# World models

World models are generative neural-network models that learn compressed spatial and temporal representations of reinforcement-learning environments and can provide inputs for an agent's policy.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## SteerTrace: Tracing How Target Attacks Steer World Action Models

DevFeed: [SteerTrace: Tracing How Target Attacks Steer World Action Models](<https://devfeed.tech/articles/steertrace-tracing-how-target-attacks-steer-world-action-models-145575.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25719>)

Author: Yukun Dai, Tianshi Wang, Fengling Li, Jingjing Li, Xin Li, Lei Zhu

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Computer vision](<https://devfeed.tech/topics/computer-vision.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [a-safety](<https://devfeed.tech/tags/a-safety.md>), [actions](<https://devfeed.tech/tags/actions.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [budget](<https://devfeed.tech/tags/budget.md>), [camera](<https://devfeed.tech/tags/camera.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [effect](<https://devfeed.tech/tags/effect.md>), [experts](<https://devfeed.tech/tags/experts.md>), [ffn](<https://devfeed.tech/tags/ffn.md>), [frame](<https://devfeed.tech/tags/frame.md>), [future](<https://devfeed.tech/tags/future.md>), [inside](<https://devfeed.tech/tags/inside.md>), [layers](<https://devfeed.tech/tags/layers.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [noise](<https://devfeed.tech/tags/noise.md>), [target](<https://devfeed.tech/tags/target.md>), [tracing](<https://devfeed.tech/tags/tracing.md>), [world](<https://devfeed.tech/tags/world.md>)

### AI overview

SteerTrace reports that a single per-model perturbation bounded at 4/255, shared across frames, tasks, and camera views, steers actions toward an attacker-chosen target across eight world action models on LIBERO and RoboTwin, bringing task success close to zero. Same-budget noise and BadWAM cause little shift toward the target. Tracing the effect shows different dependencies: under the standard attack, restoring future tokens removes 61-94% of the targeted effect in five models whose action experts attend to them. When predicted futures are kept close to clean predictions, current-frame tokens become more important in three models: restoring them removes 87-99% of the effect, versus 9-15% for future tokens. In six models, restoring 0.36% of video-expert FFN neurons in layers 10-29 removes 72-96% of the effect. The abstract also reports that residual directions 2-16 contain only 11-18% of shift energy, yet removing them cancels most of the effect.

### Source excerpt

arXiv:2611.25719v1 Announce Type: cross Abstract: World action models (WAMs) predict future videos and actions with one backbone, and their imagined future is often seen as a safety signal. We introduce SteerTrace, a targeted universal attack on WAMs that also traces its effect inside the model. One $\ell_\infty$-bounded (4/255) perturbation per model, shared across frames, tasks, and camera views, pushes the executed action toward an attacker-chosen target. On eight WAMs from four families (FastWAM, Faster-WAM, OpenWAM and LingBot-VA) on LIBERO and RoboTwin, it drives task success close to zero and moves the attacked action toward the target. Same-budget noise and BadWAM produce little action shift in the target direction. We use the signed targeted effect to measure how much of the effect each intervention removes or reproduces. Under the standard attack, restoring future tokens removes 61-94% of the targeted effect in the five models whose action experts attend to them. When the attack is constrained to keep the predicted future close to the clean prediction, the targeted effect can become more dependent on current-frame tokens. In three of these five models, restoring current-frame tokens removes 87-99% of the effect, compared with 9-15% for future tokens. In six models, restoring just 0.36% of the FFN neurons in layers 10-29 of the video expert removes 72-96% of the targeted effect. Residual directions 2-16 hold only 11-18% of the shift energy, yet removing them cancels most of the targeted effect.

## When and How Much to Imagine: Adaptive Test-Time Scaling with World Models for Visual Spatial Reasoning

DevFeed: [When and How Much to Imagine: Adaptive Test-Time Scaling with World Models for Visual Spatial Reasoning](<https://devfeed.tech/articles/when-and-how-much-to-imagine-adaptive-test-time-scaling-with-world-models-for-visual-spatial-reasoning-145315.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2602.08236>)

Author: Shoubin Yu, Yue Zhang, Zun Wang, Jaehong Yoon, Huaxiu Yao, Mingyu Ding, Mohit Bansal

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [grpo](<https://devfeed.tech/topics/grpo.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [imagination](<https://devfeed.tech/tags/imagination.md>), [scaling](<https://devfeed.tech/tags/scaling.md>), [spatial-reasoning](<https://devfeed.tech/tags/spatial-reasoning.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

Visual imagination can help spatial reasoning when a scene must be assessed from an unseen viewpoint, but unnecessary or excessive imagined evidence can waste computation or reduce accuracy. The paper introduces AVIC, which checks whether current visual evidence is sufficient before selectively invoking and scaling a world model, and AVIC-R, which learns that policy using GRPO with answer-correctness rewards and imagination-cost penalties, without labels for when or how much to imagine. On SAT, MMSI, and R2R, selective control reportedly matches or outperforms fixed imagination strategies with fewer world-model calls and language tokens. AVIC-R also reportedly surpasses GPT-4o and GPT-4.1 while using the world model less often. The supplied abstract does not provide quantitative results or benchmark-specific details.

