Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a powerful paradigm for enhancing Large Language Models (LLMs), exemplified by the success of OpenAI's o-series. In RLVR, rewards are derived from verifiable signals-such as passing unit tests in code generation or matching correct answers in mathematical reasoning. While effective, this requirement largely confines RLVR to domains with automatically checkable outcomes. To overcome this, we extend the RLVR paradigm to open-ended tasks by integrating rubric-based rewards, where carefully designed rubrics serve as structured, model-interpretable criteria for automatic scoring of subjective outputs. We construct, to our knowledge, the largest rubric reward system to date, with over 10,000 rubrics from humans, LLMs, or a hybrid human-LLM collaboration. Implementing rubric-based RL is challenging; we tackle these issues with a clear framework and present an open-sourced Qwen-30B-A3B model with notable gains: 1) With only 5K+ samples, our system improves by +5.2% on open-ended benchmarks (especially humanities), outperforming a 671B DeepSeek-V3 model by +2.4%, while preserving general and reasoning abilities. 2) Our method provides fine-grained stylistic control, using rubrics as anchors to mitigate the "AI-like" tone and produce more human-like, expressive responses. We share key lessons in rubric construction, data selection, and training, and discuss limitations and future releases.

Large language models (LLMs) have shown remarkable capabilities in isolated step-by-step reasoning tasks such as mathematics and programming, but their proficiency in long-horizon planning, where solutions require extended, structured sequences of interdependent actions, remains underex-plored. Existing benchmarks typically assess LLMs through abstract or low-dimensional algorithmic tasks, failing to capture the complexity of realistic planning environments. We introduce HeroBench, a novel benchmark designed specifically to evaluate long-horizon planning and structured reasoning within complex RPG-inspired virtual worlds. HeroBench provides a rigorously constructed dataset of tasks covering a wide range of difficulties, a simulated environment to execute and validate agent plans, and detailed analytical tools for evaluating model performance. Tasks challenge models to formulate strategic plans, efficiently gather resources, master necessary skills, craft equipment, and defeat adversaries, reflecting practical scenarios' layered dependencies and constraints. Our extensive evaluation of 25 state-of-the-art LLMs, spanning both open-source and proprietary models, including the GPT -5 family, reveals substantial performance disparities rarely observed in conventional reasoning benchmarks. Detailed error analysis further uncovers specific weaknesses in current models' abilities to generate robust high-level plans and reliably execute structured actions. HeroBench thus not only significantly advances the evaluation of LLM reasoning but also provides a flexible, scalable foundation for future research into advanced, autonomous planning in virtual environments.

Medical large vision-language Models (Med-L VLMs) have shown promise in clinical applications but suffer from factual inaccuracies and unreliable outputs, posing risks in real-world diagnostics. While retrieval-augmented generation has emerged as a potential solution, current medical multimodal RAG systems are unable to perform effective retrieval across heterogeneous sources. The irrelevance of retrieved reports affects the factuality of analysis, while insufficient knowledge affects the credibility of clinical decision-making. To bridge the gap, we construct MedAtlas, which includes extensive multimodal report repositories and diverse text corpora. Based on it, we present HeteroRAG, a novel framework that enhances Med-L VLMs through heterogeneous knowledge sources. The framework introduces Modality-specific CLIPs for effective report retrieval and a Multi-corpora Query Generator for dynamically constructing queries for diverse corpora. Incorporating knowledge from such multifaceted sources, Med-L VLM is then trained with Heterogeneous Knowledge Preference Tuning to achieve cross-modality and multi-source knowledge alignment. Extensive experiments across 12 datasets and 3 modalities demonstrate that the proposed HeteroRAG achieves state-of-the-art performance in most medical vision language benchmarks, significantly improving factual accuracy and reliability of Med-L VLMs.

