GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling
Kang He, Wenshuo Peng, Chuanhao Li, Zihui Gao, Xiaojie Xu, Zhengyuan Lin, Yuzhe Ding, Donghong Ji 等 10 位作者
AI 速览：作者构建了包含69.7K片段、1,655小时、约3.5亿帧的分钟级游戏视频数据集GameVerse，用于支撑长时程交互式世界模型训练。
Recent advances in interactive world models have made it possible to generate controllable, dynamic environments from action inputs. Yet current models struggle to maintain coherence beyond roughly one minute of rollout, a limitation rooted not in model capacity but in the scarcity of long-horizon training data. Existing gameplay video datasets typically consist of clips under ten seconds, with captions that describe whole clips in coarse and undifferentiated text, which is insufficient for controllable long-horizon generation. We address this with GameVerse, a minute-scale gameplay video dataset purpose-built for long-horizon interactive world modeling. GameVerse contains 69.7K clips spanning 1,655 hours and 349.6 million frames across diverse first- and third-person gameplay, including 1,358 hours of continuous two-minute sequences, an order of magnitude longer than prior datasets. Clips pass through a unified pipeline of scene-consistent segmentation, quality filtering, camera pose estimation, and event-based caption annotation. Each clip is paired with a temporally stable camera trajectory that remains coherent across minutes, together with captions that separate static scene context from dynamic foreground events and align to event boundaries rather than fixed time windows. Comprehensive analysis shows that GameVerse surpasses existing public datasets in scale, diversity, and annotation quality. Experiments further demonstrate that training on GameVerse substantially extends the temporal horizon of interactive world models.
One-step and few-step text-to-image generation must accurately render the details specified in the prompt despite limited opportunities for refinement. Distribution matching distillation (DMD) uses score differences to bring the generated distribution closer to a reference distribution. Following prior work, we estimate this difference with internal velocity branches, avoiding a separate fake-score network. We ask whether an explicit moment constraint improves this learned score supervision. A decomposition of the reverse-KL field into a Gaussian moment component and a non-Gaussian residual identifies statistics that can be constrained explicitly. We introduce Intrinsic Distribution Matching (IDM), which retains the full DMD correction and matches feature means and per-coordinate standard deviations using the generator's exponential-moving-average (EMA) copy, requiring no external feature encoder. Experiments use full-parameter and low-rank adaptation (LoRA) tuning on Z-Image-Turbo 6B and Qwen-Image 20B. At two network evaluations, IDM reaches 94.25% and 90.20% on Chinese and English LongText-Bench, the highest scores among the compared methods at this budget. Controlled ablations show gains from the added moment loss, while generator EMA features perform competitively with an external DINOv2 encoder across the five reported metrics. Code is available at https://github.com/inclusionAI/IDM.
Interactive video world models aim to follow user actions while maintaining scene consistency over extended interactions. History-based methods reuse generated latents to preserve scene content, but three limitations of this reuse contribute to error recirculation. 1) The latent generated at each interaction is retained as history, potentially containing blurred details, geometric distortions, or semantic drift. 2) Although retrieved historical content matches the current scene, it can still contain errors that are carried into the next chunk. 3) New outputs then become history for later interactions, potentially reinforcing the same errors through repeated reuse. To limit this cycle, we aim to predict the future scene state from the current generated latent and the next action before generating the next interaction's video chunk. This future-state prior provides guidance for the next chunk, with the aim of reducing the errors carried forward from history. To this end, we propose a plug-and-play Causal Next-Chunk Predictor (C-NCP) with cross-model adaptability and prediction-guided modulation. We integrate C-NCP into four base generators and evaluate the augmented models across three benchmarks for interactive long-video generation and action-conditioned world modeling. Extensive experiments demonstrate reduced error accumulation and improved long-horizon consistency across diverse generation settings. The project page is at https://GuangYuanKK.github.io/futureworld.
Evaluating autonomous agents requires examining whether they can translate goals into coherent action across extended interactions. In persistent environments, this requires coordinating interdependent actions under incomplete observations, accounting for delayed effects, and adapting as earlier decisions shape subsequent opportunities. We introduce ASIWorld, a survival sandbox that combines persistent world and character state with character-scoped observations and a programmable asynchronous interface to study sustained goal execution. Agents pursue interdependent production and exploration tasks through direct actions or agent-authored programs that continue running between model calls. The evaluation assesses task success, the attainment of development achievements, and survival costs, complemented by an analysis of execution behavior. In an exploratory study of nine agent configurations using a common harness, eight complete the production task and six complete both tasks. Analyses of execution trajectories and performance across difficulty levels reveal different preparation, combat, and recovery behaviors among configurations with similar task outcomes. Incomplete preparation is associated with stalled progression, while configurations that complete the tasks differ in survival costs and the fulfillment of additional achievement criteria. These findings illustrate the value of assessing sustained goal execution through both task outcomes and the behaviors and costs associated with achieving them.