### Source excerpt

arXiv:2602.08236v3 Announce Type: replace-cross Abstract: Despite rapid progress in MLLMs, visual spatial reasoning remains unreliable when correct answers depend on how a scene would appear under unseen or alternative viewpoints. Recent work addresses this by augmenting reasoning with world models for visual imagination, but questions such as when imagination is actually necessary, how much of it is beneficial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagination can increase computation and even degrade performance by introducing misleading evidence. In this work, we present an in-depth analysis of test-time visual imagination as a controllable resource for spatial reasoning. We first study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnecessary imagination affects accuracy and efficiency. To support this analysis, we then introduce AVIC, an adaptive test-time framework with world models that explicitly reasons about the sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Finally, to further learn this gating and planning behavior without any annotation of when and how much to imagine, we introduce AVIC-R, which trains the policy via GRPO from QA-correctness rewards and penalties by imagination cost. Across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our results reveal clear scenarios where imagination is critical, marginal, or detrimental, and show that selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens. Our AVIC-R surpasses strong proprietary baselines including GPT-4o and GPT-4.1 while invoking the world model less often. Overall, our findings highlight the importance of analyzing and controlling test-time imagination for efficient and reliable spatial reasoning.

## Physical Laws without Physics Loss? Probing Learned Dynamics in Video World Models through Forced Response

DevFeed: [Physical Laws without Physics Loss? Probing Learned Dynamics in Video World Models through Forced Response](<https://devfeed.tech/articles/physical-laws-without-physics-loss-probing-learned-dynamics-in-video-world-models-through-forced-response-145081.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25626>)

Author: Yutao Li, Chen Wang, Wangmeng Zuo, Hui Li

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Computer vision](<https://devfeed.tech/topics/computer-vision.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>)

Tags: [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [flow-matching](<https://devfeed.tech/tags/flow-matching.md>), [image-to-video](<https://devfeed.tech/tags/image-to-video.md>), [laws](<https://devfeed.tech/tags/laws.md>), [motion](<https://devfeed.tech/tags/motion.md>), [object](<https://devfeed.tech/tags/object.md>), [physics](<https://devfeed.tech/tags/physics.md>), [probing](<https://devfeed.tech/tags/probing.md>), [response](<https://devfeed.tech/tags/response.md>), [training](<https://devfeed.tech/tags/training.md>), [video](<https://devfeed.tech/tags/video.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

The study tests whether video world models learn dynamics beyond replay by shaking the support of suspended objects and evaluating forced responses. Six image-to-video backbones are adapted with visual and numeric conditioning using only flow-matching loss. All six reproduce linear response, including resonance, and five capture higher-order effects such as nonlinear softening, despite softening accounting for only 0.74% of the swing's squared displacement in training. Wan generalizes to unseen excitations, objects, poses, and scenes; counterfactual tests suggest it does more than replay training videos. Extensions to a double pendulum and rocking block find sustained dynamics are learned more faithfully than brief impacts. The excerpt also cautions that learned dynamics may not emerge when state estimation or propagation fails.

### Source excerpt

arXiv:2611.25626v1 Announce Type: cross Abstract: Do video world models learn physical laws without a physics loss, or merely imitate motion? Overall motion and appearance scores cannot settle it. We investigate this through nonlinear forced response: an excitation shakes the support of a suspended object. We introduce dual visual-numeric conditioning and adapt six image-to-video backbones to predict this motion using only the flow-matching loss. Wan predicts motion accurately and generalizes to unseen excitations, objects, poses, and scenes. Counterfactual tests suggest capabilities beyond replaying training videos. But what laws have they learned? We design three tiers of tests using Taylor expansions of the dynamics and geometry. All six reproduce the linear response, including resonance; five capture higher-order dynamics, including nonlinear softening. Remarkably, softening is barely present in training: only 0.74% of the swing's squared displacement. Extensions to a double pendulum and a rocking block show sustained dynamics are learned more faithfully than brief impacts. Yet learned dynamics may fail to appear when state estimation or propagation goes wrong. We turn these tests and findings into strategies to improve video training and generation.

## MultiWorldBench: Do Independently Controlled Views Describe One Shared World?

DevFeed: [MultiWorldBench: Do Independently Controlled Views Describe One Shared World?](<https://devfeed.tech/articles/multiworldbench-do-independently-controlled-views-describe-one-shared-world-145396.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.11723>)

Author: Zhangbo Xu, Ruoxi Zhang, Rui Hu, Yisong Wang, Xitong Li, Jingdong Wang

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [multiplayer](<https://devfeed.tech/topics/multiplayer.md>), [synchronization](<https://devfeed.tech/topics/synchronization.md>), [race-condition](<https://devfeed.tech/topics/race-condition.md>), [Benchmark](<https://devfeed.tech/topics/benchmark.md>)

Tags: [automatic](<https://devfeed.tech/tags/automatic.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [concurrent](<https://devfeed.tech/tags/concurrent.md>), [control](<https://devfeed.tech/tags/control.md>), [correlation](<https://devfeed.tech/tags/correlation.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [gamma](<https://devfeed.tech/tags/gamma.md>), [identity](<https://devfeed.tech/tags/identity.md>), [independent](<https://devfeed.tech/tags/independent.md>), [minecraft](<https://devfeed.tech/tags/minecraft.md>), [motion](<https://devfeed.tech/tags/motion.md>), [multiplayer](<https://devfeed.tech/tags/multiplayer.md>), [persistence](<https://devfeed.tech/tags/persistence.md>), [race-condition](<https://devfeed.tech/tags/race-condition.md>), [ranking](<https://devfeed.tech/tags/ranking.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [solaris](<https://devfeed.tech/tags/solaris.md>), [spatial](<https://devfeed.tech/tags/spatial.md>), [synchronization](<https://devfeed.tech/tags/synchronization.md>), [world](<https://devfeed.tech/tags/world.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

MultiWorldBench tests whether multiplayer world models keep independently controlled Minecraft views coherent across 495 configurations, seven task suites, and ten capabilities. Among the generated systems, Gamma-World has the highest average score (21.39), narrowly ahead of Solaris (MP) (20.88); Solaris (SP) scores 10.06, while Engine GT reaches 91.69. Gamma-World leads on several control, shared-state, and revisit capabilities, while Solaris (MP) leads on cross-view motion and race-condition consistency. All generated systems score at most 8.00 on persistence and 1.33 on structural consistency, and none succeeds in spatial reasoning or preserving building identity. Human pairwise judgments match the overall ranking and align strongly with automatic evaluation (mean dimension-level Spearman correlation 0.95).

