Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of any single memory component is hard to isolate. We introduce and instrument an alternative bounded contract: every decision is made from a fresh user message assembled by typed retrieval, with no raw cross-decision transcript appended. The prompt thus stays bounded across runs of any length, and any single layer can be ablated in isolation. We instantiate the contract in Slay the Spire 2, a closed-rule stochastic deck-building game whose runs require hundreds of tactical and strategic decisions. A public online benchmark of frontier LLMs on the same game reports zero wins at the lowest difficulty across five configurations, and the developer-reported human win rate at the same difficulty is 16

Designing a board game demands both thinking as a designer and experiencing as a player, while iterating through repeated prototyping and playtesting cycles, making it a cognitively intensive creative task well suited for human-AI collaboration. However, current systems lack end-to-end support to guide designers through the complete workflow from vague early ideation to iterative rulebook revision and audience testing. To this end, we present AutoBG, a board game design assistant built around critic-driven iterative refinement, comprising four specialized modules: BG-Ideator guides designers via multi-turn dialogue to produce structured design drafts; BG-Realizer generates complete rulebooks from drafts and revises them in a closed loop with BG-Critic, which diagnoses design flaws and gates each revision so that only verified improvements are accepted; and BG-Persona simulates individualized feedback from 150 real player profiles. Together, these modules enable designers to go from an initial idea to a polished, audience-tested rulebook within a single integrated workflow. The system is built on 2.2K structured rulebooks and 180K quality-filtered real player reviews, with task-specific training data derived for each module. Experiments on 207 held-out games show that AutoBG substantially outperforms state-of-the-art baselines (e.g., GPT-5.4), generating rulebooks that approach the quality of published games. Furthermore, a user study with 30 participants across diverse experience levels confirms that AutoBG effectively reduces blank-page anxiety, surfaces hidden design flaws, and provides highly rated, practical assistance throughout the creative process.

Game worlds have traditionally been built through labor-intensive production pipelines, making them costly to develop, difficult to customization, and expensive to modify after deployment. Recent advances in video world models offer a fundamentally different paradigm. Rather than explicitly authoring every component of a virtual environment, these models autoregressively synthesize future observations conditioned on the current world state and user interactions, enabling playable worlds to be generated online. Trained on both gameplay recordings and real-world videos, they can capture diverse visual appearances and physical dynamics, opening new opportunities for interactive applications beyond gaming, including embodied intelligence. In this paper, we present AlayaWorld, a full-stack open-source framework for building interactive generative worlds. AlayaWorld enables open-ended real-time interaction, allowing users to freely navigate and perform diverse actions such as combat, spell casting, and monster summoning. The framework unifies the complete development-from data preparation model architecture, model training, inference acceleration, and deployment-within a modular and extensible architecture. Alongside the framework, we release reproducible pipelines, reference implementations, evaluation tools, and comprehensive documentation, establishing a practical foundation for future research and real-time applications of generative world models.

Dynamical systems theory and reinforcement learning view world evolution as latent-state dynamics driven by actions, with visual observations providing partial information about the state. Recent video world models attempt to learn this action-conditioned dynamics from data. However, existing datasets rarely match the requirement: they typically lack diverse and semantically meaningful action spaces, and actions are directly tied to visual observations rather than mediated by underlying states. As a result, actions are often entangled with pixel-level changes, making it difficult for models to learn structured world dynamics and maintain consistent evolution over long horizons. In this paper, we propose WildWorld, a large-scale action-conditioned world modeling dataset with explicit state annotations, automatically collected from a photorealistic AAA action role-playing game (Monster Hunter: Wilds). WildWorld contains over 108 million frames and features more than 450 actions, including movement, attacks, and skill casting, together with synchronized per-frame annotations of character skeletons, world states, camera poses, and depth maps. We further derive WildBench to evaluate models through Action Following and State Alignment. Extensive experiments reveal persistent challenges in modeling semantically rich actions and maintaining long-horizon state consistency, highlighting the need for state-aware video generation. The project page is https://shandaai.github.io/wildworld-project/.