Alzheimer's disease (AD) diagnosis may benefit from gut metabolic information associated with brain-gut interactions, yet existing models largely analyze cerebral glucose metabolism alone. Although whole-body fluorodeoxyglucose positron emission tomography (FDG-PET) can provide paired brain-gut observations, routine brain examinations rarely include gut imaging, limiting direct use of this information for diagnosis. In this paper, we present BRIDGE (Brain and Gut Representation Integration via Distribution-Guided Enhancement), a two-stage framework for transferring brain-gut metabolic knowledge to brain-only diagnosis. Specifically, BRIDGE integrates a conditional brain-to-gut latent transport module and a disease-evidence-guided cerebral pathway to capture subject-specific cross-organ correspondence and disease-relevant cerebral representations. For the cerebral pathway, a large language model (LLM) helps organize literature-derived descriptors into human-approved semantic anchors that guide feature learning. Additionally, we introduce a counterfactual-guided residual fusion mechanism, which uses matched and non-matched virtual-gut evidence to supervise a confidence gate and dynamically regulate the contribution of transported cross-organ information. At inference, BRIDGE exploits complementary gut-related context using brain PET alone, without measured gut PET. Extensive experiments across public and private cohorts demonstrate that BRIDGE outperforms brain-only and cross-organ baselines, achieving AUC gains of 6.0% and 5.2% over its brain-only counterpart for CN/MCI and CN/AD diagnosis, respectively. Notably, its ability to transfer complementary brain-gut metabolic context while preserving brain PET as the sole inference input highlights its potential for clinically practical diagnosis. Our code is available at https://github.com/CUHK-AIM-Group/BRIDGE.
Generating high-quality and controllable camera motion is essential for AI-assisted cinematography, video synthesis, and 3D scene understanding. We introduce TKCAM, a Text- and Keyframe-conditioned CAMera-motion synthesis framework based on generative masked modeling. We represent camera dynamics using a 12-dimensional kinematic feature comprising position, velocity, and a continuous rotation representation and discretize them into hierarchical motion tokens via a Residual Vector Quantizer (RVQ). A two-stage masked transformer architecture then learns to reconstruct and refine these tokens, utilizing explicit self- and cross-attention modules for multimodal conditioning. A central feature of our framework is sparse visual keyframe conditioning: users can provide free-form text prompts together with RGB observations at selected timestamps, which provide temporally localized visual guidance for generating coherent in-between trajectories. Furthermore, to advance evaluation standards, we curate RealEstate10K-Cap, a large-scale text-camera dataset, and establish a cross-domain benchmark with a Universal CLaTr Evaluator. Extensive experiments demonstrate that TKCAM surpasses recent state-of-the-art baselines on Fréchet distance (FID), text-motion matching scores, and retrieval metrics (R@K), while additional analyses evaluate temporal smoothness and cross-domain generalization. Code and RealEstate10K-Cap dataset are available at https://github.com/linearalgebrayhz/TKCAM.
LLM-based search agents are widely used for information-seeking tasks, but their reliance on external tool returns introduces a critical security risk: web content retrieved during execution is untrusted, exposing agents to prompt injection and goal hijacking. Prior work on search-agent safety primarily focuses on static web-content injection, but modern agents issue follow-up queries and cross-check competing sources, so a single injected page is often diluted or rejected. We show that the channel delivering search and page observations is a fragile security boundary: beyond exposing the agent to a single poisoned page, a mediated search interface can repeatedly steer how the agent gathers evidence and forms its final answer. Under a constrained tool-intermediary threat model, appending only one controlled result per query can substantially increase attack success when the evidence is coordinated across the agent's trajectory. We study this setting with a strategy-driven long-horizon attack system and introduce Authority-Chain Hijack (ACH), an expert-refined strategy that turns isolated search-result and page-content manipulations into a coherent evidence chain across seemingly corroborating sources. ACH achieves the highest Overall ASR among all baselines, reaching 55.9% / 83.3% ASR / MaxN ASR on the full SafeSearch test split. We further introduce Trace-Guided Strategy Evolution (TGSE), which automatically improves attacker strategies from execution traces, replacing manual redesign with trace-driven refinement; its strongest single setting reaches 71.4% / 95.0% in held-out evaluation. Code is available at https://github.com/MBOSSLESS/Breadcrumbing-Search-Agents.
