The rapid growth of AI-generated imagery has blurred the boundary between real and synthetic content, raising critical concerns for digital integrity. Vision-language models (VLMs) offer interpretability through explanations but often fail to detect subtle artifacts in high-quality synthetic images. We propose ZoomIn, a two-stage forensic framework that improves both accuracy and interpretability. Mimicking human visual inspection, ZoomIn first scans an image to locate suspicious regions and then performs a focused analysis on these zoomed-in areas to deliver a grounded verdict. To support training, we introduce MagniFake, a dataset of 20,000 real and high-quality synthetic images annotated with bounding boxes and forensic explanations, generated through an automated VLM-based pipeline. Our approach achieves 96.39% accuracy with strong generalization across external datasets, and providing human-understandable explanations grounded in visual evidence. The rapid advancement of image generation models (Wang et al., 2025b; Li et al., 2025; Chadebec et al., 2025) has enabled the creation of AI-generated images with unprecedented photorealism, increasingly blurring the boundary between authentic and synthetic content. There is a critical need for platforms to deploy accurate and effective detection methods. However, the current landscape of detection methods is dominated by classification-based approaches. While often effective on specific datasets, these methods typically operate as "black-or gray-boxes", offering little insight into their decision-making process. This lack of explainability is coupled with poor generalizability, as models trained to detect artifacts from one generative architecture often fail when confronted with novel, unseen ones.

Recent advances in large language models (LLMs) have sparked growing interest in building generalist agents that can learn through online interactions. However, applying reinforcement learning (RL) to train LLM agents in multi-turn, multi-task settings remains challenging due to lack of scalable infrastructure and stable training algorithms. In this work, we present the AgentRL framework for scalable multi-turn, multi-task agentic RL training. On the infrastructure side, AgentRL features a fully-asynchronous generation-training pipeline for efficient multi-turn RL. To support heterogeneous environment development in multi-task RL, we design a unified function-call based API interface, containerized environment development, and a centralized controller. On the algorithm side, we propose cross-policy sampling to encourage model exploration in multi-turn settings and task advantage normalization to stabilize multi-task training. Experiments show that AgentRL, trained on open LLMs across five agentic tasks, significantly outperforms GPT-5, Clause-Sonnet-4, DeepSeek-R1, and other open-source LLM agents. Multi-task training with AgentRL matches the best results among all task-specific models. AgentRL is open-sourced at https://github.com/THUDM/AgentRL. The algorithm and framework are adopted in building \textsc{\href{https://autoglm.zhipuai.cn}{AutoGLM}}.

Large Reasoning Models (LRMs) have demonstrated strong capabilities in complex multi-step reasoning, opening new opportunities for automating optimization modeling. However, existing domain adaptation methods, originally designed for earlier instruction-tuned models, often fail to exploit the advanced reasoning patterns of modern LRMs -- In particular, we show that direct fine-tuning on traditional \textit{non-reflective} datasets leads to limited gains. To fully leverage LRMs' inherent reasoning abilities, we propose \textbf{CALM} (\textit{Corrective Adaptation with Lightweight Modification}), a framework that progressively refines LRMs within their native reasoning modes for optimization modeling tasks. In CALM, an expert intervener identifies reasoning flaws and provides concise corrective hints, which the LRM incorporates to produce improved reasoning trajectories. These interventions modify fewer than 2.6\% of generated tokens, but generate high-quality data for soft adaptation through supervised fine-tuning. The adapted model is then further improved through reinforcement learning. Building on CALM, we develop \textbf{STORM} (\textit{Smart Thinking Optimization Reasoning Model}), a 4B-parameter LRM that achieves a new state-of-the-art average accuracy of 68.9\% across five popular optimization modeling benchmarks, matching the performance of a 671B LRM. These results demonstrate that dynamic, hint-based data synthesis both preserves and amplifies the native reasoning patterns of modern LRMs, offering a more effective and scalable path towards expert-level performance on challenging optimization modeling tasks.

Loss explosions in training deep neural networks can nullify multi-million dollar training runs. Conventional monitoring metrics like weight and gradient norms are often lagging and ambiguous predictors, as their values vary dramatically across different models and even between layers of the same model, making it difficult to establish a unified standard for detecting impending failure. We introduce Spectral Alignment (SA), a novel, theoretically-grounded metric that monitors the distributional alignment between layer inputs and the principal singular vectors of weight matrices. We show that a collapse in the sign diversity of this alignment is a powerful early predictor of representational collapse and training divergence. Empirical results on language models demonstrate that monitoring the SA distribution provides a significantly earlier and clearer warning of loss explosions than traditional scalar metrics. SA's low computational overhead makes it a practical tool for safeguarding model training.

Large Multimodal Reasoning Models (LMRMs) are moving into real applications, where they must be both useful and safe. Safety is especially challenging in multimodal settings: images and text can be combined to bypass guardrails, and single objective training can cause policy drift that yields over-refusal on benign inputs or unsafe compliance on risky ones. We present COSMO-RL, a mixed reinforcement learning framework that trains reasoning oriented LMRMs under multimodal, multitask, and multiobjective signals, and we release the resulting model, COSMO-R1. Our approach aims to let safety and capability grow together in one stable pipeline rather than competing during alignment. In experiments, COSMO-R1 improves safety while maintaining-and often improving multimodal reasoning and instruction following, shows stronger robustness to multimodal jailbreaks, and reduces unnecessary refusals. The framework also transfers across backbones with consistent gains. Ablations support the design choices, indicating a simple path to advancing safety and general capability together in LMRMs.

