Decoupling

SFT and RLVR represent two fundamental yet distinct paradigms for LLM post-training, each excelling in distinct dimensions. SFT expands knowledge breadth while RLVR enhances reasoning depth. Yet integrating these complementary strengths remains a formidable challenge. Sequential training can cause catastrophic forgetting, and joint optimization often suffers from severe gradient conflicts. We analyze SFT and RLVR through the lens of task vectors and reveal three structural properties behind these failures: a 30* magnitude disparity, 45* sign interference, and heterogeneous module-wise update distributions. These findings show SFT and RLVR are difficult to integrate directly, but they also suggest that the two paradigms modify partly complementary components of the model. Motivated by these observations, we propose Decoupled Test-time Synthesis (DoTS), a post-hoc framework allows SFT and RLVR checkpoints to be trained independently and synthesizes their capabilities only at inference time via task vector arithmetic, without updating model parameters. To reduce interference, DOTS applies selective sparsification with norm-preserving rescaling. It then uses Bayesian optimization on a small set of unlabeled queries to search for combination coefficients on the Pareto frontier of consistency and perplexity. Empirically, \ours matches or exceeds the performance of training-based SFT--RLVR integration methods across multiple mathematical reasoning benchmarks, incurring only $\sim$3% of the computational cost. When applied to stronger post-trained checkpoints, DOTS surpasses SOTA models and generalizes to out-of-domain benchmarks without re-tuning. Code is available at https://github.com/chaohaoyuan/DoTS.

Subword tokenization is an essential part of modern large language models (LLMs), yet its specific contributions to training efficiency and model performance remain poorly understood. In this work, we decouple the effects of subword tokenization by isolating them within a controlled byte-level pretraining pipeline. We formulate and test hypotheses across various dimensions, including sample throughput, vocabulary scaling, and the linguistic prior of subword boundaries. By simulating these effects in a byte-level setting, we refine our understanding of why subword models outperform raw byte models and offer insights to improve the pretraining of future byte-level and subword models. Specifically, our experiments highlight the critical role of increased training throughput and the integration of subword boundaries as either explicit priors or inductive biases.

In continual learning, the primary challenge is to learn new information without forgetting old knowledge. A common solution addresses this trade-off through regularization, penalizing changes to parameters critical for previous tasks. In most cases, this regularization term is directly added to the training loss and optimized with standard gradient descent, which blends learning and retention signals into a single update and does not explicitly separate essential parameters from redundant ones. As task sequences grow, this coupling can over-constrain the model, limiting forward transfer and leading to inefficient use of capacity. We propose a different approach that separates task learning from stability enforcement via operator splitting. The learning step focuses on minimizing the current task loss, while a proximal stability step applies a sparse regularizer to prune unnecessary parameters and preserve task-relevant ones. This turns the stability-plasticity into a negotiated update between two complementary operators, rather than a conflicting gradient. We provide theoretical justification for the splitting method on the continual-learning objective, and demonstrate that our proposed solver achieves state-of-the-art results on standard benchmarks, improving both stability and adaptability without the need for replay buffers, Bayesian sampling, or meta-learning components.

Digital subtraction angiography (DSA) in coronary imaging is fundamentally challenged by physiological motion, forcing reliance on raw angiograms cluttered with anatomical noise. Existing deep learning methods often produced images with two critical clinically unacceptable flaws: persistent boundary artifacts and a loss of native tissue grayscale fidelity that undermined diagnostic confidence. We propose a novel framework termed as CDSA-Net that for the first time explicitly decouples and jointly optimizes vascular structure preservation and realistic background restoration. CDSA-Net introduces two core innovations: (i) A hierarchical geometric prior guidance (HGPG) mechanism, embedded in our coronary structure extraction network (CSENet). It synergistically combines integrated geometric prior (IGP) with gated spatial modulation (GSM) and centerline-aware topology (CAT) loss supervision, ensuring structural continuity. (ii) An adaptive noise module (ANM) within our coronary background restoration network (CBResNet). Unlike standard restoration, ANM uniquely models the stochastic nature of clinical X-ray noise, bridging the domain gap to enable seamless background intensity estimation and the complete elimination of boundary artifacts. The final subtraction is obtained by removing the restored background from the raw angiogram. Quantitatively, it significantly outperformed state-of-the-art methods in vascular intensity correlation and perceptual quality. A 25.6% improvement in morphology assessment efficiency and a 42.9% gain in hemodynamic evaluation speed set a new benchmark for utility in interventional cardiology, while maintaining diagnostic results consistent with raw angiograms. The project code is available at https://github.com/DrThink-ai/CDSA-Net.

The paper presents an approach for learning antenna Radiation Patterns (RPs) of a pair of heterogeneous quadrotor Uncrewed Aerial Vehicles (UAVs) by calibration flight data. RPs are modeled either as a Spherical Harmonics series or as a weighted average over inducing samples. Linear regression of polynomial coefficients enables decoupling of independent UAVs' RPs from the observed joint gain. A synchronized calibration trajectory provides training and testing samples in an obstacle-free anechoic altitude. Evaluation on a real-world dataset demonstrates the feasibility of learning both radiation patterns, achieving 4.56 dB RMS extrapolation error. The proposed RP learning and decoupling can be exploited in rapid recalibration upon payload changes, thereby enabling precise autonomous path planning and swarm control in real-world applications where setup changes are expected.

