arXiv subjects

Subin Kim

Publications and source records attributed to Subin Kim.

At least 19 recordsLinked to original sources

arXiv2026

Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense

Despite advances in safety alignment, large language models remain vulnerable to continuously evolving jailbreaks. Existing fine-tuned safety classifiers cannot adapt to these evolving attacks, while adaptive memory-based guardrails tend to over-refuse benign queries that resemble stored attacks. We propose Membrane, a self-evolving guardrail built on Contrastive Safety Memory (CSM): each cell pairs the conditions for blocking a harmful query with those for permitting a superficially similar benign request. Without retraining, Membrane evolves CSM by distilling each harmful interaction and its benign counterpart into a contrastive cell indexed by the underlying attack strategy, so that one cell generalizes across topical variants of the same mechanism. At inference, retrieved cells serve as grounding context for precise safety decisions. Across model-level safety on HarmBench and agent-level safety on AgentHarm, Membrane achieves the highest F1 on all six modern jailbreak attacks. Notably, benign refusal on AgentHarm stays at 7-14%, well below the 28-85% range of prior guards. Memory cells also retain 87-88% F1 under cross-attack transfer and remain stable under memory poisoning.

cs.CR↗

arXiv2026

Contrastive Conceptor Activation Steering (COAST): Unlocking Vision-Language-Action Models through Hidden States

Vision-Language-Action (VLA) models leverage powerful perceptual priors from web-scale Vision-Language Model (VLM) pre-training, yet they remain surprisingly brittle in practice, frequently failing at simple robotic tasks. To mitigate this, we propose Contrastive Conceptor Activation Steering (COAST). COAST builds on the notion of a "conceptor", a linear operator that soft-projects data into the principal components of a target distribution. COAST uses conceptors to identify success-critical subspaces for a target robotic task from a few examples of success and failure rollouts. At inference time, it steers VLA latents into these identified success subspaces to improve task outcomes. Across three architecturally distinct neural policies (flow-matching VLA, autoregressive VLA, and Diffusion Policy), COAST improves absolute mean simulation and real-robot task success rate by over 20 and 40% respectively. The activation subspace geometry reveals that failure modes share substantial structure across tasks while success representations remain largely task-specific. When tasks share similar failure modes, this structure enables previously fitted conceptors to improve performance on new tasks without refitting. Ultimately, our results suggest that current VLAs retain substantial task-relevant knowledge in their latent representations, and that the action expert's decoding bottleneck could be mitigated by steering its residual stream toward task-relevant subspaces. COAST provides a lightweight, training-free path to unlocking these latent capabilities by steering the model towards its own "success" distributions.

cs.RO↗

arXiv2026

Magnetic fields in monoclinic $\alpha$-RuCl$_3$ reveal rhombohedral inclusions underlying apparent oscillations

The majority of research on $\alpha$-RuCl$_3$ has focused on applying in-plane magnetic fields to suppress antiferromagnetic order and induce a quantum spin liquid (QSL). However, this effort has been complicated by the materials temperature-dependent crystal structure and sensitivity to strain-induced stacking disorder, making interpretation of field-induced phenomena contentious. The crystal structure of $\alpha$-RuCl$_3$ has recently been clarified as a function of temperature and sample size, motivating a reassessment of its magnetic properties and connection to proposed spin-liquid signatures. Here, we show that the monoclinic structure can be isolated in nanogram-scale crystals, enabling the study of Kitaev physics in a new regime. We focus on a structurally well-defined monoclinic crystal at low temperature and perform high-resolution magnetotropic susceptibility measurements in several crystal planes. Mapping the AFM phase boundary versus temperature, field, and orientation, we find the monoclinic phase diagram closely resembles rhombohedral crystals but is systematically shifted to higher transition temperatures and critical fields. For $B \parallel a$, we observe a two-step suppression of AFM order, indicating an intermediate ordered phase analogous to the ZZ2 phase reported in rhombohedral samples. Our results show that transitions previously observed beyond the AFM regime under in-plane fields arise from multiple shifted AFM phase boundaries associated with monoclinic inclusions, rather than non-magnetic phases. These findings indicate that features attributed to a QSL are instead due to an incomplete transition from the high-temperature monoclinic to the low-temperature rhombohedral structure. They also highlight the role of structural symmetry and sample homogeneity in interpreting field-induced phenomena in $\alpha$-RuCl$_3$ and related two-dimensional quantum magnets.