Recent advances in AI-assisted programming have empowered agents to execute complex workflows via command-line interfaces, however, existing benchmarks are limited by short task horizons, data contamination from GitHub scraping, and a lack of fine-grained evaluation metrics, fail to rigorously evaluate the long-horizon planning and execution capabilities essential for realistic software engineering. To address these gaps, we introduce LongCLI-Bench, a comprehensive benchmark designed to evaluate agentic capabilities across long-horizon, realistic tasks. We curated 20 high-quality, long-horizon tasks from over 1,000 computer science assignments and real-world workflows, covering four engineering categories: from scratch, feature addition, bug fixing, and refactoring. We propose a dual-set testing protocol for LongCLI-Bench, which measures requirement fulfillment (fail-to-pass) and regression avoidance (pass-to-pass), and incorporates step-level scoring to pinpoint execution failures. Extensive experiments reveal that even state-of-the-art agents achieve pass rates below 20

The field of 4D world modeling—aiming to jointly capture spatial geometry and temporal dynamics—has witnessed remarkable progress in recent years, driven by advances in large-scale generative models and multimodal learning. However, the development of truly general 4D world models remains fundamentally constrained by the availability of high-quality data. Existing datasets and benchmarks often lack the dynamic complexity, multi-domain diversity, and spatial-temporal annotations required to support key tasks such as 4D geometric reconstruction, future prediction, and camera-controlled video generation. To address this gap, we introduce OmniWorld, a large-scale, multi-domain, multi-modal dataset specifically designed for 4D world modeling. OmniWorld consists of a newly collected OmniWorld-Game dataset and several curated public datasets spanning diverse domains. Compared with existing synthetic datasets, OmniWorld-Game provides richer modality coverage, larger scale, and more realistic dynamic interactions. Based on this dataset, we establish a challenging benchmark that exposes the limitations of current state-of-the-art (SOTA) approaches in modeling complex 4D environments. Moreover, fine-tuning existing SOTA methods on OmniWorld leads to significant performance gains across 4D reconstruction and video generation tasks, strongly validating OmniWorld as a powerful resource for training and evaluation. We envision OmniWorld as a catalyst for accelerating the development of general-purpose 4D world models, ultimately advancing machines’ holistic understanding of the physical world.

Current AI-driven game development has made substantial progress in asset generation, gameplay design, and web-based game coding, yet project-level code engineering on professional game engines remains largely unexplored due to the absence of large-scale datasets and deterministic evaluation methods. We present JamSet and JamBench, the first project-level game code framework dataset and benchmark built on a professional game engine. Our key insight is that Game Jam competitions, community events where developers build complete games under tight time constraints, yield thousands of open-source projects suitable for this purpose. Building on the Godot engine's text-based format and headless execution mode, we design a deterministic verification pipeline from file integrity to runtime behavior collection, distilling 8,133 verified projects from over 240,000 repositories. Of these, 300 manually verified projects form JamBench; the rest constitute JamSet. JamBench defines theme-driven generation and code completion tasks, evaluated through a pipeline combining compilation pass rates, Structural Completeness Score (SCS), and Behavioral Alignment Score (BAS). Evaluation of 9 frontier models reveals a capability cliff as project scale increases, with runtime pass rates dropping from 80.4

Reinforcement learning for agentic multimodal models often suffers from interaction collapse, where models learn to reduce tool usage and multi-turn reasoning, limiting the benefits of agentic behavior. We introduce PyVision-RL, a reinforcement learning framework for open-weight multimodal models that stabilizes training and sustains interaction. Our approach combines an oversampling–filtering–ranking rollout strategy with an accumulative tool reward to prevent collapse and encourage multi-turn tool use. Using a unified training pipeline, we develop PyVision-Image and PyVision-Video for image and video understanding. For video reasoning, PyVision-Video employs on-demand context construction, selectively sampling task-relevant frames during reasoning to significantly reduce visual token usage. Experiments show strong performance and improved efficiency, demonstrating that sustained interaction and on-demand visual processing are critical for scalable multimodal agents.

When should an intelligent assistant speak up without being asked? Continuous egocentric video offers rich, evolving context that enables a new form of assistance: one that is proactive rather than merely reactive. Yet existing approaches either wait passively for user queries or treat every detected event as requiring a response, without considering the user's history, current activity, or whether assistance would actually be welcome. We reframe proactive assistance as a context-dependent decision problem: the agent must not only perceive what is happening, but reason over accumulated temporal context to determine when and whether to intervene. To this end, we present Vinci2, a proactive egocentric assistance system that advances the on-device assistant Vinci from reactive response toward proactivity. On the evaluation side, we present EgoServe, the first large-scale benchmark for proactive assistance in continuous egocentric video. EgoServe comprises over 3,000 service instances organized along 4 temporal memory horizons, ranging from immediate safety alerts to long-term habit coaching, across 10 service categories. On the modeling side, we propose EgoMemo, a training-free, memory-augmented agent that maintains three complementary memory representations: multi-scale temporal summaries, a semantic knowledge graph, and visual embedding archives. At each timestep, EgoMemo performs retrieval-augmented reasoning to determine whether assistance is warranted and, if so, produces contextually grounded responses. Experiments demonstrate that EgoMemo establishes strong baselines on EgoServe while remaining competitive on existing egocentric benchmarks. Our benchmark and code are publicly available at \href{https://sitonggong.github.io/EgoServe-page/}{Vinci2}.