In the swiftly evolving domain of cloud computing, the advent of serverless systems underscores the crucial need for predictive auto-scaling systems. This necessity arises to ensure optimal resource allocation and maintain operational efficiency in inherently volatile environments. At the core of a predictive auto-scaling system is the workload forecasting model. Existing forecasting models struggle to quickly adapt to the dynamics in online workload streams and have difficulty capturing the complex periodicity brought by fine-grained, high-frequency forecasting tasks. Addressing this, we propose a novel online ensemble model, E3Former, for online workload forecasting in large-scale predictive auto-scaling. Our model synergizes the predictive capabilities of multiple subnetworks to surmount the limitations of single-model approaches, thus ensuring superior accuracy and robustness. Remarkably, it accomplishes this with a minimal increase in computational overhead, adhering to the lean operational ethos of serverless systems. Through extensive experimentation on real-world workload datasets, we establish the efficacy of our ensemble model. In online forecasting tasks, the proposed method reduces forecast error by an average of 10%, and its effectiveness is further demonstrated through a predictive auto-scaling test in the real-life online system. Currently, our method has been deployed within ByteDance's Intelligent Horizontal Pod Auto-scaling (IHPA) platform, which supports the stable operation of over 30 applications, such as Douyin E-Comerce, TouTiao, and Volcano Engine. The predictive auto-scaling capacity reaching over 600,000 CPU cores. On the basis of essentially ensuring service quality, the predictive auto-scaling system can reduce resource utilization by over 40%.

Labeling Cadmium Zinc Telluride (CdZnTe) semiconductor images is challenging due to the low-contrast defect boundaries, necessitating annotators to cross-reference multiple views. These views share a single ground truth (GT), forming a unique ``many-to-one'' relationship. This characteristic renders advanced semi-supervised semantic segmentation (SSS) methods suboptimal, as they are generally limited by a ``one-to-one'' relationship, where each image is independently associated with its GT. Such limitation may lead to error accumulation in low-contrast regions, further exacerbating confirmation bias. To address this issue, we revisit the SSS pipeline from a group-oriented perspective and propose a human-inspired solution: the Intra-group Consistency Augmentation Framework (ICAF). First, we experimentally validate the inherent consistency constraints within CdZnTe groups, establishing a group-oriented baseline using the Intra-group View Sampling (IVS). Building on this insight, we introduce the Pseudo-label Correction Network (PCN) to enhance consistency representation, which consists of two key modules. The View Augmentation Module (VAM) improves boundary details by dynamically synthesizing a boundary-aware view through the aggregation of multiple views. In the View Correction Module (VCM), this synthesized view is paired with other views for information interaction, effectively emphasizing salient regions while minimizing noise. Extensive experiments demonstrate the effectiveness of our solution for CdZnTe materials. Leveraging DeepLabV3+ with a ResNet-101 backbone as our segmentation model, we achieve a 70.6\% mIoU on the CdZnTe dataset using only 2 group-annotated data (5\textperthousand). The code is available at \href{https://github.com/pipixiapipi/ICAF}{https://github.com/pipixiapipi/ICAF}.

The recent rise in popularity of large language models (LLMs) has prompted considerable concerns about their moral capabilities. Although considerable effort has been dedicated to aligning LLMs with human moral values, existing benchmarks and evaluations remain largely superficial, typically measuring alignment based on final ethical verdicts rather than explicit moral reasoning. In response, this paper aims to advance the investigation of LLMs' moral capabilities by examining their capacity to function as Artificial Moral Assistants (AMAs), systems envisioned in the philosophical literature to support human moral deliberation. We assert that qualifying as an AMA requires more than what state-of-the-art alignment techniques aim to achieve: not only must AMAs be able to discern ethically problematic situations, they should also be able to actively reason about them, navigating between conflicting values outside of those embedded in the alignment phase. Building on existing philosophical literature, we begin by designing a new formal framework of the specific kind of behaviour an AMA should exhibit, individu-ating key qualities such as deductive and abductive moral reasoning. Drawing on this theoretical framework, we develop a benchmark to test these qualities and evaluate popular open LLMs against it. Our results reveal considerable variability across models and highlight persistent shortcomings, particularly regarding abductive moral reasoning. Our work connects theoretical philosophy with practical AI evaluation while also emphasising the need for dedicated strategies to explicitly enhance moral reasoning capabilities in LLMs.