### Source excerpt

arXiv:2610.11723v5 Announce Type: replace Abstract: Multiplayer world models must ensure that independently controlled views remain consistent with a single shared and persistent world. We introduce MultiWorldBench, a diagnostic Minecraft benchmark comprising 495 case configurations across seven task suites and ten capabilities, including independent control, cross-view motion, shared-state synchronization, persistence, structural reasoning, concurrent interaction, and delayed revisit. Using Engine GT as a reference, we evaluate two multiplayer world models, Solaris (MP) and Gamma-World, together with a single-player generation variant, Solaris (SP). Among the generated systems, Gamma-World achieves the highest ten-capability average at 21.39, followed by Solaris (MP) at 20.88 and Solaris (SP) at 10.06, while Engine GT reaches 91.69. Gamma-World performs better on several control, shared-state, and revisit capabilities, whereas Solaris (MP) leads in cross-view motion and race-condition consistency. Nevertheless, all generated systems score at most 8.00 on state persistence and 1.33 on structural consistency, and none succeeds in spatial reasoning or building-identity preservation. Human pairwise evaluation yields the same overall system ranking and shows strong alignment with the automatic evaluation, with a mean dimension-level Spearman correlation of 0.95. These results show that current multiplayer world models still exhibit a substantial gap relative to Engine GT in maintaining a coherent shared world.

## Hierarchical Spatiotemporal World Modeling with Content-Adaptive Spatial Addressing

DevFeed: [Hierarchical Spatiotemporal World Modeling with Content-Adaptive Spatial Addressing](<https://devfeed.tech/articles/hierarchical-spatiotemporal-world-modeling-with-content-adaptive-spatial-addressing-145078.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25610>)

Author: Tiantian Li, Junfeng Lu, Yifei Wang, Zhaoyang Zhang, Lei Wang, Tianyi Qian

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [attention](<https://devfeed.tech/tags/attention.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [capture](<https://devfeed.tech/tags/capture.md>), [coding](<https://devfeed.tech/tags/coding.md>), [content](<https://devfeed.tech/tags/content.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ne](<https://devfeed.tech/tags/cs-ne.md>), [design](<https://devfeed.tech/tags/design.md>), [drift](<https://devfeed.tech/tags/drift.md>), [effective](<https://devfeed.tech/tags/effective.md>), [environments](<https://devfeed.tech/tags/environments.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [grid](<https://devfeed.tech/tags/grid.md>), [horizon](<https://devfeed.tech/tags/horizon.md>), [inductive](<https://devfeed.tech/tags/inductive.md>), [layer](<https://devfeed.tech/tags/layer.md>), [modeling](<https://devfeed.tech/tags/modeling.md>), [spatial](<https://devfeed.tech/tags/spatial.md>), [world](<https://devfeed.tech/tags/world.md>)

### AI overview

HiSTFormer targets long-horizon prediction in partially observable environments, where localization drift can erode spatial memory. It combines layer-wise local temporal RoPE with multi-periodic spatial RoPE and content-conditioned address offsets to represent recent context and long-range spatial structure. Its fixed-capacity memory bounds KV storage and per-step attention cost as sequences grow. The abstract reports that on MemoryMaze and HM3D, HiSTFormer outperforms strong world-model baselines within the training sequence length and generalizes substantially better to prediction horizons well beyond training, while remaining compact. The design also retains consistent gains as Transformer depth increases. The supplied text does not provide quantitative results or implementation details.

### Source excerpt

arXiv:2611.25610v1 Announce Type: cross Abstract: Long-horizon world modeling in partially observable environments requires compact memories that preserve spatial structure despite accumulated localization drift. Inspired by hippocampal episodic and spatial memory and the multi-periodic coding of entorhinal grid cells, we propose HiSTFormer, a hierarchical spatiotemporal world model, which adopts a layer-wise organization of local temporal RoPE and multi-periodic spatial RoPE, together with content-conditioned address offsets, to capture recent temporal context and long-range spatial structure. A fixed-capacity memory keeps KV storage and per-step attention cost bounded as the sequence grows. On MemoryMaze and HM3D, HiSTFormer outperforms strong world-model baselines within the training sequence length and generalizes substantially better to prediction horizons far beyond those seen during training, while using a compact model. The hierarchical spatiotemporal design remains effective across deeper Transformer stacks, with consistent gains as depth increases. These results show that hierarchical spatiotemporal organization and content-adaptive spatial addressing provide effective inductive biases for efficient long-horizon world modeling.