cond-mat.str-el↗

arXiv2026

Verifying random matrix product states with autoregressive local measurements

Matrix product states (MPS) are a central language for one-dimensional quantum matter and a practical target for near-term quantum simulators and variational algorithms. Yet, while substantial effort has focused on preparing MPS with shallow circuits, scalable methods to \emph{verify} that a many-body device has actually produced the intended state remain underdeveloped. Direct fidelity estimation (DFE) relies only on local Pauli measurements, but in many-body settings it suffers an exponential classical overhead from the preprocessing needed to sample Pauli strings. We eliminate this obstacle by introducing an \emph{autoregressive} importance sampler that draws Pauli strings sequentially from efficiently computable conditional distributions, reducing the per-shot classical overhead to linear scaling in the number of qubits. We further develop a grouped extension that constructs qubit-wise commuting measurement settings via a \emph{sorting string} and simultaneously estimates the entire commuting group from a single setting, significantly reducing estimator variance while preserving efficient postprocessing. Our approach extends naturally to matrix product operators (MPO), enabling scalable verification of tensor-network states and observables in long one-dimensional quantum systems. We utilize random MPS as a natural benchmark for generic 1D entangled states.

quant-ph↗

arXiv2026

ExpGuard: LLM Content Moderation in Specialized Domains

With the growing deployment of large language models (LLMs) in real-world applications, establishing robust safety guardrails to moderate their inputs and outputs has become essential to ensure adherence to safety policies. Current guardrail models predominantly address general human-LLM interactions, rendering LLMs vulnerable to harmful and adversarial content within domain-specific contexts, particularly those rich in technical jargon and specialized concepts. To address this limitation, we introduce ExpGuard, a robust and specialized guardrail model designed to protect against harmful prompts and responses across financial, medical, and legal domains. In addition, we present ExpGuardMix, a meticulously curated dataset comprising 58,928 labeled prompts paired with corresponding refusal and compliant responses, from these specific sectors. This dataset is divided into two subsets: ExpGuardTrain, for model training, and ExpGuardTest, a high-quality test set annotated by domain experts to evaluate model robustness against technical and domain-specific content. Comprehensive evaluations conducted on ExpGuardTest and eight established public benchmarks reveal that ExpGuard delivers competitive performance across the board while demonstrating exceptional resilience to domain-specific adversarial attacks, surpassing state-of-the-art models such as WildGuard by up to 8.9% in prompt classification and 15.3% in response classification. To encourage further research and development, we open-source our code, data, and model, enabling adaptation to additional domains and supporting the creation of increasingly robust guardrail models.

