtle variations in wrinkle branching, creating device-specific optical fingerprints that provide an additional layer of authentication beyond the visible security image. The films also showed excellent mechanical durability, maintaining stable wrinkle patterns and optical performance after 500 repeated bending cycles. “Instead of treating wrinkles as an unavoidable mechanical instability, we showed that they can be programmed through geometric design,” said Professor Kim. “By controlling wrinkle orientation while preserving wrinkle periodicity, we significantly broadened the optical response of wrinkle-based structural color films.” He added, "The approach could be applied to anti-counterfeiting technologies for banknotes, identification cards, luxury goods, and pharmaceutical packaging, while also providing a versatile platform for optical sensors and flexible display technologies." The findings were published in Advanced Functional Materials on June 25, 2026. Dr. Kaliannan Thiyagarajan and Sungjoon Ji from UNIST served as co-first authors of the study. The research was supported by the National Research Foundation of Korea (NRF) and UNIST. Journal Reference Kaliannan Thiyagarajan, Sungjoon Ji, Jiseok Han, et al ., “Geometrically Programmed Wrinkle Architectonics for High Angle-Dependent Structural Coloration,” Adv. Funct. Mater. , (2026).

2026.07.22

Quantum Algorithm Cuts the Cost of Training Robust AI by 80%

Abstract Robust reinforcement learning (RRL) aims to tackle unexpected environmental changes by optimizing policies against the worst case. However, RRL remains impractical due to the cost of the Max-Min optimization, where it suffers from the exhaustive query complexity for finding the worst-case (dubbed 'Min') within the environmental uncertainty set U , ie, O (|U|). By viewing this via a lens of quantum perspective, we raise a pivotal question: If we can query from the environment with quantum superpositions, is it possible to accelerate the Max-Min optimization of RRL? Our answer is 'Yes'. Our method, called quantum robust inner minimization (QRIM), encodes the uncertainty set with quantum superposition and amplifies low-return cases, thus enabling RL for solving the robust (ie, worst-case) Bellman equation. Importantly, QRIM achieves a quadratic speed-up in query complexity without altering the outer RL pipeline, ie, O (|U|). Validated through classical simulations to real quantum hardware execution, QRIM learns more robust policies with quadratically reduced queries than classical RL. Artificial intelligence systems deployed in the real world must continue to make reliable decisions even when conditions differ from those encountered during training. Robust reinforcement learning (RRL) addresses this challenge by preparing AI agents for the worst-case scenarios they may face. However, identifying those worst-case conditions requires searching a large uncertainty space, making training increasingly expensive as the number of possible environmental variations grows. To this bottleneck, Professor Sung Whan Yoon of the Graduate School of Artificial Intelligence at UNIST, in collaboration with Professor Joongheon Kim of Korea University, developed Quantum Robust Inner Minimization (QRIM), a method that replaces only the inner optimization responsible for identifying worst-case environments, while leaving the overall reinforcement learning pipeline unchanged. QRIM exploits quantum superposition to evaluate many possible environmental conditions simultaneously rather than examining each one individually. By amplifying the worst-case scenarios through a quantum search routine, the method reduces the number of required environment queries from N to √N—a quadratic improvement in query complexity—without modifying the surrounding reinforcement learning procedure. Across a series of experiments, QRIM learned more robust policies while requiring only 20–30% of the computation used by conventional methods. The researchers validated the algorithm through both classical simulations and experiments on IBM's 127-qubit quantum processor, demonstrating that the quadratic speed-up remained effective even in the presence of hardware noise. “QRIM redesigns the most computationally demanding component of RRL—the search for worst-case environments—using a quantum algorithm,” said first author Hyun Kyu Lee. "Because only the inner optimization is replaced, existing reinforcement learning algorithms can be readily incorporate the quantum module, making the approach broadly applicable across robust reinforcement learning." “This study demonstrates how quantum computing can complement artificial intelligence by addressing a well-defined computational bottleneck,” said Professor Yoon. "As quantum hardware continues to advance, approaches like QRIM could help make RRL practical for applications that require reliable decision-making under uncertainty, including robotics and autonomous driving." The findings have been accepted for presentation at the International Conference on Machine Learning (ICML) 2026, one of the world's leading conferences on machine learning. Among the quantum AI papers accepted to ICML 2026, QRIM is the only study led by a Korean research institution. The research was supported by the National Research Foundation of Korea (NRF) and the Institute for Information & Communications Technology Planning & Evaluation (IITP) through programs funded by the Ministry of Science and ICT (MSIT). Journal Reference Hyun Kyu Lee, Joongheon Kim, and Sung Whan Yoon, “Quantum Robust Inner Minimization for Reinforcement Learning with Quadratic Speed-Up in Query Complexity,” ICML, (2026) .