## PsyWorld: Profile-Conditioned World Modeling for Human Behavior Prediction

DevFeed: [PsyWorld: Profile-Conditioned World Modeling for Human Behavior Prediction](<https://devfeed.tech/articles/psyworld-profile-conditioned-world-modeling-for-human-behavior-prediction-144998.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24999>)

Author: Ziyuan Qin, Yang Qiao, Summer B. Frandsen, Michael T. Treadway, Liang Zhao

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [announce](<https://devfeed.tech/tags/announce.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [behavioral-prediction](<https://devfeed.tech/tags/behavioral-prediction.md>), [behavioral-science](<https://devfeed.tech/tags/behavioral-science.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [encoder](<https://devfeed.tech/tags/encoder.md>), [environment](<https://devfeed.tech/tags/environment.md>), [modeling](<https://devfeed.tech/tags/modeling.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [profile](<https://devfeed.tech/tags/profile.md>), [simulator](<https://devfeed.tech/tags/simulator.md>), [split](<https://devfeed.tech/tags/split.md>), [stable](<https://devfeed.tech/tags/stable.md>), [state](<https://devfeed.tech/tags/state.md>), [supervision](<https://devfeed.tech/tags/supervision.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>), [test](<https://devfeed.tech/tags/test.md>), [training](<https://devfeed.tech/tags/training.md>), [type](<https://devfeed.tech/tags/type.md>), [world](<https://devfeed.tech/tags/world.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

### AI overview

PsyWorld uses observed environment-behavior state transitions to supervise behavior prediction, rather than treating a static psychometric profile as a direct action descriptor. In training, a posterior encoder quantizes realized transitions and a prior predicts those codes from the profile and pre-transition context; the prior code and predicted next state then condition macro-level behavior and simulator-grounded execution. Evaluation uses participant-disjoint data in VEET and a structurally different synthetic environment, AgentSociety. On VEET, PsyWorld reports 0.904 Top-1 accuracy (95% CI 0.895-0.913) and 0.0033 Macro KL. Ablations link most of the Top-1 gain to continuous next-state modeling, while discrete codes and prior alignment mainly improve Top-3 and distributional alignment. The results support transition-informed supervision for prediction on unseen participants, but do not establish causal psychological mechanisms or calibrated long-horizon state dynamics.

### Source excerpt

arXiv:2611.24999v1 Announce Type: new Abstract: Understanding how stable psychometric traits and time-varying states jointly inform behavior is an important problem in computational behavioral science. Existing personalized agents commonly condition actions directly on a static profile, leaving realized state transitions unused as an explicit source of supervision. We introduce a profile-conditioned state world model that uses observed environment--behavior state transitions to learn transition-informed representations for behavioral prediction. During training, a posterior encoder quantizes realized transitions, while a prior predicts the resulting codes from the psychometric profile and pre-transition context. The prior code and predicted next state then condition macro-level behavior and simulator-grounded execution. Thus, the world model provides structured transition supervision instead of treating personality as a direct action descriptor. We evaluate the framework in VEET, a controlled behavioral simulator with human trajectories, and in AgentSociety as a structurally different synthetic environment. A participant-disjoint split of 240 training, 25 validation, and 48 test participants evaluates generalization to unseen people. On VEET, the model reaches 0.904 Top-1 accuracy (95\% CI: [0.895, 0.913]) and 0.0033 Macro KL. Ablations associate continuous next-state modeling with most of the Top-1 gain, while discrete codes and prior alignment mainly improve Top-3 and distributional alignment. These results establish transition-informed world-model supervision as useful for subject-generalized behavior prediction, without implying causal psychological mechanisms or calibrated long-horizon state dynamics.

## Drifting, Discriminating, and Directioning for Action Coordination in Higher-dimensional Action Space

DevFeed: [Drifting, Discriminating, and Directioning for Action Coordination in Higher-dimensional Action Space](<https://devfeed.tech/articles/drifting-discriminating-and-directioning-for-action-coordination-in-higher-dimensional-action-space-145048.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25385>)

Author: Taisuke Kobayashi

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Model Development](<https://devfeed.tech/topics/model-development.md>), [World models](<https://devfeed.tech/topics/world-models.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [control](<https://devfeed.tech/tags/control.md>), [coupling](<https://devfeed.tech/tags/coupling.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [policy](<https://devfeed.tech/tags/policy.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [rl](<https://devfeed.tech/tags/rl.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [spaces](<https://devfeed.tech/tags/spaces.md>), [training](<https://devfeed.tech/tags/training.md>), [variance](<https://devfeed.tech/tags/variance.md>)

### AI overview

The method combines a drifting policy model with three training and action-selection steps to capture coordination in high-dimensional reinforcement-learning actions. It co-optimizes the model's source distribution as a diagonal policy so the drifting model can represent coordination as residuals, uses adversarial training with a discriminator before and after drifting to promote that coordination, and applies best-of-N sampling during rollouts to guide action selection. Across three locomotion tasks, it reportedly improves sample efficiency and final control performance over conventional diagonal distribution modeling. The analysis attributes the gains to reduced variance in some action dimensions, effectively coupling them with other actions.

### Source excerpt

arXiv:2611.25385v1 Announce Type: new Abstract: This paper proposes a lightweight policy model to mitigate the difficulty of handling higher-dimensional action spaces in reinforcement learning (RL), alongside an algorithm to effectively train it. To address this, we focus on drifting models, which can implicitly represent complex distributions and infer actions in a single forward pass. However, while drifting models can be straightforwardly trained via backpropagation of the Q-gradient, this alone cannot be expected to appropriately capture the desired action coordination. Therefore, this paper constructs a three-step learning algorithm to encourage the action coordination. First, the source distribution of the drifting model is co-optimized as a diagonal policy model. This conditions the drifting model to attach the coordination, which cannot be expressed by a diagonal distribution, as residuals. Second, a discriminator model before and after drifting is introduced with adversarial training for a regularization term. This actively promotes the extraction of the coordination that fundamentally separates the two distributions. Third, best-of-N sampling is applied for action selection during rollouts in order to give directionality toward/beyond the regions where the coordination is insufficiently extracted/sufficiently discovered. The proposed method consistently improves the sample efficiency and the final control performance over the conventional diagonal distribution modeling across three locomotion tasks with high-dimensional action spaces. Furthermore, the analysis indicates that this improvement is achieved by reducing variance in certain action dimensions, effectively coupling them with other actions as action coordination.

