Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment

Large language models (LLMs) are commonly aligned with human preferences using reinforcement learning from human feedback (RLHF). In this method, LLM policies are generally optimized through reward maximization with Kullback-Leibler (KL) divergence regularization of the reference policy. However, KL and its $f$-divergence variants only compare token probabilities at identical indices, failing to capture semantic similarity. We propose Wasserstein Policy Regularization (WPR), a semantic-aware regularization for the RLHF framework based on the entropy-regularized Wasserstein distance, which incorporates the geometry of the token space. The dual formulation of the distance expresses the regularization as penalty terms applied to the reward via optimal dual variables, which yield a tractable objective compatible with standard RL algorithms. Empirically, our method outperforms KL- and $f$-divergence-based baselines, demonstrating the benefits of semantic-aware policy distances for alignment. Our code is available at https://github.com/aailab-kaist/WPR.

ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control

Achieving robust, human-like whole-body control on humanoid robots for agile, contact-rich behaviors remains a central challenge, demanding heavy per-skill engineering and a brittle process of tuning controllers. We introduce ZEST (Zero-shot Embodied Skill Transfer), a streamlined motion-imitation framework that trains policies via reinforcement learning from diverse sources -- high-fidelity motion capture, noisy monocular video, and non-physics-constrained animation -- and deploys them to hardware zero-shot. ZEST generalizes across behaviors and platforms while avoiding contact labels, reference or observation windows, state estimators, and extensive reward shaping. Its training pipeline combines adaptive sampling, which focuses training on difficult motion segments, and an automatic curriculum using a model-based assistive wrench, together enabling dynamic, long-horizon maneuvers. We further provide a procedure for selecting joint-level gains from approximate analytical armature values for closed-chain actuators, along with a refined model of actuators. Trained entirely in simulation with moderate domain randomization, ZEST demonstrates remarkable generality. On Boston Dynamics' Atlas humanoid, ZEST learns dynamic, multi-contact skills (e.g., army crawl, breakdancing) from motion capture. It transfers expressive dance and scene-interaction skills, such as box-climbing, directly from videos to Atlas and the Unitree G1. Furthermore, it extends across morphologies to the Spot quadruped, enabling acrobatics, such as a continuous backflip, through animation. Together, these results demonstrate robust zero-shot deployment across heterogeneous data sources and embodiments, establishing ZEST as a scalable interface between biological movements and their robotic counterparts.

TwinPurify: Purifying gene expression data to reveal tumor-intrinsic transcriptional programs via self-supervised learning

Advances in single-cell and spatial transcriptomic technologies have transformed tumor ecosystem profiling at cellular resolution. However, large scale studies on patient cohorts continue to rely on bulk transcriptomic data, where variation in tumor purity obscures tumor-intrinsic transcriptional signals and constrains downstream discovery. Many deconvolution methods report strong performance on synthetic bulk mixtures but fail to generalize to real patient cohorts because of unmodeled biological and technical variation. Here, we introduce TwinPurify, a representation learning framework that adapts the Barlow Twins self-supervised objective, representing a fundamental departure from the deconvolution paradigm. Rather than resolving the bulk mixture into discrete cell-type fractions, TwinPurify instead learns continuous, high-dimensional tumor embeddings by leveraging adjacent-normal profiles within the same cohort as "background" guidance, enabling the disentanglement of tumor-specific signals without relying on any external reference. Benchmarked against multiple large cancer cohorts across RNA-seq and microarray platforms, TwinPurify outperforms conventional representation learning baselines like auto-encoders in recovering tumor-intrinsic and immune signals. The purified embeddings improve molecular subtype and grade classification, enhance survival model concordance, and uncover biologically meaningful pathway activities compared to raw bulk profiles. By providing a transferable framework for decontaminating bulk transcriptomics, TwinPurify extends the utility of existing clinical datasets for molecular discovery.

