Large reasoning models that use long chainof-thought excel at problem-solving yet waste compute on redundant checks.Curbing this overthinking is hard: training-time length penalties can cripple ability, while inferencetime early-exit adds system overhead.To bridge this gap, we propose Step-GRPO, a novel post-training framework that internalizes dynamic early-exit capabilities directly into the model.Step-GRPO shifts the optimization objective from raw tokens to semantic steps by utilizing linguistic markers to structure reasoning.We introduce a Dynamic Truncated Rollout mechanism that exposes the model to concise high-confidence trajectories during exploration, synergized with a Step-Aware Relative Reward that dynamically penalizes redundancy based on group-level baselines.Extensive experiments across three model sizes on diverse benchmarks demonstrate that Step-GRPO achieves a superior accuracy-efficiency tradeoff.On Qwen3-8B, our method reduces token consumption by 32.0% compared to the vanilla model while avoiding the accuracy degradation observed in traditional length-penalty methods.

Yusong Hu, Runmin Ma, Yue Fan, Jinxin Shi, Zongsheng Cao, Yuhao Zhou, Jiakang Yuan, Shuaiyu Zhang, Shiyang Feng, Xiangchao Yan, Shufei Zhang, Wenlong Zhang, Lei Bai, Bo Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yiqun Zhang, Peng Ye, Xiaocui Yang, Shi Feng, Shufei Zhang, Lei Bai, Wanli Ouyang, Shuyue Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Nuclear reactor FSI simulation faces two key challenges: "Mapping wall" bottleneck in data transfer across non-matching mesh coupling interfaces; Low hardware utilization from multi-physics solvers’ heterogeneous core tasks (compute- vs. memory-intensive). Therefore, an innovative FSI framework integrating two strategies is proposed: Scalable radial basis function mapping—restructuring the global problem into massive independent subproblems via task partitioning, preallocation, and multi-granularity load balancing to eliminate communication overhead; Dependency-aware multi-stream optimization—deeply overlapping heterogeneous solver tasks to maximize hardware utilization. It first achieves parameter transfer across ∼90,000 non-matching coupling interfaces in China Experimental Fast Reactor, with 86.36% strong scaling and 94.01% weak scaling. The combined optimizations yield ∼60% performance gain, increase strong scaling by over 20 percentage points, and achieve high weak scaling of ∼97%. Moreover, the FSI results align well with publicly available data, verifying its correctness.

Large Language Models (LLMs) have achieved remarkable success and have been applied across various scientific fields, including chemistry. However, many chemical tasks require the processing of visual information, which cannot be successfully handled by existing chemical LLMs. This brings a growing need for models capable of integrating multimodal information in the chemical domain. In this paper, we introduce ChemVLM, an open-source chemical multimodal large language model specifically designed for chemical applications. ChemVLM is trained on a carefully curated bilingual multimodal dataset that enhances its ability to understand both textual and visual chemical information, including molecular structures, reactions, and chemistry examination questions. We develop three datasets for comprehensive evaluation, tailored to Chemical Optical Character Recognition (OCR), Multimodal Chemical Reasoning (MMCR), and Multimodal Molecule Understanding tasks. We benchmark ChemVLM against a range of open-source and proprietary multimodal large language models on various tasks. Experimental results demonstrate that ChemVLM achieves competitive performance across all evaluated tasks.

Text-to-image diffusion model is a popular paradigm that synthesizes personalized images by providing a text prompt and a random Gaussian noise. While people observe that some noises are ``golden noises'' that can achieve better text-image alignment and higher human preference than others, we still lack a machine learning framework to obtain those golden noises. To learn golden noises for diffusion sampling, we mainly make three contributions in this paper. First, we identify a new concept termed the \textit{noise prompt}, which aims at turning a random Gaussian noise into a golden noise by adding a small desirable perturbation derived from the text prompt. Following the concept, we first formulate the \textit{noise prompt learning} framework that systematically learns ``prompted'' golden noise associated with a text prompt for diffusion models. Second, we design a noise prompt data collection pipeline and collect a large-scale \textit{noise prompt dataset}~(NPD) that contains 100k pairs of random noises and golden noises with the associated text prompts. With the prepared NPD as the training dataset, we trained a small \textit{noise prompt network}~(NPNet) that can directly learn to transform a random noise into a golden noise. The learned golden noise perturbation can be considered as a kind of prompt for noise, as it is rich in semantic information and tailored to the given text prompt. Third, our extensive experiments demonstrate the impressive effectiveness and generalization of NPNet on improving the quality of synthesized images across various diffusion models, including SDXL, DreamShaper-xl-v2-turbo, and Hunyuan-DiT. Moreover, NPNet is a small and efficient controller that acts as a plug-and-play module with very limited additional inference and computational costs, as it just provides a golden noise instead of a random noise without accessing the original pipeline.

Di Zhang, Jianbo Wu, Jingdi Lei, Tong Che, Jiatong Li, Tong Xie, Xiaoshui Huang, Shufei Zhang, Marco Pavone, Yuqiang Li, Wanli Ouyang, Dongzhan Zhou. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Retrosynthetic planning is a fundamental task in chemical discovery. Due to the vast combinatorial search space, identifying viable synthetic routes remains a significant challenge--even for expert chemists. Recent advances in Large Language Models (LLMs), particularly equipped with reinforcement learning, have demonstrated strong human-like reasoning and planning abilities, especially in mathematics and code problem solving. This raises a natural question: Can the reasoning capabilities of LLMs be harnessed to develop an AI chemist capable of learning effective policies for multi-step retrosynthesis? In this study, we introduce Retro-R1, a novel LLM-based retrosynthesis agent trained via reinforcement learning to design molecular synthesis pathways. Unlike prior approaches, which typically rely on single-turn, question-answering formats, Retro-R1 interacts dynamically with plug-in single-step retrosynthesis tools and learns from environmental feedback. Experimental results show that Retro-R1 achieves a 55.79\% pass@1 success rate, surpassing the previous state of the art by 8.95\%. Notably, Retro-R1 demonstrates strong generalization to out-of-domain test cases, where existing methods tend to fail despite their high in-domain performance. Our work marks a significant step toward equipping LLMs with advanced, chemist-like reasoning abilities, highlighting the promise of reinforcement learning for enabling data-efficient, generalizable, and sophisticated scientific problem-solving in LLM-based agents.