2026.07.21

Understanding What AI Models Have Learned

Abstract The explosive growth of open-source model repositories has created a Model Jungle, where checkpoints are frequently shared without adequate documentation or metadata. While weight-space learning offers a pathway to identify and analyze these models directly from their parameters, processing full-scale weights is computationally prohibitive. Probing-based methods have emerged as a lightweight alternative, extracting permutation-equivariant representations via learnable probe vectors. However, existing probing methods are limited by a single-view design: they capture first-order structures but fail to encode the rich, higher-order correlation patterns inherent in row-column interactions. To bridge this gap, we introduce MVProbe, a multi-perspective probing framework that synthesizes first-order signals with interaction-aware (Gram-based) views. Our approach is theoretically grounded; We analyze the scaling laws of different probing orders to derive a principled standardization and fusion strategy that ensures balanced contributions from all branches. On the Model Jungle benchmark, MVProbe consistently outperforms the state-of-the-art ProbeX across diverse architectures, including discriminative backbones (ResNet, SupViT, MAE, DINO) and large-scale generative LoRA adapters (Stable Diffusion LoRA). Not every AI model comes with a manual. As open-source model repositories continue to grow, many checkpoints are shared with little information about how they were trained or what data they learned from. As a result, identifying suitable models for reuse—or tracing where they came from—is becoming increasingly difficult. To address this challenge, Professor Jaejun Yoo and his team at the UNIST Graduate School of Artificial Intelligence developed MVProbe (Multi-view Probe), a method that identifies what an AI model has learned by analyzing its internal parameters alone, without relying on metadata or executing the model itself. Conventional probing methods analyze model weights from only a single perspective, limiting their ability to distinguish between closely related models. MVProbe overcomes this limitation by examining neural network parameters from four complementary perspectives, capturing not only individual parameter values but also the relationships among them to build a more complete representation of a model. The researchers also showed why existing probing methods struggle to distinguish similar models, demonstrating mathematically that important information embedded in model parameters is overlooked when weights are analyzed from only a single perspective. By integrating multiple complementary views, MVProbe overcomes this limitation while remaining reliable regardless of which neural network layer is analyzed. The team evaluated MVProbe on the Model Jungle benchmark, a large collection of modern AI models spanning image recognition networks and large generative models. Across all evaluated architectures, the method consistently outperformed the previous state-of-the-art approach, improving identification accuracy by as much as 5.96 percentage points while maintaining stable performance across different network layers. Beyond making large model repositories easier to search and organize, MVProbe could support a wide range of applications, including model reuse, provenance analysis, copyright verification, and AI security auditing—all without requiring models to be executed. “As more pretrained AI models become publicly available, the ability to identify and reuse the right one is becoming increasingly important,” said Professor Yoo. "MVProbe makes it possible to search and manage large model repositories by reading the information already encoded in a model's parameters, without running the model itself." The study was co-first authored by Eunwoo Heo and Kyungguk Seo of the UNIST Graduate School of Artificial Intelligence. It has been accepted for presentation at the International Conference on Machine Learning (ICML) 2026. This work was supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP) and the National Research Foundation of Korea (NRF). Journal Reference Eunwoo Heo, Kyeongkook Seo, and Jaejun Yoo, "What Linear Probes Miss: Multi-View Probing for Weight-Space Learning,"

2026.07.20

New Material Enables More Precise Patterning of Organic Electronics

Abstract Direct photopatterning with photoactivated cross-linkers provides a powerful route to highly integrated, solution-processed electronics. We introduce Diazo-6Bx, a six-branched diazo-based cross-linker that combines high miscibility with organic semiconductors and efficient C–H insertion reactivity. Comparative studies with azide-functionalized cross-linkers (phenyl- and tetrafluorophenyl-based structural analogues, namely, Ph-6Bx and 6Bx, respectively) reveal that the fluorine-free structure of Diazo-6Bx and its superior compatibility with host materials enable highly uniform cross-linking. This results in high-fidelity photopatterning with steep taper angles, excellent pattern retention, and minimal line-width variation. In solution-processed organic thin-film transistors (OTFTs), Diazo-6Bx enables the fabrication of unit devices that outperform those based on 6Bx, as well as large-area, high-resolution transistor arrays and logic circuits with excellent operational stability and scalability. These results demonstrate that a careful cross-linker design will play a crucial role in patterning next-generation electronic materials in extremely high fidelity. Printing electronic circuits from solution offers a promising route to lightweight, flexible electronics, but producing fine circuit patterns without sacrificing device performance remains a persistent manufacturing challenge. Professor BongSoo Kim and his research team in the Department of Chemistry at UNIST, working with collaborators at Yonsei University, have developed a new photo-crosslinker that enables more precise patterning of organic semiconductor circuits while improving the stability of the resulting devices. The material, Diazo-6Bx, was designed to blend more uniformly with hydrocarbon-based organic semiconductors than existing photo-crosslinkers. During ultraviolet patterning, it forms cross-linked networks that preserve circuit geometry, producing cleaner edges and more consistent feature sizes. In photopatterning experiments, Diazo-6Bx reduced line-width variation from 10.3 μm to 2.5 μm—roughly one quarter that of a widely used conventional photo-crosslinker. Circuit edges also became significantly sharper, with taper angles increasing from 67.9° to 87°, approaching nearly vertical sidewalls. The improved material compatibility also reduced charge-trapping defects, resulting in more stable transistor performance during prolonged operation. Organic thin-film transistors fabricated using Diazo-6Bx maintained more stable electrical characteristics under prolonged operation than devices produced with conventional crosslinking materials. The researchers also demonstrated that both p-type and n-type organic semiconductors could be patterned sequentially using the same solvent process. Using the same process, the researchers fabricated an 84-transistor organic thin-film transistor (OTFT) array together with complementary NOT, NAND, and NOR logic gates, demonstrating the scalability of the approach for complex organic electronic circuits. “This photo-crosslinker improves both patterning precision and device stability,” said Professor BongSoo Kim of the Department of Chemistry at UNIST. "Its compatibility with both p-type and n-type organic semiconductors simplifies solution processing and could help simplify manufacturing of large-area flexible displays and wearable electronic devices." The study was conducted in collaboration with Professor Jeong Ho Cho of Yonsei University. The findings were published in ACS Nano on June 23, 2026. The research was supported by the Samsung Future Technology Incubation Program and the National Research Foundation of Korea (NRF) through funding from the Ministry of Science and ICT (MSIT). Journal Reference Hyobin Ham, In Cheol Kwak, Daeyeon Lee, et al. , “Diazo-6Bx, a Six-Branched Diazo Cross-Linker, Enables High-Fidelity Patterning for Solution-Processed Electronics with Stable Operation,” ACS Nano, (2026).

2026.07.16