The Missing Half: Unveiling Training-time Implicit Safety Risks Beyond Deployment

Safety risks of AI models have been widely studied at deployment time, such as jailbreak attacks that elicit harmful outputs. In contrast, safety risks emerging during training remain largely unexplored. Beyond explicit reward hacking that directly manipulates explicit reward functions in reinforcement learning, we study implicit training-time safety risks: harmful behaviors driven by a model's internal incentives and contextual background information. For example, during code-based reinforcement learning, a model may covertly manipulate logged accuracy for self-preservation. We present the first systematic study of this problem, introducing a taxonomy with five risk levels, ten fine-grained risk categories, and three incentive types. Extensive experiments reveal the prevalence and severity of these risks: notably, Llama-3.1-8B-Instruct exhibits risky behaviors in 74.4% of training runs when provided only with background information. We further analyze factors influencing these behaviors and demonstrate that implicit training-time risks also arise in multi-agent training settings. Our results identify an overlooked yet urgent safety challenge in training.

FactNet: A Billion-Scale Knowledge Graph for Multilingual Factual Grounding

While LLMs exhibit remarkable fluency, their utility is often compromised by factual hallucinations and a lack of traceable provenance. Existing resources for grounding mitigate this but typically enforce a dichotomy: they offer either structured knowledge without textual context (e.g., knowledge bases) or grounded text with limited scale and linguistic coverage. To bridge this gap, we introduce FactNet, a massive, open-source resource designed to unify 1.7 billion atomic assertions with 3.01 billion auditable evidence pointers derived exclusively from 316 Wikipedia editions. Unlike recent synthetic approaches, FactNet employs a strictly deterministic construction pipeline, ensuring that every evidence unit is recoverable with byte-level precision. Extensive auditing confirms a high grounding precision of 92.1%, even in long-tail languages. Furthermore, we establish FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking. FactNet provides the community with a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems.

Feasible strategies for conflict resolution within intuitionistic fuzzy preference-based conflict situations

In three-way conflict analysis, preference-based conflict situations characterize agents' attitudes towards issues by formally modeling their preferences over pairs of issues. However, existing preference-based conflict models rely exclusively on three qualitative relations, namely, preference, converse, and indifference, to describe agents' attitudes towards issue pairs, which significantly limits their capacity in capturing the essence of conflict. To overcome this limitation, we introduce the concept of an intuitionistic fuzzy preference-based conflict situation that captures agents' attitudes towards issue pairs with finer granularity than that afforded by classical preference-based models. Afterwards, we develop intuitionistic fuzzy preference-based conflict measures within this framework, and construct three-way conflict analysis models for trisecting the set of agent pairs, the agent set, and the issue set. Additionally, relative loss functions built on the proposed conflict functions are employed to calculate thresholds for three-way conflict analysis. Finally, we present adjustment mechanism-based feasible strategies that simultaneously account for both adjustment magnitudes and conflict degrees, together with an algorithm for constructing such feasible strategies, and provide an illustrative example to demonstrate the validity and effectiveness of the proposed model.

Learning Query-Specific Rubrics from Human Preferences for DeepResearch Report Generation

Nowadays, training and evaluating DeepResearch-generated reports remain challenging due to the lack of verifiable reward signals. Accordingly, rubric-based evaluation has become a common practice. However, existing approaches either rely on coarse, pre-defined rubrics that lack sufficient granularity, or depend on manually constructed query-specific rubrics that are costly and difficult to scale. In this paper, we propose a pipeline to train human-preference-aligned query-specific rubric generators tailored for DeepResearch report generation. We first construct a dataset of DeepResearch-style queries annotated with human preferences over paired reports, and train rubric generators via reinforcement learning with a hybrid reward combining human preference supervision and LLM-based rubric evaluation. To better handle long-horizon reasoning, we further introduce a Multi-agent Markov-state (MaMs) workflow for report generation. We empirically show that our proposed rubric generators deliver more discriminative and better human-aligned supervision than existing rubric design strategies. Moreover, when integrated into the MaMs training framework, DeepResearch systems equipped with our rubric generators consistently outperform all open-source baselines on the DeepResearch Bench and achieve performance comparable to that of leading closed-source models.

No Global Plan in Chain-of-Thought: Uncover the Latent Planning Horizon of LLMs

This work stems from prior complementary observations on the dynamics of Chain-of-Thought (CoT): Large Language Models (LLMs) is shown latent planning of subsequent reasoning prior to CoT emergence, thereby diminishing the significance of explicit CoT; whereas CoT remains critical for tasks requiring multi-step reasoning. To deepen the understanding between LLM's internal states and its verbalized reasoning trajectories, we investigate the latent planning strength of LLMs, through our probing method, Tele-Lens, applying to hidden states across diverse task domains. Our empirical results indicate that LLMs exhibit a myopic horizon, primarily conducting incremental transitions without precise global planning. Leveraging this characteristic, we propose a hypothesis on enhancing uncertainty estimation of CoT, which we validate that a small subset of CoT positions can effectively represent the uncertainty of the entire path. We further underscore the significance of exploiting CoT dynamics, and demonstrate that automatic recognition of CoT bypass can be achieved without performance degradation. Our code, data and models are released at https://github.com/lxucs/tele-lens.