## Fast-JEPA: Advancing the Success-Latency Frontier for Latent World Models

DevFeed: [Fast-JEPA: Advancing the Success-Latency Frontier for Latent World Models](<https://devfeed.tech/articles/fast-jepa-advancing-the-success-latency-frontier-for-latent-world-models-145611.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25927>)

Author: Yao Zhang, Bin Yu, Shijie Lian, Haishan Liu, Changti Wu, Yu Wu, Kai Chen

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [candidate](<https://devfeed.tech/tags/candidate.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [execution](<https://devfeed.tech/tags/execution.md>), [faster](<https://devfeed.tech/tags/faster.md>), [latency](<https://devfeed.tech/tags/latency.md>), [planning](<https://devfeed.tech/tags/planning.md>), [search](<https://devfeed.tech/tags/search.md>), [shift](<https://devfeed.tech/tags/shift.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

Fast-JEPA lowers the cost of latent-world-model planning by using a goal-directed plan to guide search and directly predicting complete candidate outcomes, avoiding recursive state rollout during refinement. Across PushT, Reacher, Cube, and TwoRoom, it reports 96.50% average success and 2.41x faster planning than INTACT; search-budget sweeps indicate competitive success and low latency across candidate counts and refinement depths. The authors also introduce DynaGoal to test planning under dynamic out-of-distribution conditions, deadlines, and controlled response delays. Fast-JEPA reports 73.83% average success on DynaGoal-Shift and 87.67% on DynaGoal-Deadline. These results are those reported in the supplied abstract.

### Source excerpt

arXiv:2611.25927v1 Announce Type: cross Abstract: Latent world models predict the outcomes of action sequences, but comparing candidate plans through recursive rollout remains expensive at test time. We introduce Fast-JEPA, a latent world model planning framework that advances the success-latency Pareto frontier of predictive search. Fast-JEPA reorganizes predictive computation to support effective action-sequence search at low planning cost. It constructs a goal-directed plan to guide search and directly predicts the outcomes of complete candidates for efficient comparison, avoiding recursive state rollout during refinement. Across PushT, Reacher, Cube, and TwoRoom, Fast-JEPA achieves 96.50% average success rate with 2.41x faster planning than INTACT. Search-budget sweeps show that its competitive success and low latency extend across candidate counts and refinement depths. We further introduce DynaGoal, a benchmark for evaluating world-model planning under dynamic out-of-distribution conditions, execution deadlines, and controlled response delays. Fast-JEPA achieves average success rates of 73.83% on DynaGoal-Shift and 87.67% on DynaGoal-Deadline, demonstrating the practical value of low-latency predictive planning under changing control conditions and execution-time constraints. Our code is available at https://github.com/Sofilyzia/Fast-JEPA.

## What One Sample Cannot See: Why Fewer Denoising Steps Can Score Higher on Physics Benchmarks

DevFeed: [What One Sample Cannot See: Why Fewer Denoising Steps Can Score Higher on Physics Benchmarks](<https://devfeed.tech/articles/what-one-sample-cannot-see-why-fewer-denoising-steps-can-score-higher-on-physics-benchmarks-145083.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25632>)

Author: Yutao Li, Chen Wang

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>)

Tags: [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [image-to-video](<https://devfeed.tech/tags/image-to-video.md>), [metric](<https://devfeed.tech/tags/metric.md>), [models](<https://devfeed.tech/tags/models.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [systems](<https://devfeed.tech/tags/systems.md>), [video](<https://devfeed.tech/tags/video.md>), [videos](<https://devfeed.tech/tags/videos.md>), [world](<https://devfeed.tech/tags/world.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

Physics-IQ can reward video generators for producing less varied futures because it compares each generated video with a single recording, even when a scene has multiple plausible outcomes. The authors introduce Kernel Physics-IQ, which uses two independent samples per condition and is designed to remove this incentive in expectation. In tests, two-step sampling lowered Kernel Physics-IQ scores across all five image-to-video models; for CogVideoX, the official score rose by 0.055 while the proposed score fell by 0.131. The authors also report that changing the sampling schedule exposes differences missed by official scores, and that in controlled physical systems with known outcome probabilities, their metric favors the true distribution over collapsed predictions.

### Source excerpt

arXiv:2611.25632v1 Announce Type: cross Abstract: When a storm can head north or south, a forecast down the middle may score better under squared error than either plausible path. The same paradox appears when video generators are tested as world models: fewer denoising steps can raise physics scores even as videos blur or stop moving early. Why? Physics-IQ compares each generated video with one recording, although a scene can have multiple valid futures. An exact decomposition shows its scores can reward reduced variation among generated futures, even when their distribution becomes less accurate. We introduce Kernel Physics-IQ to remove this incentive in expectation, needing only two independent samples per condition. Two-step sampling lowers Kernel Physics-IQ in all five image-to-video models. For CogVideoX, the official score rises by 0.055 while ours falls by 0.131. Changing the sampling schedule reveals further differences official scores miss. In controlled physical systems with known outcome probabilities, our metric favors the true distribution over collapsed predictions. In short, a world model should be judged not only by how closely one video matches what happened, but by how faithfully its predictions represent what could have happened.