Unsupervised vision-language pre-training (UVLP) is a promising way to alleviate the reliance on large-scale paired image-text data, which is expensive and time-consuming to obtain. Existing UVLP approaches are mainly generation-based or retrieval-based: the former produces well-aligned but overly simplistic pairs, while the latter provides richer samples but suffers from weak alignment. To tackle these problems, we propose a method to Retrieve and then Generate image-text pairs from Semantic Concepts (RGSC). Specifically, we extract semantic concepts from both images and texts, which enables well alignment in retrieval while preserving diverse information for generation. Then, we propose SeCoCap, which reconstructs complete captions from semantic concepts retrieved by images, enabling unsupervised vision-language alignment without paired data. Extensive experiments demonstrate that RGSC achieves a new state-of-the-art on four downstream tasks, including VQA, NLVR2, visual entailment, and image-text retrieval.

Composed Image Retrieval (CIR) retrieves target images based on the composition of a reference image and a textual modification. Existing CIR approaches commonly assume that training triplets are perfectly aligned, yet large-scale web-sourced data inevitably contain noise. Such noise may originate from the reference image, the modification text, or the target image. However, current approaches lack mechanisms to distinguish and handle these different types of noise. To address this problem, we propose a novel noisy correspondence decomposition framework that categorizes noisy triplets according to different noise types. Specifically, it identifies text noise by calculating the visual alignment between reference and target images, and detects reference-image noise by comparing fused features against the reference and target representations. Building on this framework, we introduce a multi-path composed image retrieval (MPCIR) method that routes triplets along different learning paths based on their noise type, enabling the model to extract reliable supervision from partially misaligned data. Finally, to bridge the gap between clean benchmarks and real-world data, we introduce a reconstruction framework that transforms existing annotations into realistic datasets. This yields Real-CIRR and Real-FashionIQ, two new datasets that reflect naturally occurring noisy correspondences in real-world web search scenarios. Extensive experiments on four CIR datasets show that MPCIR delivers strong robustness under both synthetic and real-world noise.

Multimodal large language models (MLLMs), which integrate language and visual cues for problem-solving, are crucial for advancing artificial general intelligence (AGI). However, current benchmarks for measuring the intelligence of MLLMs suffer from limited scale, narrow coverage, and unstructured knowledge, offering only static and undifferentiated evaluations. To bridge this gap, we introduce MDK12-Bench, a large-scale multidisciplinary benchmark built from real-world K–12 exams spanning six disciplines with 141K instances and 6,225 knowledge points organized in a six-layer taxonomy. Covering five question formats with difficulty and year annotations, it enables comprehensive evaluation to capture the extent to which MLLMs perform over four dimensions: 1) difficulty levels, 2) temporal (cross-year) shifts, 3) contextual shifts, and 4) knowledge-driven reasoning. We propose a novel dynamic evaluation framework that introduces unfamiliar visual, textual, and question form shifts to challenge model generalization while improving benchmark objectivity and longevity by mitigating data contamination. We further evaluate knowledge-point reference-augmented generation (KP-RAG) to examine the role of knowledge in reasoning. Key findings reveal limitations in current MLLMs in multiple aspects and provide guidance for enhancing model reasoning, robustness, and AI-assisted education.

Streaming autoregressive diffusion makes minute-scale video synthesis practical, but its bounded context and fixed denoising schedule allocate resources uniformly across a highly non-stationary sequence. A rolling key-value cache forgets distant visual evidence even when that evidence remains important, while every generated chunk receives the same number of denoising passes irrespective of its actual difficulty. We introduce Surprise Forcing, a training-free framework that treats both limitations as online resource-allocation problems. A Surprise-Gated Memory Bank summarizes evicted frames with value-token descriptors, evaluates them using complementary global-deviation and nearest-neighbor novelty signals, and regulates admission through a feedback-controlled budget in normalized score space. Priority-based replacement and relevance-aware routing then keep the external memory compact and useful. In parallel, Surprise-Aware Denoising estimates chunk difficulty from the maximum adjacent-frame cosine distance after the first denoising pass and uses a local percentile scheduler to skip intermediate steps for comparatively easy chunks. Experiments on VBench, VBench-Long, and VBench-2.0 show that the proposed allocation strategy improves long-horizon consistency and visual quality while retaining real-time streaming throughput.