cs.CL↗

arXiv2026

BankMathBench: A Benchmark for Numerical Reasoning in Banking Scenarios

Large language models (LLMs)-based chatbots are increasingly being adopted in the financial domain, particularly in digital banking, to handle customer inquiries about products such as deposits, savings, and loans. However, these models still exhibit low accuracy in core banking computations-including total payout estimation, comparison of products with varying interest rates, and interest calculation under early repayment conditions. Such tasks require multi-step numerical reasoning and contextual understanding of banking products, yet existing LLMs often make systematic errors-misinterpreting product types, applying conditions incorrectly, or failing basic calculations involving exponents and geometric progressions. However, such errors have rarely been captured by existing benchmarks. Mathematical datasets focus on fundamental math problems, whereas financial benchmarks primarily target financial documents, leaving everyday banking scenarios underexplored. To address this limitation, we propose BankMathBench, a domain-specific dataset that reflects realistic banking tasks. BankMathBench is organized in three levels of difficulty-basic, intermediate, and advanced-corresponding to single-product reasoning, multi-product comparison, and multi-condition scenarios, respectively. When trained on BankMathBench, open-source LLMs exhibited notable improvements in both formula generation and numerical reasoning accuracy, demonstrating the dataset's effectiveness in enhancing domain-specific reasoning. With tool-augmented fine-tuning, the models achieved average accuracy increases of 57.6%p (basic), 75.1%p (intermediate), and 62.9%p (advanced), representing significant gains over zero-shot baselines. These findings highlight BankMathBench as a reliable benchmark for evaluating and advancing LLMs' numerical reasoning in real-world banking scenarios.

cs.CL↗

arXiv2026

Sample thickness dependence of structural and magnetic properties in $\alpha$-RuCl$_3$

The layered transition metal trihalide $\alpha$-RuCl$_3$ has been studied extensively in recent years as a promising candidate for a proximate Kitaev quantum spin liquid state. In high quality samples, a complete structural transition from room-temperature C2/m to low-temperature R$\bar{3}$ is consistently observed, with a single magnetic transition to antiferromagnetic ordering at $\sim$7K. However, magnetic and physical properties have been shown to depend heavily on both sample size and sample quality, with small and damaged samples exhibiting incomplete structural transitions and multiple magnetic anomalies. Although large high quality samples have been well studied, an understanding of the features attributed to low quality or small sample size is limited. Here, we probe the structural and magnetic transitions of $\alpha$-RuCl$_3$ single crystal samples via magnetic susceptibility through a range of thickness, manipulated through careful mechanical exfoliation. We present a non-destructive protocol for exfoliating crystals and show success to 30 $\mu$m, where sample quality is observed to improve with successive cleaving. Higher temperature magnetic features at 10 K/12 K are found to emerge through cleaving, both with and without induced sample damage. In both cases, we link these additional magnetic features to a persistence of C2/m structure to the low-temperature regime.

cond-mat.mtrl-sci↗

arXiv2025

Rethinking Prompt Design for Inference-time Scaling in Text-to-Visual Generation

Achieving precise alignment between user intent and generated visuals remains a central challenge in text-to-visual generation, as a single attempt often fails to produce the desired output. To handle this, prior approaches mainly scale the visual generation process (e.g., increasing sampling steps or seeds), but this quickly leads to a quality plateau. This limitation arises because the prompt, crucial for guiding generation, is kept fixed. To address this, we propose Prompt Redesign for Inference-time Scaling, coined PRIS, a framework that adaptively revises the prompt during inference in response to the scaled visual generations. The core idea of PRIS is to review the generated visuals, identify recurring failure patterns across visuals, and redesign the prompt accordingly before regenerating the visuals with the revised prompt. To provide precise alignment feedback for prompt revision, we introduce a new verifier, element-level factual correction, which evaluates the alignment between prompt attributes and generated visuals at a fine-grained level, achieving more accurate and interpretable assessments than holistic measures. Extensive experiments on both text-to-image and text-to-video benchmarks demonstrate the effectiveness of our approach, including a 15% gain on VBench 2.0. These results highlight that jointly scaling prompts and visuals is key to fully leveraging scaling laws at inference-time. Visualizations are available at the website: https://subin-kim-cv.github.io/PRIS.