## World Models Favor Learned Priors over Visual Evidence

DevFeed: [World Models Favor Learned Priors over Visual Evidence](<https://devfeed.tech/articles/world-models-favor-learned-priors-over-visual-evidence-145665.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26366>)

Author: Han Xiao, Zhiyong Bao, Xiaotong Liu, Yanhao Chen, Qingqiang Wu

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bias](<https://devfeed.tech/tags/bias.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [inference](<https://devfeed.tech/tags/inference.md>), [motion](<https://devfeed.tech/tags/motion.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [training](<https://devfeed.tech/tags/training.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

The abstract reports that world models can favor common physical rules over visual evidence when training data are imbalanced. TailWorld evaluates transfer of rules demonstrated in short calibration videos to a different initial state, varying training imbalance and calibration informativeness separately; it measures bias by comparing predictions with futures under the demonstrated and a common rule. The proposed RIFT method learns how rules affect object motion, infers the rule from calibration interactions, and controls the influence of training frequency. On TailWorld, RIFT improves prediction accuracy and physical consistency across imbalances and distribution shifts, including for rare dynamics.

### Source excerpt

arXiv:2611.26366v1 Announce Type: new Abstract: World models are increasingly used to simulate physical interactions, but training a separate model for each environment limits scalability. Cross-environment prediction requires identifying the current physical rule, which appearance alone may not reveal. We therefore study visual calibration: using object motion and interactions in a short video as direct evidence to infer the current rule and predict motion under that rule from a different initial state. However, differences in data availability across physical conditions can yield a long-tailed training distribution, biasing models toward common rules even when evidence supports rare ones. To study this challenge, we introduce TailWorld, which evaluates the transfer of demonstrated rules across initial states while independently controlling training imbalance and calibration informativeness. The benchmark compares predictions with two reference futures from the same prediction initial state, one following the demonstrated rule and the other a common rule, to measure prediction bias toward common dynamics. We further propose Rule Inference for Transfer (RIFT), which learns how different rules affect object motion, infers the current rule from calibration interactions, and applies it to prediction from a new initial state. RIFT explicitly controls the influence of training frequency on rule inference, allowing sufficiently informative observational evidence to support rare rules. Experiments on TailWorld show that RIFT improves prediction accuracy and physical consistency across varying training imbalances and distribution shifts, with gains in predicting rare dynamics.

## MAWAM: A Multi-Agent World-Action Model for Coordinated Manipulation

DevFeed: [MAWAM: A Multi-Agent World-Action Model for Coordinated Manipulation](<https://devfeed.tech/articles/mawam-a-multi-agent-world-action-model-for-coordinated-manipulation-145467.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25105>)

Author: Xianzhe Fan, Chi-Pin Huang, Fu-En Yang, Jing Wang, Yecheng Wu, Yu-Cheng Chou, Johan Bjorck, Linxi "Jim" Fan, Jan Kautz, Yu-Chiang Frank Wang, Zhiding Yu, Hengshuang Zhao, Zhiqi Li

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [Robot Manipulation](<https://devfeed.tech/topics/robot-manipulation.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [manipulation](<https://devfeed.tech/tags/manipulation.md>), [model](<https://devfeed.tech/tags/model.md>), [multi-agent](<https://devfeed.tech/tags/multi-agent.md>), [robotics](<https://devfeed.tech/tags/robotics.md>)

### AI overview

MAWAM predicts a robot team's physical future alongside its actions: it jointly generates each robot's future views and executable action chunks, using cross-agent attention while keeping separate control interfaces. On three real-robot tasks, it reports 67.8% mean success, 18.9 percentage points above MA-VLA trained on the same demonstrations and cameras. In simulation, it reports 90.3% mean success across eight tasks with teams of up to four arms. Ablations indicate both cross-agent attention and team-stage future-video supervision matter: removing attention reduced four-arm relay success from 95.0% to 39.0%, while removing that video supervision reduced Place Food success from 78.0% to 47.0%.

### Source excerpt

arXiv:2611.25105v1 Announce Type: cross Abstract: Every move in a robot team changes the world in which its teammates must act. Coordinated manipulation therefore calls for predicting the team's actions together with the physical future they jointly create. We introduce MAWAM, a multi-agent world-action model that jointly generates each robot's future views and executable action chunks in a single sample. Built from a pretrained single-robot world-action backbone, MAWAM jointly denoises all robots' action and future-view tokens through cross-agent attention, while preserving a separate control interface for each robot. MAWAM achieves the highest success on all three real-robot tasks (laundry-bag collection, bath-towel folding, and umbrella handover), with 67.8% mean success, 18.9 percentage points above MA-VLA trained on the same demonstrations and cameras. In simulation, MAWAM reaches 90.3% mean success on eight tasks with teams of up to four arms. Ablations show that cross-agent attention and team-stage future-video supervision both improve coordinated control: without attention, a four-arm relay falls from 95.0% to 39.0%, and without this supervision, Place Food falls from 78.0% to 47.0%.

## MADJEPA: Enhanced Multi-Stage Joint-Embedding Predictive Architecture for Medical Anomaly Detection

DevFeed: [MADJEPA: Enhanced Multi-Stage Joint-Embedding Predictive Architecture for Medical Anomaly Detection](<https://devfeed.tech/articles/madjepa-enhanced-multi-stage-joint-embedding-predictive-architecture-for-medical-anomaly-detection-145021.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25194>)

Author: Ioana-Sabina Stoian, Florinel-Alin Croitoru, Vlad Hondru, Andreea Iuliana Ionescu, Andrei Paun, Radu Tudor Ionescu

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Medical imaging](<https://devfeed.tech/topics/medical-imaging.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [Predictive AI](<https://devfeed.tech/topics/predictive-ai.md>)

Tags: [ablation](<https://devfeed.tech/tags/ablation.md>), [anomaly-detection](<https://devfeed.tech/tags/anomaly-detection.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [eess-iv](<https://devfeed.tech/tags/eess-iv.md>), [medical](<https://devfeed.tech/tags/medical.md>), [self-supervised-learning](<https://devfeed.tech/tags/self-supervised-learning.md>), [training](<https://devfeed.tech/tags/training.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

MADJEPA uses a three-stage training pipeline to detect anomalies from healthy medical images without labeled abnormal data. First, I-JEPA learns representations of normal training images; target encoder distillation is intended to limit over-generalization to abnormal data; then pseudo-anomaly classification calibrates predictions by learning a tight boundary around the normal-data manifold. Across six datasets spanning MRI, OCT, X-ray and histopathology, the authors report competitive results and first place on three datasets. Ablations indicate that I-JEPA alone was insufficient for competitive performance. The abstract does not provide dataset-specific scores or further experimental details.