Current semantic segmentation models are very data-hungry and require massive costly pixel-wise human annotations. Generative data augmentation, which scales the train set using generative models, provides a potential remedy. In this paper, we propose MatchMask, a novel mask-centric generative data augmentation approach tailored for label-scarce semantic segmentation. By leveraging a limited set of labeled semantic masks, MatchMask generates diverse, realistic, and well-aligned image-mask pairs, thereby enhancing the performance of semantic segmentation models. Specifically, to adapt existing text-to-image models for semantic image synthesis in the few-shot setting, we first propose a Gradient Probe Method to investigate the role of each layer in the diffusion model. On this basis, a lightweight LoRA-style adapter is designed for critical layers to enable efficient adaptation, coupled with a Layer-adaptive Cross-attention Fusion mechanism. Meanwhile, we present a robust relative filtering principle to suppress incorrectly synthesized regions. Moreover, the proposed approach is extended to MatchMask++ in the semi-supervised setting to take advantage of additional unlabeled data. Experimental results on PASCAL VOC, COCO and ADE20K demonstrate that MatchMask remarkably enhances the performance of segmentation models, surpassing prior data augmentation techniques in various benchmarks, e.g., 67.5%->74.3% mIoU on PASCAL VOC. Our code will be made publicly available.

Existing mainstream video customization methods focus on generating identity-consistent videos based on given reference images and textual prompts. Benefiting from the rapid advancement of joint audio-video generation, this paper proposes a more compelling new task: sync audio-video customization, which aims to synchronously customize both video identity and audio timbre. Specifically, given a reference image I^r and a reference audio A^r, this novel task requires generating videos that maintain the identity of the reference image while imitating the timbre of the reference audio, with spoken content freely specifiable through user-provided textual prompts. To this end, we propose OmniCustom, a powerful DiT-based audio-video customization framework that can synthesize a video following reference image identity, audio timbre, and text prompts all at once in a zero-shot manner. Our framework is built on three key contributions. First, identity and audio timbre control are achieved through separate reference identity and audio LoRA modules that operate through self-attention layers within the base audio-video generation model. Second, we introduce a contrastive learning objective alongside the standard flow matching objective. It uses predicted flows conditioned on reference inputs as positive examples and those without reference conditions as negative examples, thereby enhancing the model ability to preserve identity and timbre. Third, we train OmniCustom on our constructed large-scale, high-quality audio-visual human dataset. Extensive experiments demonstrate that OmniCustom outperforms existing methods in generating audio-video content with consistent identity and timbre fidelity.

Interactive game world models typically autoregress visual observations directly in pixel or latent space, forcing structured properties such as pose, geometry, and occlusion to be implicitly maintained by the same generative sequence. Over long horizons, errors in these latent world properties accumulate, making consistency and controllability fragile. We explicitly model the evolving world state, delegate exact geometric computation to a fixed, zero-parameter renderer, and leave the neural model to synthesize appearance. We instantiate this idea as Marionette, a world model for interactive games with articulated characters. First, a two-stage autoregressive dynamics model predicts an explicit and interpretable 276-dimensional 3D world state comprising multi-entity articulated skeletons, metric root trajectories, and rotations. Second, a zero-parameter graphics bridge converts the predicted state into pose-control videos, computing world-space geometry and occlusion in closed form. Third, a control-conditioned video-diffusion observation model synthesizes photorealistic RGB observations from the resulting structured controls. Our experiments establish two properties of Marionette. First, the predicted world state is directly controllable. Forcing a mismatched action stream changes root-aligned joint error by 31