Jingong Chen, Linlian Jiang, Guangyuan Li, Xinxin Zuo, Yang Wang, Chulwoo Pack, Kyle Gao, Kwanghee Won
Feed-forward 3D Gaussian Splatting predicts a scene from a few images, but our predictor is less accurate than its own Gaussians optimized per scene on those images. This gap lies mostly in the shape of the Gaussians, their scale and rotation, since optimizing the shape alone recovers almost all of it and opacity and color recover less. Gradient steps on the input views correct this shape, and the model holds these views at inference. Log-anisotropy correction changes the axis ratios of each Gaussian in log space at constant volume, and in-forward gradient refinement renders the Gaussians into the input views and applies the gradient of the photometric loss with a learned recurrent update, first to opacity and color and then to the shape. GradSplat is a feed-forward model that refines its own Gaussians by learned gradient steps inside its forward pass, with no target view. On both complete six-view RealEstate10K protocols, it has the best PSNR and SSIM among nine methods and keeps its accuracy from $6$ to $42$ input views. Without fine-tuning, it is the most accurate on DL3DV and has the best PSNR and SSIM on ScanNet++. The code will be released at https://github.com/Jingong-Chen/GradSplat/.
Hazel H. Kim, Jakub Binkowski, Ekaterina Fadeeva, Juan David Guerra, Yijie Hao, Zheng Yi Ho, Baixiang Huang, Buyun Liang 等 25 位作者
Hallucination remains one of the most critical challenges in deploying large language models (LLMs) reliably and safely. Yet the term is used inconsistently across the research community. We argue for a strict definition of hallucinations in LLMs. First, hallucinations should be confident incorrectness, kept separate from expressed uncertainty (UNK)1. UNK reveals that the model may be wrong, whereas a hallucination does not reveal it and is therefore more dangerous in practice. Conflating the two obscures their distinct behaviors and makes them hard to study. Second, hallucinations should be split by a simple re-elicitation test: does any semantically equivalent prompt recover the correct response? This test probes whether the relevant knowledge2 is accessible. If none does, the knowledge is likely absent or inaccessible, and the error type is a confabulation (CF) that likely requires external knowledge or retraining; if some does, the knowledge is accessible from the model but applied inconsistently, and the error type is a semantic illusion (SI) that simple prompt or decoding changes can mitigate3. Re-evaluating 14 recent hallucination methods under this taxonomy yields three findings. (1) Detection methods flag UNK more often than hallucinated responses (i.e., SI and CF). (2) Mitigation methods are far more effective for SI than for CF. (3) Latent signals are more informative than output-level signals for CF. An audit of recent hallucination papers shows that most conflate these categories. We recommend reporting CF, SI, and UNK separately, and introduce a practical labeling protocol that needs only a modest number of semantically equivalent prompts.
Zhiliang Chen, Alfred Wei Lun Leong, Shao Yong Ong, Apivich Hemachandra, Gregory Kang Ruey Lau, Chuan-Sheng Foo, Zhengyuan Liu, Nancy F. Chen 等 9 位作者
Fine-tuning an LLM to maximize performance on a downstream task requires finding the optimal data and model configuration. This is a black-box optimization problem that Bayesian optimization (BO) addresses by sequentially evaluating training configurations and using observed performance feedback to adaptively guide the search towards better configurations. However, directly applying BO to the joint data-model space suffers from two fundamental challenges: the prohibitive cost of full LLM training at each iteration, and the large number of BO iterations required for effective exploration due to the high problem dimensionality. This paper introduces JoBS, a BO-based approach that co-optimizes data and model configurations without requiring a full training run at every iteration. JoBS allocates an initial portion of the optimization budget to learn a scaling-law-inspired performance predictor that estimates fully trained LLM performance from only a small number of training steps. The remaining budget is then used to run BO with this predictor, eliminating the need for full training runs and enabling JoBS to explore significantly more data and model configurations within the same budget. We analyze JoBS's average regret and derive the optimal budget allocation between predictor learning and BO iterations. Empirically, JoBS outperforms independent data and model optimization methods and existing multi-fidelity BO baselines across a diverse set of downstream tasks. Our code is available at https://github.com/chenzhiliang94/BO-for-LLMs.
Ptychography is a computational imaging method that reconstructs a sample from overlapping diffraction measurements. This technique relies on these overlapping measurements to provide the redundancy information needed for robust reconstruction, but dense scanning increases acquisition time and data volume. Reducing scan overlaps can improve measurement efficiency but weakens the physical constraints available for reconstruction. Diffusion models provide strong learned prior information, offering a promising approach to compensate for the loss of measurement redundancy. However, effectively utilizing this prior information under weak measurement constraints remains a challenge. We propose a progressive coupling posterior sampling (PCPS) method, which allows for a broader exploration of the solution space under high uncertainty while maintaining stronger diffusion-guided consistency as the reconstruction becomes more refined. Experiments on real X-ray ptychographic data at various scan overlaps demonstrate that PCPS consistently improves reconstruction accuracy compared to traditional iterative methods and existing diffusion posterior samplers, especially at low overlaps. These results show that diffusion priors can reduce the measurement redundancy required for accurate ptychography reconstruction.