VIRO: Robust and Efficient Neuro-Symbolic Reasoning with Verification for Referring Expression Comprehension

Referring Expression Comprehension (REC) aims to localize the image region corresponding to a natural-language query. Recent neuro-symbolic REC approaches leverage large language models (LLMs) and vision-language models (VLMs) to perform compositional reasoning, decomposing queries 4 structured programs and executing them step-by-step. While such approaches achieve interpretable reasoning and strong zero-shot generalization, they assume that intermediate reasoning steps are accurate. However, this assumption causes cascading errors: false detections and invalid relations propagate through the reasoning chain, yielding high-confidence false positives even when no target is present in the image. To address this limitation, we introduce Verification-Integrated Reasoning Operators (VIRO), a neuro-symbolic framework that embeds lightweight operator-level verifiers within reasoning steps. Each operator executes and validates its output, such as object existence or spatial relationship, thereby allowing the system to robustly handle no-target cases when verification conditions are not met. Our framework achieves state-of-the-art performance, reaching 61.1% balanced accuracy across target-present and no-target settings, and demonstrates generalization to real-world egocentric data. Furthermore, VIRO shows superior computational efficiency in terms of throughput, high reliability with a program failure rate of less than 0.3%, and scalability through decoupled program generation from execution.

Where It Moves, It Matters: Referring Surgical Instrument Segmentation via Motion

Enabling intuitive, language-driven interaction with surgical scenes is a critical step toward intelligent operating rooms and autonomous surgical robotic assistance. However, the task of referring segmentation, localizing surgical instruments based on natural language descriptions, remains underexplored in surgical videos, with existing approaches struggling to generalize due to reliance on static visual cues and predefined instrument names. In this work, we introduce SurgRef, a novel motion-guided framework that grounds free-form language expressions in instrument motion, capturing how tools move and interact across time, rather than what they look like. This allows models to understand and segment instruments even under occlusion, ambiguity, or unfamiliar terminology. To train and evaluate SurgRef, we present Ref-IMotion, a diverse, multi-institutional video dataset with dense spatiotemporal masks and rich motion-centric expressions. SurgRef achieves state-of-the-art accuracy and generalization across surgical procedures, setting a new benchmark for robust, language-driven surgical video segmentation.

ProFit: Leveraging High-Value Signals in SFT via Probability-Guided Token Selection

Supervised fine-tuning (SFT) is a fundamental post-training strategy to align Large Language Models (LLMs) with human intent. However, traditional SFT often ignores the one-to-many nature of language by forcing alignment with a single reference answer, leading to the model overfitting to non-core expressions. Although our empirical analysis suggests that introducing multiple reference answers can mitigate this issue, the prohibitive data and computational costs necessitate a strategic shift: prioritizing the mitigation of single-reference overfitting over the costly pursuit of answer diversity. To achieve this, we reveal the intrinsic connection between token probability and semantic importance: high-probability tokens carry the core logical framework, while low-probability tokens are mostly replaceable expressions. Based on this insight, we propose ProFit, which selectively masks low-probability tokens to prevent surface-level overfitting. Extensive experiments confirm that ProFit consistently outperforms traditional SFT baselines on general reasoning and mathematical benchmarks.

Mikasa: A Character-Driven Emotional AI Companion Inspired by Japanese Oshi Culture

Recent progress in large language models and multimodal interaction has made it possible to develop AI companions that can have fluent and emotionally expressive conversations. However, many of these systems have problems keeping users satisfied and engaged over long periods. This paper argues that these problems do not come mainly from weak models, but from poor character design and unclear definitions of the user-AI relationship. I present Mikasa, an emotional AI companion inspired by Japanese Oshi culture-specifically its emphasis on long-term, non-exclusive commitment to a stable character-as a case study of character-driven companion design. Mikasa does not work as a general-purpose assistant or a chatbot that changes roles. Instead, Mikasa is designed as a coherent character with a stable personality and a clearly defined relationship as a partner. This relationship does not force exclusivity or obligation. Rather, it works as a reference point that stabilizes interaction norms and reduces the work users must do to keep redefining the relationship. Through an exploratory evaluation, I see that users describe their preferences using surface-level qualities such as conversational naturalness, but they also value relationship control and imaginative engagement in ways they do not state directly. These results suggest that character coherence and relationship definition work as latent structural elements that shape how good the interaction feels, without users recognizing them as main features. The contribution of this work is to show that character design is a functional part of AI companion systems, not just decoration. Mikasa is one example based on a specific cultural context, but the design principles-commitment to a consistent personality and clear relationship definition-can be used for many emotionally grounded AI companions.

