Zhou Zhao

Visual generation is becoming a general component of AI infrastructure, supporting content creation across design, entertainment, education, science, and industry. Generative and multimodal models have substantially improved single-pass synthesis. At the same time, multi-step visual creation requires spatial precision, identity persistence, temporal coherence, editable structure, factual fidelity, and physical plausibility throughout the creation process. Errors in intermediate artifacts can propagate across subsequent steps, making creation control a central requirement. Model scaling, prompt refinement, additional conditioning, and repeated sampling improve individual outputs; coordinating these capabilities across a task requires runtime evidence to influence subsequent creation decisions. Within AI-generated content (AIGC), we use Agentic AIGC to describe goal-driven content creation organized through a multi-step control loop. This survey focuses on visual generation agents: systems that use runtime evidence from evolving visual artifacts, task environments, and interaction histories to change later creation-relevant action selection or parameterization, including generation, editing, verification, recovery, replanning, and stopping. The defining behavioral property is therefore an observation--action dependency: runtime observations change later action selection. Whether those decisions improve task or constraint outcomes is evaluated separately with matched tools and resource budgets. We organize this emerging field around the control mechanisms and behavioral evidence that connect generative capabilities to autonomous visual creation. Specifically, we formalize this boundary and characterize autonomy through five cumulative levels (L1--L5). A six-component analytical framework (goal understanding, specification, and planning; memory; tool; perception; action; and cross-task self-improvement) structures a cross-domain comparison spanning image, video, 3D/CAD, scientific visualization, structured documents, UI/Web, and related visual-creation domains. Building on this comparison, we analyze training from trajectory supervision and policy initialization through reinforcement learning from multimodal and executable feedback to experience reuse with skill abstraction and transfer. We organize evaluation into four complementary levels: artifact quality, goal and constraint satisfaction, trajectory and decision quality, and system and human-centered evaluation. This synthesis distinguishes behavioral evidence from evidence that feedback improves a creation outcome. Finally, we identify research directions in the scope of visual agency, reliable long-horizon control, cross-task improvement, and evaluation.

Code summarization plays a vital role in program comprehension and software maintenance by generating natural language descriptions to summarize the semantics of code. While Large Language Models (LLMs) have shown remarkable performance in this area, recent empirical studies reveal a critical limitation: LLMs are prone to hallucinations, producing summaries that are factually inaccurate or unfaithful to the source code, potentially misleading developers. In this paper, we propose to unveil, detect, and mitigate hallucinations in LLM-based code summarization. First, we construct Hallu-Eval, a novel dataset for unveiling hallucination phenomena and rigorously evaluating the effectiveness of hallucination detection and mitigation in LLM-based code summarization. It comprises both original code snippets to capture naturally occurring hallucinations and their semantically perturbed counterparts, which are designed to systematically induce challenging logical hallucinations, all complemented with manual hallucination annotations on a curated testbed of 800 code-summary pairs. Next, we propose Hallu-Det, a synergistic approach that combines direct entity-level detection to identify explicit hallucinations with a synonymous mutation-based refinement to reliably confirm or refute more ambiguous cases. Finally, we introduce Hallu-Shield, an inference-time mitigation approach that leverages an external value model to guide LLMs toward producing more faithful summaries without costly retraining of the LLM itself. Extensive experiments show that Hallu-Eval effectively triggers hallucinations, increasing the hallucination rate of models such as Qwen2.5-Coder-7B from 17% to 97% on perturbed code. Our detection approach, Hallu-Det, achieves the best performance among baselines, reaching an F1-score of 0.95 for summaries generated by Qwen2.5-Coder-7B. Moreover, our mitigation method, Hallu-Shield, reduces hallucination rates. For example, it lowers the rate from 66% to 59%, a 10.6% relative reduction, on DeepSeek-Coder-6.7B, while simultaneously improving summary quality, achieving a 74.0% win rate evaluated by an LLM-as-a-judge majority vote ensemble.

Recent advancements in multi-modal large language models (MLLMs) have shown strong potential for 3D scene understanding. However, existing methods struggle with fine-grained object grounding and contextual reasoning, limiting their ability to interpret and interact with complex 3D environments. In this paper, we present Chat-Scene++, an MLLM framework that represents 3D scenes as context-rich object sequences. By structuring scenes as sequences of objects with contextual semantics, Chat-Scene++ enables object-centric representation and interaction. It decomposes a 3D scene into object representations paired with identifier tokens, allowing LLMs to follow instructions across diverse 3D vision-language tasks. To capture inter-object relationships and global semantics, Chat-Scene++ extracts context-rich object features using large-scale pre-trained 3D scene-level and 2D image-level encoders, unlike the isolated per-object features in Chat-Scene. Its flexible object-centric design also supports grounded chain-of-thought (G-CoT) reasoning, enabling the model to distinguish objects at both category and spatial levels during multi-step inference. Without the need for additional task-specific heads or fine-tuning, Chat-Scene++ achieves state-of-the-art performance on five major 3D vision-language benchmarks: ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D. These results highlight its effectiveness in scene comprehension, object grounding, and spatial reasoning. Additionally, without reconstructing 3D worlds through computationally expensive processes, we demonstrate its applicability to real-world scenarios using only 2D inputs. Code will be made available at https://github.com/ZzZZCHS/Chat-Scene https://github.com/ZzZZCHS/Chat-Scene.

This work presents WorldCompass, a novel Reinforcement Learning (RL) post-training framework for the long-horizon, interactive video-based world models, enabling them to explore the world more accurately and consistently based on interaction signals. To effectively "steer" the world model's exploration, we introduce three core innovations tailored to the autoregressive video generation paradigm: 1) Clip-level rollout Strategy: We generate and evaluate multiple samples at a single target clip, which significantly boosts rollout efficiency and provides fine-grained reward signals. 2) Complementary Reward Functions: We design reward functions for both interaction-following accuracy and visual quality, which provide direct supervision and effectively suppress reward-hacking behaviors. 3) Efficient RL Algorithm: We employ the negative-aware fine-tuning strategy coupled with various efficiency optimizations to efficiently and effectively enhance model capacity. Evaluations on the SoTA open-source world model, WorldPlay, demonstrate that WorldCompass significantly improves interaction accuracy and visual fidelity across various scenarios.