Image set classification (ISC), which can be viewed as a task of comparing similarities between sets consisting of unordered heterogeneous images with variable quantities and qualities, has attracted growing research attention in recent years. How to learn effective feature representations and how to explore the similarities between different image sets are two key yet challenging issues in this field. However, existing traditional ISC methods classify image sets based on raw pixel features, ignoring the importance of feature learning. Existing deep ISC methods can learn deep features, but they fail to adaptively adjust the features when measuring set distances, resulting in limited performance in few-shot ISC. To address the above issues, this paper combines traditional ISC methods with deep models and proposes a novel few-shot ISC approach called Deep Class-specific Collaborative Representation (DCSCR) network to simultaneously learn the frame- and concept-level feature representations of each image set and the distance similarities between different sets. Specifically, DCSCR consists of a fully convolutional deep feature extractor module, a global feature learning module, and a class-specific collaborative representation-based metric learning module. The deep feature extractor and global feature learning modules are used to learn (local and global) frame-level feature representations, while the class-specific collaborative representation-based metric learning module is exploit to adaptively learn the concept-level feature representation of each image set and thus obtain the distance similarities between different sets by developing a new CSCR-based contrastive loss function. Extensive experiments on several well-known few-shot ISC datasets demonstrate the effectiveness of the proposed method compared with some state-of-the-art image set classification algorithms.

Y uanxin Wei 1 Lansong Diao 2 Bujiao Chen 2 Shenggan Cheng 3 Zhengping Qian 2 Wenyuan Y u 2 Nong Xiao 1 Wei Lin 2 Jiangsu Du 1 1 Sun Y at-sen University 2 Alibaba Group 3 National University of Singapore Corresponding authors Abstract Leveraging the Transformer architecture and the diffusion process, video DiT models have emerged as a dominant approach for high-quality video generation. Caching, a widely adopted optimization method in DiT models, leverages the redundancy in the diffusion process to skip computations in different granularities (e.g., step, cfg, block). Nevertheless, existing caching methods are limited to single-granularity strategies, struggling to balance generation quality and inference speed in a flexible manner. In this work, we propose Mix-Cache, a training-free caching-based framework for efficient video DiT inference. It first distinguishes the interference and boundary between different caching strategies, and then introduces a context-aware cache triggering strategy to determine when caching should be enabled, along with an adaptive hybrid cache decision strategy for dynamically selecting the optimal caching granularity. Extensive experiments on diverse models demonstrate that, MixCache can significantly accelerate video generation (e.g., 1.94 speedup on Wan 14B, 1.97 speedup on HunyuanVideo) while delivering both superior generation quality and inference efficiency compared to baseline methods. 1 Introduction Diffusion Transformer (DiT) (Peebles and Xie 2023) has revolutionized video generation by integrating the scalability of the Transformer architecture (V aswani et al. 2017) with the power of the diffusion process (Ho, Jain, and Abbeel 2020; Rombach et al. 2022), enabling unprecedented quality in video generation. Cutting-edge video DiT models emerge including SD3.0 (Esser et al. 2024), Sora (Brooks et al. 2024), CogVideoX (Y ang et al. 2024b) and Wan (Wang et al. 2025). These models have facilitated the development of many meaningful applications, such as text-to-video generation (Khachatryan et al. 2023; Wu et al. 2023), video editing (Wang et al. 2023; Jiang et al. 2025) and video continuation (Y ang et al. 2024b).

Test-time Adaptation (TTA) poses a challenge, requiring models to dynamically adapt and perform optimally on shifting target domains. This task is particularly emphasized in real-world driving scenes, where weather domain shifts occur frequently. To address such dynamic changes, our proposed method, TTA-DAME, leverages source domain data augmentation into target domains. Additionally, we introduce a domain discriminator and a specialized domain detector to mitigate drastic domain shifts, especially from daytime to nighttime conditions. To further improve adaptability, we train multiple detectors and consolidate their predictions through Non-Maximum Suppression (NMS). Our empirical validation demonstrates the effectiveness of our method, showing significant performance enhancements on the SHIFT Benchmark.

Masked diffusion language models (MDMs) have recently gained traction as a viable generative framework for natural language. This can be attributed to its scalability and ease of training compared to other diffusion model paradigms for discrete data, establishing itself as the state-of-the-art non-autoregressive generator for discrete data. Diffusion models, in general, have shown excellent ability to improve the generation quality by leveraging inference-time scaling either by increasing the number of denoising steps or by using external verifiers on top of the outputs of each step to guide the generation. In this work, we propose a verifier-based inference-time scaling method that aids in finding a better candidate generation during the denoising process of the MDM. Our experiments demonstrate the application of MDMs for standard text-style transfer tasks and establish MDMs as a better alternative to autoregressive language models. Additionally, we show that a simple soft-value-based verifier setup for MDMs using off-the-shelf pre-trained embedding models leads to significant gains in generation quality even when used on top of typical classifier-free guidance setups in the existing literature.