Multimodal Large Language Models (MLLMs) are increasingly deployed in human-facing roles where personality perception is critical, yet existing benchmarks evaluate this capability solely on numerical Big Five score prediction, leaving open whether models truly perceive personality through behavioral understanding or merely prejudge through superficial pattern matching. We address this gap with three contributions. (i) A new task: we formalize Grounded Personality Reasoning (GPR), which requires MLLMs to anchor each Big Five rating in observable evidence through a chain of rating, reasoning, and grounding. (ii) A new dataset: we release MM-OCEAN (1,104 videos, 5,320 MCQs), produced by a multi-agent pipeline with human verification, with timestamped behavioral observations, evidence-grounded trait analyses, and seven categories of cue-grounding MCQs. (iii) Benchmark and analysis: we design a three-tier evaluation (rating, reasoning, grounding) plus four sample-level failure-mode metrics: Prejudice Rate (PR), Confabulation Rate (CR), Integration-failure Rate (IR), and Holistic-grounding Rate (HR), and benchmark 27 MLLMs (13 closed, 14 open). The analysis uncovers a striking Prejudice Gap: across the field, 51

Multimodal large language models (MLLMs) have achieved remarkable success in multimodal applications. Their reliance on large-scale, internet-sourced data raises significant concerns regarding security and privacy. These concerns are especially critical when MLLM-based systems are deployed in agentic edge networks, where limited computational resources make retraining impractical for addressing privacy and security issues. Machine Unlearning (MU) has emerged as a critical technique for dealing with these issues. It enables the selective removal of targeted information from pre-trained models without costly retraining. However, the evaluation of MU for MLLMs in agentic edge networks remains inadequate. Existing benchmarks focus narrowly on entities and overlook unlearning of broader visual concepts and their semantic couplings. To bridge this gap, we introduce IEBench. This benchmark enables thorough assessment of MU for MLLM-based systems in agentic edge networks. IEBench features a fictitious dataset of identity entities and corresponding event scenes. It supports the evaluation of unlearning across these distinct yet entangled concepts. We use IEBench to evaluate eight MU methods and reveal their individual strengths and weaknesses. Our findings show that unlearning one concept can unintentionally degrade performance on related concepts within the same image. We term this challenge cross concept interference. Furthermore, we demonstrate the difficulty of unlearning identity and event concepts simultaneously. We also explore potential methods to mitigate these conflicting objectives.

Scientific illustrations are essential tools for communicating research findings, especially in natural science, where they visualize complex concepts and processes. As Text-to-Image (T2I) models become increasingly capable, researchers have started to use them for scientific illustration generation. However, existing benchmarks often assess outputs at a holistic level, overlooking fine-grained elements, while scientific reasoning ability and output conciseness remain under-quantified. We introduce FEPBench, a benchmark built from carefully selected high-quality scientific illustrations across multiple disciplines and layout types. With the assistance of multimodal large language models (MLLMs) and human experts, we provide fine-grained atom set annotations and systematically evaluate T2I models along three dimensions: instruction faithfulness, reasoning enrichment, and semantic precision. Our evaluation further decomposes model performance across visual, textual, relation, and layout elements. Results show that even state-of-the-art (SOTA) closed-source models, such as GPT Image 2 and Nano Banana Pro, still suffer from text-rendering bottlenecks, limited reasoning enrichment, and difficulty balancing generation richness with precision. These findings provide practical guidance for improving and deploying T2I models in scientific illustration generation. Benchmark data, atom set annotations, and evaluation code will be released by us.

Unlike text- or image-driven video generation, an interactive world model is driven by actions: the user acts, and the world responds. Two obstacles stand in the way of fair and comprehensive evaluation. First, models take actions in incompatible formats—captions, camera trajectories, action functions—so no shared protocol has been established. Second, while existing benchmarks have advanced world memory and visual quality, action following is reduced to trajectory or direction error, which collapses a whole path into one number: not how quickly the world reacts to a command switch, nor how cleanly it moves along the commanded axis. WorldMark removes both obstacles. Per-model adapters translate a shared WASD-style vocabulary into each model's native control format, so ten heterogeneous models receive semantically identical instructions across 500 standardized cases spanning styles, viewpoints, and difficulty tiers; a new model costs one adapter. On this common ground we characterize action dynamics through a control-systems lens—direction accuracy, direction purity, response latency, and motion stability, each resolved per axis—alongside suites for world memory and visual quality. Together they expose differences existing protocols cannot see: the fastest responders are often the least stable, a trade-off no single action metric captures; per-axis resolution reveals models that follow translation almost perfectly while barely responding to rotation; the model with the best perceptual and aesthetic quality ranks last in translational direction accuracy and latency; and stylized scenes cost every model global consistency while leaving action dynamics largely intact. We will release all data, evaluation code, and model outputs.