cs.CV↗

arXiv2025

Phonon Hall Viscosity and the Intrinsic Thermal Hall Effect of $\alpha$-RuCl$_3$

The thermal Hall effect has been observed in a wide variety of magnetic insulators, yet its origins remains controversial. While some studies attribute the effect to intrinsic mechanism, such as heat carriers with Berry curvature, others propose extrinsic mechanisms, such as heat carriers scattering off crystal defects. Even the nature of the heat carriers is unknown: magnons, phonons, and fractionalized spin excitations have all been proposed. Resolving these issues is essential for the study of quantum spin liquids, and particularly for $\alpha$-RuCl$_3$, where a quantized thermal Hall effect has been attributed to Majorana edge modes. Here, we use ultrasonic measurements of the acoustic Faraday effect to demonstrate that the phonons in $\alpha$-RuCl$_3$ have Hall viscosity -- a non-dissipative viscosity that rotates phonon polarizations and deflects phonon heat currents. We show that phonon Hall viscosity produces an intrinsic thermal Hall effect that quantitatively accounts for a significant fraction of the measured thermal Hall effect in $\alpha$-RuCl$_3$: the thermal Hall effect in $\alpha$-RuCl$_3$ is due to phonons \textit{and} it is intrinsic. More broadly, we demonstrate that the acoustic Faraday effect is a powerful tool for detecting phonon Hall viscosity and the associated phonon Berry curvature, offering a new way to uncover and study exotic states of matter that elude conventional experiments.

cond-mat.str-el↗

arXiv2025

FontAdapter: Instant Font Adaptation in Visual Text Generation

Text-to-image diffusion models have significantly improved the seamless integration of visual text into diverse image contexts. Recent approaches further improve control over font styles through fine-tuning with predefined font dictionaries. However, adapting unseen fonts outside the preset is computationally expensive, often requiring tens of minutes, making real-time customization impractical. In this paper, we present FontAdapter, a framework that enables visual text generation in unseen fonts within seconds, conditioned on a reference glyph image. To this end, we find that direct training on font datasets fails to capture nuanced font attributes, limiting generalization to new glyphs. To overcome this, we propose a two-stage curriculum learning approach: FontAdapter first learns to extract font attributes from isolated glyphs and then integrates these styles into diverse natural backgrounds. To support this two-stage training scheme, we construct synthetic datasets tailored to each stage, leveraging large-scale online fonts effectively. Experiments demonstrate that FontAdapter enables high-quality, robust font customization across unseen fonts without additional fine-tuning during inference. Furthermore, it supports visual text editing, font style blending, and cross-lingual font transfer, positioning FontAdapter as a versatile framework for font customization tasks.

cs.CV↗

arXiv2025

A Smooth Sea Never Made a Skilled SAILOR: Robust Imitation via Learning to Search

The fundamental limitation of the behavioral cloning (BC) approach to imitation learning is that it only teaches an agent what the expert did at states the expert visited. This means that when a BC agent makes a mistake which takes them out of the support of the demonstrations, they often don't know how to recover from it. In this sense, BC is akin to giving the agent the fish -- giving them dense supervision across a narrow set of states -- rather than teaching them to fish: to be able to reason independently about achieving the expert's outcome even when faced with unseen situations at test-time. In response, we explore learning to search (L2S) from expert demonstrations, i.e. learning the components required to, at test time, plan to match expert outcomes, even after making a mistake. These include (1) a world model and (2) a reward model. We carefully ablate the set of algorithmic and design decisions required to combine these and other components for stable and sample/interaction-efficient learning of recovery behavior without additional human corrections. Across a dozen visual manipulation tasks from three benchmarks, our approach SAILOR consistently out-performs state-of-the-art Diffusion Policies trained via BC on the same data. Furthermore, scaling up the amount of demonstrations used for BC by 5-10x still leaves a performance gap. We find that SAILOR can identify nuanced failures and is robust to reward hacking. Our code is available at https://github.com/arnavkj1995/SAILOR .