Given a map description through global traversal navigation instructions (e.g., visiting each room sequentially with action signals such as north, west, etc.), an LLM can often infer the implicit spatial layout of the environment and answer user queries by providing a shortest path from a start to a destination (for instance, navigating from the lobby to a meeting room via the hall and elevator). However, such context-dependent querying becomes incapable as the environment grows much longer, motivating the need for incremental map construction that builds a complete topological graph from stepwise observations. We propose a framework for LLM-driven construction and map repair, designed to detect, localize, and correct structural inconsistencies in incrementally constructed navigation graphs. Central to our method is the Version Control, which records the full history of graph edits and their source observations, enabling fine-grained rollback, conflict tracing, and repair evaluation. We further introduce an Edge Impact Score to prioritize minimal-cost repairs based on structural reachability, path usage, and conflict propagation. To properly evaluate our approach, we create a refined version of the MANGO benchmark dataset by systematically removing non-topological actions and inherent structural conflicts, providing a cleaner testbed for LLM-driven construction and map repair. Our approach significantly improves map correctness and robustness, especially in scenarios with entangled or chained inconsistencies. Our results highlight the importance of introspective, history-aware repair mechanisms for maintaining coherent spatial memory in LLM agents.

Deep neural networks have achieved remarkable success in computer vision; however, their black-box nature in decision-making limits interpretability and trust, particularly in safety-critical applications. Interpretability is crucial in domains where errors have severe consequences. Existing models not only lack transparency but also risk exploiting unreliable or misleading features, which undermines both robustness and the validity of their explanations. Concept Bottleneck Models (CBMs) aim to improve transparency by reasoning through human-interpretable concepts. Still, they require costly concept annotations and lack spatial grounding, often failing to identify which regions support each concept. We propose SEG-MIL-CBM, a novel framework that integrates concept-guided image segmentation into an attention-based multiple instance learning (MIL) framework, where each segmented region is treated as an instance and the model learns to aggregate evidence across them. By reasoning over semantically meaningful regions aligned with high-level concepts, our model highlights task-relevant evidence, down-weights irrelevant cues, and produces spatially grounded, concept-level explanations without requiring annotations of concepts or groups. SEG-MIL-CBM achieves robust performance across settings involving spurious correlations (unintended dependencies between background and label), input corruptions (perturbations that degrade visual quality), and large-scale benchmarks, while providing transparent, concept-level explanations.

Syntax highlighting is a critical feature in modern software development environments, enhancing code readability and developer productivity. However, delivering accurate highlighting in real time remains challenging for online and web-based development tools due to strict time and memory constraints on backend services. These systems must serve highlights rapidly and frequently, even when code is partially valid or invalid. This has led to on-the-fly syntax highlighting, where visual annotations are generated just before content is served, often at high request rates and under incomplete input conditions. To meet these demands efficiently, state-of-the-art models use deep learning to learn the behavior of brute-force syntax highlighting resolvers, tools that are easy to implement but too slow for production. Through the Deep Abstraction process, brute-force strategies are encoded into fast statistical models that achieve both high accuracy and low-latency inference. Despite their success, such models face key challenges: they support only one programming language per model, require large datasets from slow brute-force generators, and involve resource-intensive training. In multi-language environments, this means maintaining multiple independent models, increasing system complexity and operational cost. This work addresses these issues by introducing a unified model capable of highlighting up to six mainstream programming languages, reducing deployment complexity by a factor of six and improving performance on unseen languages. A novel normalization technique significantly enhances model generalization, while few-shot learning experiments show that a small number of oracle samples can replace large datasets, minimizing dependence on brute-force generators. Combined, these innovations enable efficient, scalable, and cost-effective syntax highlighting across diverse programming languages.

Diffusion and flow-based non-autoregressive (NAR) models have shown strong promise in large language modeling, however, their potential for automatic speech recognition (ASR) remains largely unexplored. We propose Drax, a discrete flow matching framework for ASR that enables efficient parallel decoding. To better align training with inference, we construct an audio-conditioned probability path that guides the model through trajectories resembling likely intermediate inference errors, rather than direct random noise to target transitions. Our theoretical analysis links the generalization gap to divergences between training and inference occupancies, controlled by cumulative velocity errors, thereby motivating our design choice. Empirical evaluation demonstrates that our approach attains recognition accuracy on par with state-of-the-art speech models while offering improved accuracy-efficiency trade-offs, highlighting discrete flow matching as a promising direction for advancing NAR ASR.

Diffusion-based Large Language Models (dLLMs) have emerged as a competitive alternative to autoregressive models, offering unique advantages through bidirectional attention and parallel generation paradigms. However, the generation results of current parallel decoding methods deviate from stepwise decoding, introducing potential performance degradation, which limits their practical deployment. To address this problem, we propose \textbf{S}elf \textbf{S}peculative \textbf{D}ecoding (SSD), a lossless inference acceleration method that leverages the dLLM itself as both speculative decoding drafter and verifier without auxiliary modules. SSD introduces a self-drafting mechanism where the model generates predictions for multiple positions, then verifies them through hierarchical verification trees in a single forward pass. Unlike traditional speculative decoding that requires separate draft models, SSD eliminates model redundancy and memory overhead by exploiting the dLLM's inherent parallel prediction capability for multiple positions. This self-speculative approach allows the model to progressively verify and accept multiple tokens in a single forward pass. Our experiments demonstrate that SSD achieves up to 3.46$\times$ speedup while keeping the output identical to stepwise decoding on open source models such as LLaDA and Dream. Code will be made publicly available on GitHub.