### Source excerpt

arXiv:2611.25194v1 Announce Type: cross Abstract: Anomaly detection in medical images is a complex task, requiring the development of a robust model using only healthy patient data. Without access to labeled data, modern approaches typically employ self-supervised learning to obtain a model of normality. Since world models promise to build strong internal representations of their environment, we consider them as suitable candidates for the medical anomaly detection environment, namely medical investigations of healthy patients. To this end, we propose MADJEPA, a medical anomaly detector based on an enhanced multi-stage joint-embedding predictive architecture (JEPA). We design a multi-stage training pipeline to produce a robust detector: (1) image-based JEPA (I-JEPA) training aims to learn a strong representation of normal training data; (2) target encoder distillation aims to modulate over-generalization of I-JEPA representations to abnormal data; (3) pseudo-anomaly classification aims to calibrate anomaly predictions by learning a tight boundary around the normal data manifold. To demonstrate our design, we carry out comprehensive experiments on six datasets covering various medical image domains (e.g.~MRI, OCT, X-ray, histopathology), comparing our MADJEPA with numerous state-of-the-art methods. The results show that our carefully designed world model, MADJEPA, attains competitive results, ranking first on three datasets. Our ablation studies further confirm the utility of each training stage, indicating that I-JEPA alone is not sufficient to obtain competitive results in medical anomaly detection. We publicly release our code at https://github.com/Sabina085/madjepa.

## Bradley-Terry Training Can Distort Outcome Probabilities of Initially Accurate Language World Models

DevFeed: [Bradley-Terry Training Can Distort Outcome Probabilities of Initially Accurate Language World Models](<https://devfeed.tech/articles/bradley-terry-training-can-distort-outcome-probabilities-of-initially-accurate-language-world-models-145074.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25594>)

Author: Xinze Shi, Litian Zhang, Zhenjie Su, Zelong Yu, Yuting Zhang, Qihua Feng, Zifeng Kang, Chaozhuo Li, Zhengxuan Hu

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Calibrated uncertainty](<https://devfeed.tech/topics/calibrated-uncertainty.md>)

Tags: [bradley-terry](<https://devfeed.tech/tags/bradley-terry.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distributions](<https://devfeed.tech/tags/distributions.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [probability](<https://devfeed.tech/tags/probability.md>), [robot](<https://devfeed.tech/tags/robot.md>), [robust](<https://devfeed.tech/tags/robust.md>), [shift](<https://devfeed.tech/tags/shift.md>), [training](<https://devfeed.tech/tags/training.md>), [utility](<https://devfeed.tech/tags/utility.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

In a simulated robot push, fitting a language world model with a Bradley-Terry objective distorted its initially accurate outcome probabilities: 33-43 percentage points shifted to outcomes absent from the enumerated branches, despite 2.4 pp hazard error on unseen layouts before feedback. The paper attributes the shift to the objective: its precomputed optimum predicts the trained result, while a likelihood that accounts for the judge's rule avoids the shift using the same comparisons. The effect recurs on unseen conditions and under varied candidate laws and denser enumeration; about 29 pp remain after excluding physically impossible events. Decision costs depend on those events: with them, inherited thresholds cause false rejects and utility losses of about 3 pp per decision on trained conditions and 1-2 pp on unseen ones. Without them, average utility loss is below 1 pp over six seeds, though small action-choice regret remains.

### Source excerpt

arXiv:2611.25594v1 Announce Type: new Abstract: Comparative feedback, in which a judge says which of two candidate outcomes is closer to what happened, is a practical way to train language world models that assign probabilities to an action's outcomes. We show that, fitted with a Bradley-Terry objective, it need not teach the true outcome probabilities and can distort accurate ones. On a simulated robot push with enumerated outcome distributions, this objective takes a model that is accurate on unseen layouts (hazard error 2.4 percentage points, pp) and moves 33 to 43 pp of its probability onto outcomes that none of a condition's enumerated branches produces. The cause is the objective, not the comparisons: its optimum, computed before training, predicts where the model ends up, and a likelihood that knows the judge's rule learns from the same comparisons without the shift. The shift is robust: it recurs on unseen conditions, under five candidate laws and a three-fold denser enumeration, and about 29 pp remain without the physically impossible events. Its cost is conditional: with those events it inflates hazard probabilities, and a threshold inherited from the pre-feedback model costs about 3 pp of utility per decision on trained and 1 to 2 pp on unseen conditions, through false rejects; without them the utility loss averages below 1 pp over six seeds, though a small regret in action choice remains. Audits of readout, response and decision trace the effect from probabilities to actions.