Zhefan Rao, Meng Chu, Xuan Billy Zhang, Jingfan Chen, Guohao Cai, Runyi Li, Shiyi Du, Xu Huang 等 92 位作者
World Models (WMs) can support Multimodal Agents (MAs) through simulated experience, while Agent trajectories can inform World Model refinement. Evaluating this interaction requires separate evidence about action selection and state evolution. We introduce Dual Exam, which tests Agents under fixed environment response mechanisms and World Models under fixed action interventions across physical, digital, social, and scientific domains. The Agent track compares 10 configurations on 1,088 items in 15 reporting categories. The World Model track contains 400 cases across eight scenarios; 15 configurations have full coverage and one covers six scenarios. GPT 6 Astra leads the case-weighted Agent Index at 31.3, and GPT-Image-2 leads the full-coverage WM Overall at 21.7. These scores use different tasks and metrics and are not directly comparable. A 150-case pilot of one coupled pairing records zero task successes and motivates component-level diagnosis. Results reveal uneven task performance and failures to preserve task-relevant state. Their value for predicting coupled success or sustained reciprocal learning remains to be established.
Xinyuan Li, Matthew Zahr, Panos Stinis, Zecheng Zhang
Multi-fidelity learning uses low-fidelity information (coarse-scale PDE models) together with limited high-fidelity data (fine-scale solutions) to obtain high-accuracy PDE solutions. It is a data-driven approach to downscale reduced-order models to solve PDEs efficiently and accurately. Standard multi-fidelity learning focuses on one operator or PDE; given a new PDE, retraining using additional coarse-fine solution pairs may be required, which is costly. In this work, we propose an in-context multi-fidelity learning framework for learning multiple multi-fidelity operators with one shared model. The model is trained using paired low- and high-fidelity examples as demonstrations from multiple operators, together with a new coarse-scale solution query, to predict the corresponding fine-scale solution. Once trained, given a new PDE operator, a few (fewer than five) coarse-fine demonstration pairs, and a low-cost coarse query, the model predicts the corresponding fine-scale solution without fine-tuning. When the target operator appears in the training dataset but has a different prediction goal, such as predicting beyond the training time interval, no additional demonstrations are needed. Instead, the model uses training coarse-fine pairs from that PDE and information from other training operators. Testing then only requires generating a low-cost coarse-scale solution query, without fine-tuning. Notably, our smallest model contains only 255,364 trainable parameters, corresponding to approximately 456 parameters per training operator for 560 operators.
Shuolin Xu, Guanyuan Li, Boyu Li, Kelu Yao, Xiaogang Xu, Hongchuan Yu, Shuicheng Yan, Xiaobin Hu
Injecting new subjects naturally during streaming video generation requires integrating newly supplied visual references into the existing video while preserving established subjects and the scene. Encoding references independently of the current scene provides little guidance for this integration, while keeping their features in persistent memory can interfere with established context. We propose ReferStream, a framework that requires no additional training and combines reference construction using the current scene with transient cache injection. Joint Latent places reference latents in target regions of the latest scene latent, allowing reference features to be extracted in context. Cache Restoration uses these features to condition the injection chunk, then restores the overwritten global cache entries while retaining the newly generated history. This procedure incorporates new visual conditions without retraining the generator or revising emitted frames. We further introduce ReferStream Bench, a new benchmark with scheduled reference arrivals, evolving instructions, and explicit annotations of subject presence for InShot and MultiShot generation. Evaluation follows the event schedule and measures both the fidelity of new subjects and the preservation of established subjects, alongside video quality, text adherence, and count accuracy. Experiments on ReferStream Bench show higher scores for text adherence, subject similarity, and count accuracy than the compared baselines, including offline methods given all references in advance. Project page: https://xushuolin.github.io/ReferStream/.