StyleDecoupler: Generalizable Artistic Style Disentanglement

Representing artistic style is challenging due to its deep entanglement with semantic content. We propose StyleDecoupler, an information-theoretic framework that leverages a key insight: multi-modal vision models encode both style and content, while uni-modal models suppress style to focus on content-invariant features. By using uni-modal representations as content-only references, we isolate pure style features from multi-modal embeddings through mutual information minimization. StyleDecoupler operates as a plug-and-play module on frozen Vision-Language Models without fine-tuning. We also introduce WeART, a large-scale benchmark of 280K artworks across 152 styles and 1,556 artists. Experiments show state-of-the-art performance on style retrieval across WeART and WikiART, while enabling applications like style relationship mapping and generative model evaluation. We release our method and dataset at this url.

Moaw: Unleashing Motion Awareness for Video Diffusion Models

Video diffusion models, trained on large-scale datasets, naturally capture correspondences of shared features across frames. Recent works have exploited this property for tasks such as optical flow prediction and tracking in a zero-shot setting. Motivated by these findings, we investigate whether supervised training can more fully harness the tracking capability of video diffusion models. To this end, we propose Moaw, a framework that unleashes motion awareness for video diffusion models and leverages it to facilitate motion transfer. Specifically, we train a diffusion model for motion perception, shifting its modality from image-to-video generation to video-to-dense-tracking. We then construct a motion-labeled dataset to identify features that encode the strongest motion information, and inject them into a structurally identical video generation model. Owing to the homogeneity between the two networks, these features can be naturally adapted in a zero-shot manner, enabling motion transfer without additional adapters. Our work provides a new paradigm for bridging generative modeling and motion understanding, paving the way for more unified and controllable video learning frameworks.

FGTBT: Frequency-Guided Task-Balancing Transformer for Unified Facial Landmark Detection

Recently, deep learning based facial landmark detection (FLD) methods have achieved considerable success. However, in challenging scenarios such as large pose variations, illumination changes, and facial expression variations, they still struggle to accurately capture the geometric structure of the face, resulting in performance degradation. Moreover, the limited size and diversity of existing FLD datasets hinder robust model training, leading to reduced detection accuracy. To address these challenges, we propose a Frequency-Guided Task-Balancing Transformer (FGTBT), which enhances facial structure perception through frequency-domain modeling and multi-dataset unified training. Specifically, we propose a novel Fine-Grained Multi-Task Balancing loss (FMB-loss), which moves beyond coarse task-level balancing by assigning weights to individual landmarks based on their occurrence across datasets. This enables more effective unified training and mitigates the issue of inconsistent gradient magnitudes. Additionally, a Frequency-Guided Structure-Aware (FGSA) model is designed to utilize frequency-guided structure injection and regularization to help learn facial structure constraints. Extensive experimental results on popular benchmark datasets demonstrate that the integration of the proposed FMB-loss and FGSA model into our FGTBT framework achieves performance comparable to state-of-the-art methods. The code is available at https://github.com/Xi0ngxinyu/FGTBT.

Dual-Stream Collaborative Transformer for Image Captioning

Current region feature-based image captioning methods have progressed rapidly and achieved remarkable performance. However, they are still prone to generating irrelevant descriptions due to the lack of contextual information and the over-reliance on generated partial descriptions for predicting the remaining words. In this paper, we propose a Dual-Stream Collaborative Transformer (DSCT) to address this issue by introducing the segmentation feature. The proposed DSCT consolidates and then fuses the region and segmentation features to guide the generation of caption sentences. It contains multiple Pattern-Specific Mutual Attention Encoders (PSMAEs) and Dynamic Nomination Decoders (DNDs). The PSMAE effectively highlights and consolidates the private information of two representations by querying each other. The DND dynamically searches for the most relevant learning blocks to the input textual representations and exploits the homogeneous features between the consolidated region and segmentation features to generate more accurate and descriptive caption sentences. To the best of our knowledge, this is the first study to explore how to fuse different pattern-specific features in a dynamic way to bypass their semantic inconsistencies and spatial misalignment issues for image captioning. The experimental results from popular benchmark datasets demonstrate that our DSCT outperforms the state-of-the-art image captioning models in the literature.

Supervision-by-Hallucination-and-Transfer: A Weakly-Supervised Approach for Robust and Precise Facial Landmark Detection

High-precision facial landmark detection (FLD) relies on high-resolution deep feature representations. However, low-resolution face images or the compression (via pooling or strided convolution) of originally high-resolution images hinder the learning of such features, thereby reducing FLD accuracy. Moreover, insufficient training data and imprecise annotations further degrade performance. To address these challenges, we propose a weakly-supervised framework called Supervision-by-Hallucination-and-Transfer (SHT) for more robust and precise FLD. SHT contains two novel mutually enhanced modules: Dual Hallucination Learning Network (DHLN) and Facial Pose Transfer Network (FPTN). By incorporating FLD and face hallucination tasks, DHLN is able to learn high-resolution representations with low-resolution inputs for recovering both facial structures and local details and generating more effective landmark heatmaps. Then, by transforming faces from one pose to another, FPTN can further improve landmark heatmaps and faces hallucinated by DHLN for detecting more accurate landmarks. To the best of our knowledge, this is the first study to explore weakly-supervised FLD by integrating face hallucination and facial pose transfer tasks. Experimental results of both face hallucination and FLD demonstrate that our method surpasses state-of-the-art techniques.