With the global proliferation of Large Language Models (LLMs), cultural safety, defined as the ability to generate respectful and appropriate responses across diverse cultures, becomes critical for responsible AI deployment. However, existing research often treats cultural safety and cultural knowledge in isolation. It remains unclear whether cultural safety is grounded in understanding varying cultural knowledge to enable LLMs to adaptively yield respectful and appropriate responses across diverse cultures, which trigger ethical concerns in cross-cultural scenarios. In this work, we introduce AdaCultureSafe, a dataset designed to jointly evaluate cultural safety and knowledge. Through AdaCultureSafe, we reveal a critical insight: \textit{a significant decoupling exists between cultural safety and cultural knowledge proficiency. Although LLMs possess rich cultural knowledge, they fail to leverage it to improve cultural safety.} LLMs tend to rely on generic safety rather than safety based on culture-specific knowledge. Motivated by this, we propose a knowledge-grounded method that elicits internal cultural knowledge in LLMs during response generation. Experimental results demonstrate that our approach effectively improves cultural safety. Our work can aid better understanding the landscape of cultural safety of LLMs.

Multi-modal scene reconstruction integrating RGB and thermal infrared data is essential for robust environmental perception across diverse lighting and weather conditions. However, extending 3D Gaussian Splatting (3DGS) to multi-spectral scenarios remains challenging. Current approaches often struggle to fully leverage the complementary information of multi-modal data, typically relying on mechanisms that either tend to neglect cross-modal correlations or leverage shared representations that fail to adaptively handle the complex structural correlations and physical discrepancies between spectrums. To address these limitations, we propose ThermoSplat, a novel framework that enables deep spectral-aware reconstruction through active feature modulation and adaptive geometry decoupling. First, we introduce a Spectrum-Aware Adaptive Modulation that dynamically conditions shared latent features on thermal structural priors, effectively guiding visible texture synthesis with reliable cross-modal geometric cues. Second, to accommodate modality-specific geometric inconsistencies, we propose a Modality-Adaptive Geometric Decoupling scheme that learns independent opacity offsets and executes an independent rasterization pass for the thermal branch. Additionally, a hybrid rendering pipeline is employed to integrate explicit Spherical Harmonics with implicit neural decoding, ensuring both semantic consistency and high-frequency detail preservation. Extensive experiments on the RGBT-Scenes dataset demonstrate that ThermoSplat achieves state-of-the-art rendering quality across both visible and thermal spectrums.

Distracted driving is a major cause of traffic collisions, calling for robust and scalable detection methods. Vision-language models (VLMs) enable strong zero-shot image classification, but existing VLM-based distracted driver detectors often underperform in real-world conditions. We identify subject-specific appearance variations (e.g., clothing, age, and gender) as a key bottleneck: VLMs entangle these factors with behavior cues, leading to decisions driven by who the driver is rather than what the driver is doing. To address this, we propose a subject decoupling framework that extracts a driver appearance embedding and removes its influence from the image embedding prior to zero-shot classification, thereby emphasizing distraction-relevant evidence. We further orthogonalize text embeddings via metric projection onto Stiefel manifold to improve separability while staying close to the original semantics. Experiments demonstrate consistent gains over prior baselines, indicating the promise of our approach for practical road-safety applications. Code is available at https://github.com/mtakamichi/ZVL-DDD

Virtual Try-On (VTON) synthesizes realistic images of a person wearing a target garment, with broad applications in e-commerce and fashion. Diffusion-based dual-UNet methods achieve strong results but double the parameters by dedicating a separate network to garment conditioning. Spatial concatenation offers a simpler single-network alternative, yet both UNet- and DiT-based instantiations report that full fine-tuning is ineffective, and the community has settled for attention-only training. We ask: why does full fine-tuning fail, and can this be resolved? Through what is, to our knowledge, the first visualization study of dual-UNet reference network behavior, we identify a unifying insight: garment conditioning must be decoupled from the denoising process. Spatial concatenation violates this by embedding the garment within the denoising target, causing three conflicts: guidance leakage, gradient competition, and train-test discrepancy. We derive three design principles to restore this decoupling and implement them as a pure recipe atop a standard architecture with no modification. The resulting model, DeCo-VTON (860M params), achieves single-network state of the art, matching the dual-UNet state of the art at half the cost while being preferred in human evaluation.

Large Reasoning Models (LRMs) can exhibit step-by-step reasoning, reflection, and backtracking, but these behaviors are often unregulated, leading to overthinking. As a result, LRMs continue generating redundant reasoning even after reaching high-confidence conclusions. This increases inference cost and latency, limiting practical deployment. The root cause is the absence of an intrinsic mechanism to monitor the reasoning state and decide when to continue, backtrack, or stop. We propose MERA, a meta-cognitive reasoning framework that decouples reasoning from control to enable independent optimization of control strategies. MERA constructs high-quality reasoning-control supervision data via a takeover-based pipeline, and transforms long-horizon traces into structured reasoning-control alternating sequences for training. The model is trained with supervised fine-tuning to internalize the structured separation, and further optimized with Control-Segment Policy Optimization (CSPO), which combines segment-wise GRPO with control masking to focus learning on control segments. Experiments across reasoning benchmarks show that MERA improves both efficiency and accuracy.