Hanwen Cui, Yuanyuan Li, Qing Guo, Zhida Zhang, Chao Chen, Xi Chen, Jie Cao
World Action Models (WAMs) combine future-video prediction with robot action learning, yet improvements in task success alone do not reveal how predictive supervision manifests in execution behavior or whether its benefits persist under task and environmental shifts. We conduct controlled future-video prediction loss ablations within FastWAM and OpenWAM, keeping model architecture, initialization, robot training data, and training steps fixed within each model family. Under clean conditions on seen tasks, predictive supervision increases success rates by 24.38 and 72.91 percentage points, respectively. Trajectory clustering further reveals higher success rates within several coarse trajectory clusters, including cases where cluster frequencies change little, indicating that the gains are not solely reflected in changes in the prevalence of trajectory patterns. However, these benefits do not consistently persist under distribution shifts: improvements on unseen tasks vary across models and tasks, while both models exhibit low success rates under the evaluated environmental randomizations, without a consistent advantage from predictive supervision. We further introduce future-video prediction as an auxiliary training objective for $π_{0.5}$, where a lightweight Transformer predictor used only during training increases success from 82.94% to 89.19%. These findings support future-video prediction as effective auxiliary supervision for action learning, while distinguishing improved execution on familiar tasks from reliable transfer and showing that such gains need not rely on a large pretrained video diffusion transformer (DiT) in the evaluated setting.
World-action models (WAMs) incorporate the prediction of future observations into robot policy learning. However, most WAMs build on billion-parameter vision-language or video-generation backbones, which incur high training cost and inference latency. In this letter, we present GAWM (Geometry-Aware World Model), a compact multi-task WAM that predicts future observations in the latent space of a frozen DINOv3 encoder. GAWM compresses the features of each camera view into a small set of tokens. A lightweight predictor, conditioned on the task rather than on actions, forecasts these tokens at two future horizons. The forecast is supervised with the dense patch features of future frames, which requires it to encode the spatial layout of the future scene. An action decoder directly regresses a chunk of future actions from the current and predicted tokens. With fewer than 100M parameters and no robot or video pretraining, GAWM outperforms billion-parameter WAMs on RoboTwin 2.0 and performs comparably to billion-parameter VLAs on LIBERO. On a dual-arm robot, it achieves lower action-prediction error on held-out demonstrations than a fine-tuned billion-parameter VLA, with substantially lower inference latency. Ablation studies show that the latent world model substantially improves performance when training data or training time is limited. Code is available at https://github.com/gaoxiang12/ga-wm.
Privacy compliance requires re-identification models to selectively forget individuals while preserving their ability to recognize others. For this, we present MoME, a switchable framework that can adapt pre-trained backbones for privacy compliance, allowing the same model to switch between Full mode for standard inference and Core mode for unlearned inference. MoME splits one backbone into core parameters that preserve retrieval and zero-initialized parasitic parameters that carry the full-data knowledge, and training them under gradient isolation so the core never receives a forget-set gradient. Disabling the branch yields Core, which satisfies a request to delete a person's training data, and enabling it yields Full, which retains what the deleted data taught, so honoring the request is a flag flip rather than a retraining run. We prove Core independent of the forget set up to an audited set of coupling channels and introduce the rank-based Forgetting-Rank statistic to audit that independence in a trained model. On Market-1501, VeRi-776 and RegDB, Core reads at a retrained reference under three membership attacks while retrieving above it, and the retrieval advantage carries to ResNet and CLIP backbones. Code and weights are available at https://github.com/Jingong-Chen/MoME.
Test-time training (TTT) enables a deployed language model to internalize incoming context on-the-fly into a small set of fast-weight parameters, rather than merely treating it as transient context for attention. In-Place TTT recently made this idea practical for large language models by repurposing the down-projection of standard MLP blocks as fast weights. However, we find that its benefit largely disappears on realistic web long-context question answering. We attribute this to In-Place TTT's uniform write schedule: every chunk is written with equal strength, so irrelevant content dilutes query-relevant information in fast-weight memory. Building on this insight, we propose Select-TTT, a model-agnostic, query-conditioned rule for In-Place TTT that scales each chunk's write strength by its relevance to the user query. This relevance signal requires no supervision and no new parameters: it is read directly from the model's frozen output head, as the probability mass that a chunk's hidden states assign to the query tokens. A temperature-scaled softmax turns these scores into learning-rate multipliers, leaving the chunk-wise, context-parallel computation of In-Place TTT intact. Select-TTT consistently improves long-context question answering on two benchmarks and two backbones. These findings highlight the value of query-conditioned fast weight writes for practical test-time training on long-context question answering, which is also potentially promising for long-horizon agent tasks involving multiple user queries over evolving contexts.