cs.LG↗

arXiv2025

MIRROR: Multimodal Cognitive Reframing Therapy for Rolling with Resistance

Recent studies have explored the use of large language models (LLMs) in psychotherapy; however, text-based cognitive behavioral therapy (CBT) models often struggle with client resistance, which can weaken therapeutic alliance. To address this, we propose a multimodal approach that incorporates nonverbal cues, which allows the AI therapist to better align its responses with the client's negative emotional state. Specifically, we introduce a new synthetic dataset, Mirror (Multimodal Interactive Rolling with Resistance), which is a novel synthetic dataset that pairs each client's statements with corresponding facial images. Using this dataset, we train baseline vision language models (VLMs) so that they can analyze facial cues, infer emotions, and generate empathetic responses to effectively manage client resistance. These models are then evaluated in terms of both their counseling skills as a therapist, and the strength of therapeutic alliance in the presence of client resistance. Our results demonstrate that Mirror significantly enhances the AI therapist's ability to handle resistance, which outperforms existing text-based CBT approaches. Human expert evaluations further confirm the effectiveness of our approach in managing client resistance and fostering therapeutic alliance.

cs.CV↗

arXiv2025

Tuning-Free Multi-Event Long Video Generation via Synchronized Coupled Sampling

While recent advancements in text-to-video diffusion models enable high-quality short video generation from a single prompt, generating real-world long videos in a single pass remains challenging due to limited data and high computational costs. To address this, several works propose tuning-free approaches, i.e., extending existing models for long video generation, specifically using multiple prompts to allow for dynamic and controlled content changes. However, these methods primarily focus on ensuring smooth transitions between adjacent frames, often leading to content drift and a gradual loss of semantic coherence over longer sequences. To tackle such an issue, we propose Synchronized Coupled Sampling (SynCoS), a novel inference framework that synchronizes denoising paths across the entire video, ensuring long-range consistency across both adjacent and distant frames. Our approach combines two complementary sampling strategies: reverse and optimization-based sampling, which ensure seamless local transitions and enforce global coherence, respectively. However, directly alternating between these samplings misaligns denoising trajectories, disrupting prompt guidance and introducing unintended content changes as they operate independently. To resolve this, SynCoS synchronizes them through a grounded timestep and a fixed baseline noise, ensuring fully coupled sampling with aligned denoising paths. Extensive experiments show that SynCoS significantly improves multi-event long video generation, achieving smoother transitions and superior long-range coherence, outperforming previous approaches both quantitatively and qualitatively.

cs.CV↗

arXiv2025

Multimodal Cognitive Reframing Therapy via Multi-hop Psychotherapeutic Reasoning

Previous research has revealed the potential of large language models (LLMs) to support cognitive reframing therapy; however, their focus was primarily on text-based methods, often overlooking the importance of non-verbal evidence crucial in real-life therapy. To alleviate this gap, we extend the textual cognitive reframing to multimodality, incorporating visual clues. Specifically, we present a new dataset called Multi Modal-Cognitive Support Conversation (M2CoSC), which pairs each GPT-4-generated dialogue with an image that reflects the virtual client's facial expressions. To better mirror real psychotherapy, where facial expressions lead to interpreting implicit emotional evidence, we propose a multi-hop psychotherapeutic reasoning approach that explicitly identifies and incorporates subtle evidence. Our comprehensive experiments with both LLMs and vision-language models (VLMs) demonstrate that the VLMs' performance as psychotherapists is significantly improved with the M2CoSC dataset. Furthermore, the multi-hop psychotherapeutic reasoning method enables VLMs to provide more thoughtful and empathetic suggestions, outperforming standard prompting methods.

cs.CL↗

arXiv2024

Neural network-based recognition of multiple nanobubbles in graphene

We present a machine learning method for swiftly identifying nanobubbles in graphene, crucial for understanding electronic transport in graphene-based devices. Nanobubbles cause local strain, impacting graphene's transport properties. Traditional techniques like optical imaging are slow and limited for characterizing multiple nanobubbles. Our approach uses neural networks to analyze graphene's density of states, enabling rapid detection and characterization of nanobubbles from electronic transport data. This method swiftly enumerates nanobubbles and surpasses conventional imaging methods in efficiency and speed. It enhances quality assessment and optimization of graphene nanodevices, marking a significant advance in condensed matter physics and materials science. Our technique offers an efficient solution for probing the interplay between nanoscale features and electronic properties in two-dimensional materials.