Recent studies explore how to obtain adversarial robustness for unsupervised domain adaptation (UDA). These efforts are however dedicated to achieving an optimal trade-off between accuracy and robustness on a given or seen target domain but ignore the robust generalization issue over unseen adversarial data. Consequently, degraded performance will be often observed when existing robust UDAs are applied to future adversarial data. In this work, we make a first attempt to address the robust generalization issue of UDA. We conjecture that the poor robust generalization of present robust UDAs may be caused by the large distribution gap among adversarial examples. We then provide an empirical and theoretical analysis showing that this large distribution gap is mainly owing to the discrepancy between feature-shift distributions. To reduce such discrepancy, a novel Anchored Feature-Shift Regularization (AFSR) method is designed with a certificated robust generalization bound. We conduct a series of experiments on benchmark UDA datasets. Experimental results validate the effectiveness of our proposed AFSR over many existing robust UDA methods.

The latest heuristic for handling the domain shift in un-supervised domain adaptation tasks is to reduce the data distribution discrepancy using adversarial learning. Recent studies improve the conventional adversarial domain adaptation methods with discriminative information by integrating the classifier’s outputs into distribution divergence measurement. However, they still suffer from the equilibrium problem of adversarial learning in which even if the discriminator is fully confused, sufficient similarity between two distributions cannot be guaranteed. To overcome this problem, we propose a novel approach named feature gradient distribution alignment (FGDA)1. We demonstrate the rationale of our method both theoretically and empirically. In particular, we show that the distribution discrepancy can be reduced by constraining feature gradients of two domains to have similar distributions. Meanwhile, our method enjoys a theoretical guarantee that a tighter error upper bound for target samples can be obtained than that of conventional adversarial domain adaptation methods. By integrating the proposed method with existing adversarial domain adaptation models, we achieve state-of-the-art performance on two real-world benchmark datasets.

While adversarial training becomes one of the most promising defending approaches against adversarial attacks for deep neural networks, the conventional wisdom through robust optimization may usually not guarantee good generalization for robustness. Concerning with robust generalization over unseen adversarial data, this paper investigates adversarial training from a novel perspective of shift consistency in latent space. We argue that the poor robust generalization of adversarial training is owing to the significantly dispersed latent representations generated by training and test adversarial data, as the adversarial perturbations push the latent features of natural examples in the same class towards diverse directions. This is underpinned by the theoretical analysis of the robust generalization gap, which is upper-bounded by the standard one over the natural data and a term of feature inconsistent shift caused by adversarial perturbation {–} a measure of latent dispersion. Towards better robust generalization, we propose a new regularization method {–} shift consistency regularization (SCR) {–} to steer the same-class latent features of both natural and adversarial data into a common direction during adversarial training. The effectiveness of SCR in adversarial training is evaluated through extensive experiments over different datasets, such as CIFAR-10, CIFAR-100, and SVHN, against several competitive methods.

Deep neural network with multi-scale feature fusion has achieved great success in human pose estimation. However, drawbacks still exist in these methods: 1) they consider multi-scale features equally, which may over-emphasize redundant features; 2) preferring deeper structures, they can learn features with the strong semantic representation, but tend to lose natural discriminative information; 3) to attain good performance, they rely heavily on pretraining, which is time-consuming, or even unavailable practically. To mitigate these problems, we propose a novel comprehensive recalibration model called Pyramid GAting Network (PGA-Net) that is capable of distillating, selecting, and fusing the discriminative and attention-aware features at different scales and different levels (i.e., both semantic and natural levels). Meanwhile, focusing on fusing features both selectively and comprehensively, PGA-Net can demonstrate remarkable stability and encouraging performance even without pre-training, making the model can be trained truly from scratch. We demonstrate the effectiveness of PGA-Net through validating on COCO and MPII benchmarks, attaining new state-of-the-art performance. https://github.com/ssr0512/PGA-Net

Adversarial examples, referred to as augmented data points generated by imperceptible perturbations of input samples, have recently drawn much attention. Well-crafted adversarial examples may even mislead state-of-the-art deep neural network (DNN) models to make wrong predictions easily. To alleviate this problem, many studies have focused on investigating how adversarial examples can be generated and/or effectively handled. All existing works tackle this problem in the Euclidean space. In this paper, we extend the learning of adversarial examples to the more general Riemannian space over DNNs. The proposed work is important in that (1) it is a generalized learning methodology since Riemmanian space will be degraded to the Euclidean space in a special case; (2) it is the first work to tackle the adversarial example problem tractably through the perspective of Riemannian geometry; (3) from the perspective of geometry, our method leads to the steepest direction of the loss function, by considering the second order information of the loss function. We also provide a theoretical study showing that our proposed method can truly find the descent direction for the loss function, with a comparable computational time against traditional adversarial methods. Finally, the proposed framework demonstrates superior performance over traditional counterpart methods, using benchmark data including MNIST, CIFAR-10 and SVHN.