Panoramic multi-object tracking is important for industrial safety monitoring, wide-area robotic perception, and infrastructure-light deployment in large workspaces. In these settings, the sensing system must provide full-surround coverage, metric geometric cues, and stable target association under wide field-of-view distortion and occlusion. Existing image-plane trackers are tightly coupled to the camera projection and become unreliable in panoramic imagery, while conventional Euclidean 3D formulations introduce redundant directional parameters and do not naturally unify angular, scale, and depth estimation. In this paper, we present $\mathbf{S^3KF}$, a panoramic 3D multi-object tracking framework built on a motorized rotating LiDAR and a quad-fisheye camera rig. The key idea is a geometry-consistent state representation on the unit sphere $\mathbb{S}^2$, where object bearing is modeled by a two-degree-of-freedom tangent-plane parameterization and jointly estimated with box scale and depth dynamics. Based on this state, we derive an extended spherical Kalman filtering pipeline that fuses panoramic camera detections with LiDAR depth observations for multimodal tracking. We further establish a map-based ground-truth generation pipeline using wearable localization devices registered to a shared global LiDAR map, enabling quantitative evaluation without motion-capture infrastructure. Experiments on self-collected real-world sequences show decimeter-level planar tracking accuracy, improved identity continuity over a 2D panoramic baseline in dynamic scenes, and real-time onboard operation on a Jetson AGX Orin platform. These results indicate that the proposed framework is a practical solution for panoramic perception and industrial-scale multi-object tracking.The project page can be found at https://kafeiyin00.github.io/S3KF/.
Self-disclosures on the Web support diverse psychological analyses, from recognizing expressed experiences to identifying symptom-relevant information. We study a shared, question-conditioned decision interface for this heterogeneous problem in content analysis. We organize PsyJudgeBench, comprising 13 annotated datasets and 82 questions across six psychological areas, while preserving source-specific criteria and masks for unobserved labels. We introduce PsyJev, a System-1 model that directly produces distributions over answers specified in each request. Joint supervised training establishes the shared judgment function, followed by Judgment-Anchored RLCD (JA-RLCD), which adapts logit-space reward optimization to heterogeneous psychological judgments through balanced supervision, selective retention of correct SFT predictions, and question-group reward optimization. The Qwen3.5-based 2B, 4B, and 9B models achieve 64.33, 66.21, and 67.56 average test F1, respectively. The 9B model exceeds the strongest external comparison by 2.54 points. Controlled 2B comparisons examine the training stages and anchored refinement. Further analyses examine question-conditioned judgment, inference cost, and probability quality, complemented by qualitative comparisons across models.
As agents collaborate on longer tasks, responsibility may pass between models before the work is complete. The receiving model must build on prior decisions, observations, and actions while addressing the remaining requirements. Evaluating this transition requires connecting the information preserved at handoff to how the receiver continues the work, a relationship that task-success and memory-retention scores alone do not capture. We introduce the Handoff State Continuity Benchmark (HSC-Bench), comprising 1,039 tasks that evaluate two complementary aspects of task takeover: recovering relevant repository evidence and continuing from established conversational, tool, and file-work state. Comparisons across six receiving models show that structured handoff state improves repository evidence recovery for every model and continuation performance for most models. The strongest receivers differ between the two tracks, showing that evidence-recovery rankings alone do not characterize performance when continuing from inherited state. By combining these comparisons with checks of subsequent responses and actions, HSC-Bench evaluates both access to prior evidence and adherence to established commitments and completed work. This provides a basis for selecting receiving models and assessing handoff representations according to the work that must be carried forward.
My Chiffon Nguyen, Tianyi Hu, Vinod Anbalagan, Henry Gagnier, Pouya Sadeghi, Farah Abdou, Tidor-Vlad Pricope, Diego MMS Oliveira 等 164 位作者
As LLMs expand their language coverage, evaluating how well they understand and represent diverse cultures becomes increasingly important. However, many benchmarks remain English-centric, with limited cultural coverage and a focus on factual knowledge over its application in real-world cultural contexts. We propose CulturalRiddles, a benchmark of original, culturally grounded riddles created entirely by humans through a participatory community effort, spanning 62 cultures (45 countries, 51 languages). These riddles are deeply rooted in their respective cultures and require abductive and compositional reasoning, alongside detailed factual and social knowledge, language skills, and an understanding of culturally specific associations. We benchmark 28 LLM configurations using both automatic and large-scale human evaluation. Even frontier models achieve at most 22% accuracy in some communities. We publicly release all artifacts alongside an open contribution process to expand the benchmark's linguistic and cultural coverage.
Video models answer from sampled frames and may miss the decisive moment. A common remedy is to commit to an answer, re-examine the video, and confirm or revise it, a loop now mainly learned with reinforcement learning (RL). We find that RL erodes the initial answer, a phenomenon we term confidence erosion, where confidence means the label-verified reliability of the initial answer instead of self-reported certainty: because errors can be repaired later, a correct initial answer earns less reward, and the initial answer stalls even as the final answer's gain over it rises. We propose the principle confident first, reflective second: the initial answer must stay reliable on its own, above a floor set by a reference policy, and reflection may only add to it. Our method, CoFire, implements this principle with one mechanism for each erosion channel: reward incentives, parameter drift, and gradient interference. CoFire reaches 67.0% mean accuracy on Video-MME, MLVU, and LongVideoBench, 1.7 points above the strongest compared 7B method, and improves initial-answer and final accuracy over outcome-reward RL by 4.1 and 3.7 points on average: a stronger initial answer carries through to a stronger final one.