cond-mat.mes-hall↗

arXiv2024

Re-investigation of Moment Direction in a Kitaev Material $\alpha$-RuCl$_{3}$

We report X-ray diffraction and resonant elastic X-ray scattering (REXS) studies on two $\alpha$-RuCl$_{3}$ crystals with distinct magnetic transition temperatures: T$_{N}$=7.3K and 6.5K. We find that the sample with T$_{N}$=6.5K exhibits a high degree of structural twinning at low temperature, whereas the T$_{N}$=7.3K sample primarily comprises a single domain of R$\bar{3}$. Notwithstanding, both samples exhibit an identical zigzag magnetic structure, with magnetic moments pointing away from the honeycomb plane by $\alpha=31(2)^{\circ}$. We argue that the identical ordered moment directions in these samples suggest that the intralayer magnetic Hamiltonian remains mostly unchanged regardless of T$_{N}$.

cond-mat.str-el↗

arXiv2024

Dr. Strategy: Model-Based Generalist Agents with Strategic Dreaming

Model-based reinforcement learning (MBRL) has been a primary approach to ameliorating the sample efficiency issue as well as to make a generalist agent. However, there has not been much effort toward enhancing the strategy of dreaming itself. Therefore, it is a question whether and how an agent can "dream better" in a more structured and strategic way. In this paper, inspired by the observation from cognitive science suggesting that humans use a spatial divide-and-conquer strategy in planning, we propose a new MBRL agent, called Dr. Strategy, which is equipped with a novel Dreaming Strategy. The proposed agent realizes a version of divide-and-conquer-like strategy in dreaming. This is achieved by learning a set of latent landmarks and then utilizing these to learn a landmark-conditioned highway policy. With the highway policy, the agent can first learn in the dream to move to a landmark, and from there it tackles the exploration and achievement task in a more focused way. In experiments, we show that the proposed model outperforms prior pixel-based MBRL methods in various visually complex and partially observable navigation tasks.

cs.LG↗

arXiv2024

Static and fluctuating zigzag order, and possible signatures of Kitaev physics, in torque measurements of ${\alpha}$-RuCl${_3}$

We have measured magnetic torque on a $T_N=7$ K single crystal of $\alpha$-RuCl$_3$, as a function of the field angle in the $ab$-plane, focusing on temperatures between 2 and 20 K and fields from 0 to 9 T. We find a number of features, many of which can be classified by their angular periodicity. The sample shows an oscillation with a period of 180$^\circ$ (i.e.\ two-fold periodicity) and within the magnetically ordered zigzag phase there is a 60$^\circ$ period (i.e.\ six-fold) sawtooth pattern, which can be explained by reorientation of the zigzag domains as the crystal rotates in the applied field. We argue that the six-fold sawtooth and the two-fold sinusoidal signals arise from distinct regions of the crystal. Suppressing the zigzag order with an applied field above $\sim8$ T at low temperature, a six-fold {\sl sinusoidal} signal remains, suggesting that there is fluctuating zigzag order in the putative field-induced quantum spin liquid state. Finally, in testing theoretical results which predict a torque response with divergent slope across C$_2$-preserving $b$-axes (B1-axis), we find no features like that predicted for Ising topological order. Instead we find features at low temperatures and fields just above the zigzag phase across the non-C$_2$-preserving $b$-axes (B2-axes). Interpretation of this feature is complicated by the development of other similar signatures nearby at slightly lower fields, and by clear enhancement with thermal cycling. Additionally, we contrast the torque response of $T_N \sim$ 7 K and 14 K samples.

cond-mat.str-el↗