In Korea, Nuclear Power Plants continually prepare and revise the final safety analysis report (FSAR), which is essential for predicting the source-term information of radioactive materials throughout their lifecycle. This report provides detailed information on the source-terms of Chalk River Unidentified Deposits (CRUD). The management of radioactive corrosion products, present on nuclear fuel surfaces and within the primary system, is crucial for ensuring equipment integrity and controlling worker exposure during operation. Unlike earlier studies that focused solely on the radioactive inventory of major components, this study established an integrated evaluation framework by coupling the CRUDTRAN, ORIGEN2, and PHITS codes to quantitatively assess the relative contribution of the CRUD source-term on the core support barrel (CSB) to worker doses, in comparison with the activation source-terms of the reactor pressure vessel (RPV). By calculating these specific source-terms and conducting subsequent radiation exposure assessments, this work underscores the critical significance of the CRUD source-term during the decommissioning phase. The comparative dose assessment revealed that in the absence of system decontamination, the radiation exposure contributed by the CRUD source-term exceeded that from the structural activation of the RPV across the range of deposition-fraction assumptions, from about 1.5 times under the minimum assumption to about 117 times under the maximum assumption. These quantitative findings indicate that CRUD is a major source of worker exposure during the decommissioning phase, which should not be neglected relative to the activation of internal components.
Although the integral configuration of the i-SMR eliminates large diameter primary coolant pipes, a loss-of-coolant accident (LOCA) involving smaller connecting lines can still compromise core cooling and progress to a severe accident if passive safety systems fail. Previous i-SMR severe accident studies have mainly relied on individual analysis codes, while code-to-code comparisons for other reactors have shown differences in predicted accident progression due to code-specific physical models. Therefore, this study compares CINEMA and MELCOR to examine the consistency of severe accident predictions and identify code-dependent differences under conservative i-SMR conditions. A 2-in modular makeup and purification system letdown line loss-of-coolant accident was selected, assuming unavailability of the passive auxiliary feedwater system and failure of all passive emergency core cooling system valves, while the passive containment cooling system operated as designed. Major geometric data, initial thermal-hydraulic conditions, and fission product release assumptions were harmonized for consistent comparison. Both codes predicted similar early thermal-hydraulic behavior, but differences appeared after core degradation began, particularly in the corium relocation process, lower plenum corium configuration, and fission product behavior. Despite these differences, both codes predicted that lower head and containment vessel integrity were maintained over 72 h, and that environmental releases of major volatile fission products were effectively limited by deposition and pool retention in the containment vessel. The results provide a valuable reference for identifying code-dependent phenomenological uncertainties in i-SMR severe accident assessment.
Vision-Language-Action (VLA) models have recently achieved notable progress in end-to-end autonomous driving by integrating perception, reasoning, and control within a unified multimodal framework. However, they often lack explicit modeling of temporal dynamics and global world consistency, which limits their foresight and safety. In contrast, world models can simulate plausible future scenes but generally struggle to reason about or evaluate the imagined future they generate. In this work, we present VLA-World, a simple yet effective VLA world model that unifies predictive imagination with reflective reasoning to improve driving foresight. VLA-World first uses an action-derived feasible trajectory to guide the generation of the next-frame image, capturing rich spatial and temporal cues that describe how the surrounding environment evolves. The model then reasons over this self-generated future imagined frame to refine the predicted trajectory, achieving higher performance and better interpretability. To support this pipeline, we curate nuScenes-GR-20K, a generative reasoning dataset derived from nuScenes, and employ a three-stage training strategy that includes pretraining, supervised fine-tuning, and reinforcement learning. Extensive experiments demonstrate that VLA-World consistently surpasses state-of-the-art VLA and world-model baselines on both planning and future-generation benchmarks. Project page: https://vlaworld.github.io
Long-context inference in large language models (LLMs) is constrained by the quadratic time and memory complexity of attention. Top- k sparse attention mitigates this bottleneck by restricting each query to attend to only the k most relevant keys, but its mathematical guarantees remain incomplete despite widespread use in practice (e.g., DeepSeek's DSA [1] and NSA [2], Kimi's MoBA [3]). We quantify the theoretical approximation gap induced by top-k sparse attention through the total variation (TV) distance tau sic TV(P; (P) over cap), where P and (P) over cap are respectively the full attention's and top-k sparse attention's softmax distributions (attention weights) over queries and keys. We first prove an exact identity connecting the TV distance to the reverse Kullback-Leibler (KL) divergence tau - 1-e(-KL((P) over cap parallel to P)). This equality explains why TV distance is a natural error metric when sparse attention is trained via KL( (P) over cap parallel to P), and it tightens classical Bretagnolle-Huber (B-H) inequality and Pinsker's inequality in information theory for top-k sparse attention. We then bound the attention output gap based on query-key term tau and value geometry terms diam(H,T) (V) and Var(P) (V). We further derive a theoretical bound for TV distance tau in terms of the boundary attention-score gap Delta k sic sk -sk+1 center dot Finally, we conduct numerical validation on Qwen3-8B to assess the tightness of our theoretical bounds.
Post training via GRPO has demonstrated remarkable effectiveness in improving the generation quality of flow-matching models. However, GRPO suffers from inherently low sample efficiency due to its on-policy training paradigm. To address this limitation, we present OP-GRPO, the first Off-Policy GRPO framework tailored for flow-matching models. First, we actively select high-quality trajectories and adaptively incorporate them into a replay buffer for reuse in subsequent training iterations. Second, to mitigate the distribution shift introduced by off-policy samples, we propose a sequence-level importance sampling correction that preserves the integrity of GRPO's clipping mechanism while ensuring stable policy updates. Third, we theoretically and empirically show that late denoising steps yield ill-conditioned off-policy ratios, and mitigate this by truncating trajectories at late steps. Across image and video generation benchmarks, OP-GRPO achieves comparable or superior performance to Flow-GRPO with only 34.2