Multi-path reasoning methods such as self-consistency (SC) sample $K$ reasoning paths and choose the most frequent answer. However, their gains quickly plateau as $K$ increases, and locating this saturation point currently requires fitting a scaling model to accuracies measured at several values of $K$ or stopping online per query. We formalize multi-path LLM reasoning as a diversity combining problem from wireless communications: each path is a noisy channel observation, and the pairwise correlation of path correctness caps the design-effect effective sample size of the vote at a finite ceiling. Generalized least squares (GLS) analysis shows that, for paths with equal gains, the optimal linear combiner of latent embeddings is uniform exactly when every row of the path covariance has the same sum, as under exchangeability, supporting majority vote as the natural default in standard SC while leaving room for weighting or pruning under heterogeneous prompt-template branches. Across 5 models and 12 benchmarks, prompt-template diversity reduces path correlation in $55$ of $60$ model-benchmark cells, with a mean relative reduction of $16\%$ and reductions in every task domain. We derive an Adaptive-K rule that uses a four-path pilot to select $K^*$, retaining $96$--$103\%$ of MV@$K{=}32$ accuracy across Math, QA, and NLU.
Zhuohang Li, Yiding Ma, Yu Shang, Lei Jin, Shengyuan Wang, Jianjie Fang, Zhaolu Wang, Xvyuan Liu 等 29 位作者
World models have emerged as a central paradigm for embodied intelligence, enabling agents to predict action-conditioned future and reason about environmental dynamics. However, existing embodied world model benchmarks are still largely confined to vision-only prediction, offline embodied applications, and simulator-based evaluation, making them insufficient for assessing increasingly comprehensive world models. In this work, we introduce WorldArena 2.0, an expanded benchmark that systematically broadens embodied world model evaluation along three dimensions: modality, functionality, and platform. Along the modality dimension, WorldArena 2.0 extends evaluation from vision-only to visuotactile modalities, enabling assessment of multimodal perception and prediction. Along the functionality dimension, it extends beyond policy evaluation and planning to assess world models as interactive RL environments for policy optimization. Along the platform dimension, it moves beyond simulator-only evaluation to a diverse suite of simulated and real-world robotic settings across multiple embodiments. Under a standardized protocol, WorldArena 2.0 comprehensively evaluates perceptual quality, interactive utility, and cross-platform performance, providing a comprehensive testbed for tracking progress toward embodied world models. The project page including codes and demos is available at: https://world-arena.ai/.
X-Era AI Lab Team, :, Jianhao Lai, Zhihao Zhan, Zhiquan Liao, Qichang Li, Shubin Chen, Zhiyuan Liu 等 28 位作者
World-action models (WAMs) offer a promising framework for embodied intelligence, yet mainstream approaches remain centered on video foundation models or vision-language models. Even when geometric features extracted by 3D models such as VGGT are incorporated, they are typically passed into a video model or VLM that serves as the primary backbone, leaving 3D geometry as an auxiliary input or prediction target rather than the native foundation for world modeling. We present E0.5, a geometry-native world-action model built directly on 3D foundation models. E0.5 introduces two key innovations. First, it adopts a 3D foundation model as the primary backbone, making 3D geometry the native representation for both world modeling and action prediction. Second, it explicitly predicts future 3D states together with actions, enabling the model to capture geometric world dynamics beyond action-only modeling or future video prediction. Experiments on RoboTwin 2.0 first demonstrate the effectiveness of E0.5 without embodied-data pretraining, achieving success rates of 89.90\% and 50.09\% under clean and randomized settings, respectively. Building on these results, we further pretrain E0.5 on more than 160k robot trajectories from AgiBotWorld-Beta, jointly learning robot actions and future geometry within the pretrained 3D foundation. This embodied pretraining improves mean real-world manipulation success rates by approximately 15 and 7 percentage points on AgileX Piper and Zerith H1, respectively. Together, these results demonstrate the potential of geometry-native foundations to acquire robot interaction knowledge and transfer it to physical manipulation. E0.5 embodies this geometry-native approach as a member of the VWA family of world-action models.