The Confidence Dichotomy: Analyzing and Mitigating Miscalibration in Tool-Use Agents

Autonomous agents based on large language models (LLMs) are rapidly evolving to handle multi-turn tasks, but ensuring their trustworthiness remains a critical challenge. A fundamental pillar of this trustworthiness is calibration, which refers to an agent's ability to express confidence that reliably reflects its actual performance. While calibration is well-established for static models, its dynamics in tool-integrated agentic workflows remain underexplored. In this work, we systematically investigate verbalized calibration in tool-use agents, revealing a fundamental confidence dichotomy driven by tool type. Specifically, our pilot study identifies that evidence tools (e.g., web search) systematically induce severe overconfidence due to inherent noise in retrieved information, while verification tools (e.g., code interpreters) can ground reasoning through deterministic feedback and mitigate miscalibration. To robustly improve calibration across tool types, we propose a reinforcement learning (RL) fine-tuning framework that jointly optimizes task accuracy and calibration, supported by a holistic benchmark of reward designs. We demonstrate that our trained agents not only achieve superior calibration but also exhibit robust generalization from local training environments to noisy web settings and to distinct domains such as mathematical reasoning. Our results highlight the necessity of domain-specific calibration strategies for tool-use agents. More broadly, this work establishes a foundation for building self-aware agents that can reliably communicate uncertainty in high-stakes, real-world deployments.

ExposeAnyone: Personalized Audio-to-Expression Diffusion Models Are Robust Zero-Shot Face Forgery Detectors

Detecting unknown deepfake manipulations remains one of the most challenging problems in face forgery detection. Current state-of-the-art approaches fail to generalize to unseen manipulations, as they primarily rely on supervised training with existing deepfakes or pseudo-fakes, which leads to overfitting to specific forgery patterns. In contrast, self-supervised methods offer greater potential for generalization, but existing work struggles to learn discriminative representations only from self-supervision. In this paper, we propose ExposeAnyone, a fully self-supervised approach based on a diffusion model that generates expression sequences from audio. The key idea is, once the model is personalized to specific subjects using reference sets, it can compute the identity distances between suspected videos and personalized subjects via diffusion reconstruction errors, enabling person-of-interest face forgery detection. Extensive experiments demonstrate that 1) our method outperforms the previous state-of-the-art method by 4.22 percentage points in the average AUC on DF-TIMIT, DFDCP, KoDF, and IDForge datasets, 2) our model is also capable of detecting Sora2-generated videos, where the previous approaches perform poorly, and 3) our method is highly robust to corruptions such as blur and compression, highlighting the applicability in real-world face forgery detection.

RGBT-Ground Benchmark: Visual Grounding Beyond RGB in Complex Real-World Scenarios

Visual Grounding (VG) aims to localize specific objects in an image according to natural language expressions, serving as a fundamental task in vision-language understanding. However, existing VG benchmarks are mostly derived from datasets collected under clean environments, such as COCO, where scene diversity is limited. Consequently, they fail to reflect the complexity of real-world conditions, such as changes in illumination, weather, etc., that are critical to evaluating model robustness and generalization in safety-critical applications. To address these limitations, we present RGBT-Ground, the first large-scale visual grounding benchmark built for complex real-world scenarios. It consists of spatially aligned RGB and Thermal infrared (TIR) image pairs with high-quality referring expressions, corresponding object bounding boxes, and fine-grained annotations at the scene, environment, and object levels. This benchmark enables comprehensive evaluation and facilitates the study of robust grounding under diverse and challenging conditions. Furthermore, we establish a unified visual grounding framework that supports both uni-modal (RGB or TIR) and multi-modal (RGB-TIR) visual inputs. Based on it, we propose RGBT-VGNet, a simple yet effective baseline for fusing complementary visual modalities to achieve robust grounding. We conduct extensive adaptations to the existing methods on RGBT-Ground. Experimental results show that our proposed RGBT-VGNet significantly outperforms these adapted methods, particularly in nighttime and long-distance scenarios. All resources will be publicly released to promote future research on robust visual grounding in complex real-world environments.