Clinical expertise is inherently distributed: it resides not in any single dataset or institution but in the accumulated reasoning experiences of thousands of clinicians navigating diverse patient populations, resource constraints, and evolving treatment protocols. Yet current approaches to scaling medical AI, from centralized large language models to federated learning, remain anchored to a data-centric paradigm that treats knowledge as static parameters to be aggregated. In this Perspective, we argue for a fundamental shift from data aggregation to experience aggregation. We propose Decentralized Agentic Systems (DAS), a framework in which distributed AI agents operating across healthcare institutions synthesize collective clinical intelligence by sharing abstracted reasoning trajectories rather than raw data or model parameters. Unlike federated learning, which averages numerical gradients across homogeneous model architectures, DAS exchanges structured records of clinical reasoning, specifically the sequence of observations, hypotheses, tool invocations, and outcome feedback that constitute diagnostic experience. We examine the architectural principles, deployment considerations, governance requirements, and critical open challenges of this framework, laying the groundwork for a collaborative medical AI ecosystem where distributed expertise autonomously adapts to the complexities of longitudinal patient care.
Recent advances in generative video models have improved visual realism, making detection increasingly dependent on subtle temporal inconsistencies. However, these signals are weak and spatially localized, while most of the video content is redundant. This creates a fundamental mismatch: existing detectors rely on heavy backbones to model full videos, which leads to high computational cost and limited scalability. In this work, we rethink AI-generated video detection from a representation perspective. Instead of modeling entire videos, we restructure them into compact temporal slice representations that isolate informative temporal evolution. We introduce a lightweight framework using frequency-aware temporal slice consistency learning, where informative spatial rows and columns are aggregated over time to form slice-domain inputs. This reparameterization suppresses redundant appearance while preserving discriminative temporal dynamics, enabling efficient and targeted detection. By constructing slices from both high- and low-frequency regions, our method exposes complementary temporal artifacts, including unstable details and abnormal motion smoothness. We instantiate this idea in a Temporal Slice Consistency Network, which integrates task-driven slice localization, hierarchical positional encoding, and a lightweight Transformer to model cross-slice and cross-frequency dependencies. We further introduce AI-Artist, a new benchmark with artist-curated videos from recent high-fidelity generators, including Seedance and Kling. Experiments on GenVideo and AI-Artist show that our method achieves strong performance while requiring over 400$\times$ fewer model-forward FLOPs and 3.6$\times$ faster model inference than the strong video-based baseline ReStraV (2.9$\times$ faster end to end, including preprocessing), enabling scalable and practical deployment.
Large vehicles in cluttered environments are often blocked by movable obstacles, motivating mobile robot teams to clear traversable corridors. Navigation Among Movable Obstacles (NAMO) approaches often abstract obstacle displacements while neglecting robot dimensions, obstacle mass, contact geometry and forces, yielding physically infeasible plans. PushAround is a physics-informed hybrid search that couples geometric connectivity reasoning with short-horizon simulation to construct executable clearing plans. Repeated W-clearance connectivity graph (WCCG) queries expose frontier gaps requiring widening and rank them by estimated downstream costs. The search generates direct or recursive pushing candidates, selecting obstacles, directions, contacts and forces, and admits physics-evaluated candidates. Numerical experiments validate WCCG against standard connectivity checks and demonstrate scaling to denser clutter and teams of up to four robots, with robustness to simulated execution-side physical-parameter mismatch.
Zefeng He, Yichen He, Yuan Lin, Yang Yang, Yu Cheng
Large language models (LLMs) can produce plausible but incorrect answers when they are uncertain, which can mislead users and undermine model reliability. Training models to abstain with reinforcement learning is a natural approach, but most existing methods use a fixed error penalty and therefore cannot adapt their behavior to different confidence requirements at inference time. We study \emph{explicit confidence targets}, where a user specifies a target \(p\) in the prompt and the model directly chooses whether to answer or abstain, with the goal of achieving an accuracy of at least \(p\) among answered questions. However, learning to respond to different confidence targets with a single model has proven difficult, and prior attempts have failed to distinguish different targets. We show that this failure comes from two problems: the reward signal that distinguishes different values of $p$ is often lost during optimization, while the original error penalty grows without bound as $p$ approaches one. We propose RLEC (\textbf{R}einforcement \textbf{L}earning with \textbf{E}xplicit \textbf{C}onfidence Targets), which preserves the learning signal needed to distinguish confidence targets, introduces a stable reward family that keeps the desired answer condition unchanged, and uses reward scheduling to balance effective learning and training stability. Together, these designs enable the first successful training of a single LLM for explicit confidence target control. Extensive experiments across diverse benchmarks and model scales show that RLEC largely preserves general capabilities and achieves competitive calibration performance. RLEC also supports fine-grained target control without an external decision system or threshold calibration after training. By making explicit confidence targets part of the trained policy, RLEC opens a new and more direct path toward behavioral calibration.