## Homeostatic Continual Learning

DevFeed: [Homeostatic Continual Learning](<https://devfeed.tech/articles/homeostatic-continual-learning-145370.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.13771>)

Author: Yue Jin

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [outlier-detection](<https://devfeed.tech/topics/outlier-detection.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [animals](<https://devfeed.tech/tags/animals.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [concepts](<https://devfeed.tech/tags/concepts.md>), [concurrently](<https://devfeed.tech/tags/concurrently.md>), [context](<https://devfeed.tech/tags/context.md>), [continual-learning](<https://devfeed.tech/tags/continual-learning.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [data](<https://devfeed.tech/tags/data.md>), [environment](<https://devfeed.tech/tags/environment.md>), [features](<https://devfeed.tech/tags/features.md>), [impact](<https://devfeed.tech/tags/impact.md>), [learning](<https://devfeed.tech/tags/learning.md>), [model](<https://devfeed.tech/tags/model.md>), [nodes](<https://devfeed.tech/tags/nodes.md>), [outliers](<https://devfeed.tech/tags/outliers.md>), [policy](<https://devfeed.tech/tags/policy.md>), [transformer](<https://devfeed.tech/tags/transformer.md>), [works](<https://devfeed.tech/tags/works.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

### AI overview

Homeostatic Continual Learning proposes incremental learning for an autonomous agent with multiple intents in changing contexts. When a context change produces an unexpected intent outcome, the agent looks for outliers in its environment data, then incrementally learns relevant features and updates its model and policy. The proposal aims to extend performance across contexts without catastrophic forgetting. For agents pursuing many intents, it may also support a transformer-like world model that factors objects into features, groups them into comparable concept instances, and maps concepts to intents; the author describes this model as semantic and explainable. The abstract presents a proposal and identifies practical work still needed; it does not provide experimental results.

### Source excerpt

arXiv:2609.13771v2 Announce Type: replace Abstract: I formulate a Continual Learning problem for an autonomous agent that imitates the one animals face. The agent has intents to satisfy. It experiences contexts where different combinations of features are active and impact intent outcomes. I propose a method named ``Homeostatic Continual Learning'' that imitates the one animals use. The agent learns incrementally environment data and features relevant for their intents from the impact of contexts, gradually completes its model and policy and performs well in more and more contexts. The core of the method is to find outliers in the environment data when the agent experiences an outlier in its intent outcome after the context changes. This method enables the agent to learn continuously in changing contexts without catastrophic forgetting. When pursuing many intents concurrently, the agent may build a transformer-like world model with the method where it factorizes the objects in the world into features, abstract objects into comparable instances of concepts and map concepts to intents through features. The world model is composed of semantic nodes and explainable by nature. I discuss the works needed to render the method practical, the connections to many fields in Artificial Intelligence and the broader implications of the method.

## Business World Model

DevFeed: [Business World Model](<https://devfeed.tech/articles/business-world-model-145348.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.10044>)

Author: Cecil Pang, Hiroki Sayama

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Simulation](<https://devfeed.tech/topics/simulation.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [acting](<https://devfeed.tech/tags/acting.md>), [actions](<https://devfeed.tech/tags/actions.md>), [agents](<https://devfeed.tech/tags/agents.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [business](<https://devfeed.tech/tags/business.md>), [case](<https://devfeed.tech/tags/case.md>), [case-study](<https://devfeed.tech/tags/case-study.md>), [churn](<https://devfeed.tech/tags/churn.md>), [competition](<https://devfeed.tech/tags/competition.md>), [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [constraint](<https://devfeed.tech/tags/constraint.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [model](<https://devfeed.tech/tags/model.md>), [world](<https://devfeed.tech/tags/world.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

### AI overview

The paper proposes a Business World Model for planning in business settings where outcomes depend on context such as customer behavior, pricing, competition, and regulation. It links states, dynamics, and feasible actions to business entities and their relationships so agents can simulate action sequences and assess trade-offs under uncertainty and constraints. In the reference design, LLM agents interpret natural-language goals and explain results; a deterministic, metadata-driven engine combining machine-learning models, business rules, and action spaces performs calculations. A digital-subscription churn case study reports recommendations that meet a churn-reduction target while respecting a revenue constraint. The supplied text is an abstract and does not provide implementation details or case-study measurements.

### Source excerpt

arXiv:2606.10044v4 Announce Type: replace Abstract: World model has emerged as a powerful paradigm in artificial intelligence, enabling agents to represent their environments, predict future states, and evaluate possible actions before acting. However, existing world models have largely been developed for visual or physical domains such as computer vision, robotics, gaming, and autonomous driving, which are governed by relatively stable dynamics. These formulations are not directly applicable to business practice, where the relevant environment is semantic, organizational, and market driven. Outcomes depend on context-sensitive factors such as customer behavior, pricing, competition, and regulation. This paper introduces the concept and architecture of a Business World Model, a world model specialized for business and organizational environments. We propose a business-semantics-centric formulation in which states, dynamics, and feasible actions are linked to business entities, their attributes, and their relationships. Within this framework, intelligent agents can simulate alternative action sequences, estimate their effects on future business outcomes, and evaluate trade-offs under uncertainty and business constraints. We present a reference design and open-source implementation in which LLM agents interpret natural-language business goals and explain results, while a deterministic, metadata-driven simulation engine integrating machine learning models, business rules, and action spaces performs all calculations. A churn case study of a digital subscription business demonstrates that the system recommends options meeting the churn-reduction target within the revenue constraint. This work establishes a conceptual foundation for autonomous business systems capable of moving from instruction-based execution toward goal-driven planning, optimization, and execution.

## WorldRuler: Interactive World Modeling under Configurable Rules

DevFeed: [WorldRuler: Interactive World Modeling under Configurable Rules](<https://devfeed.tech/articles/worldruler-interactive-world-modeling-under-configurable-rules-142841.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24869>)

Author: Mingliang Zhai, Chuanhao Li, Jiaming Tan, Zhen Li, Yuwei Wu, Yunde Jia, Zian Meng, Yuanyang Yin, Kaipeng Zhang

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [AI for interactive applications](<https://devfeed.tech/topics/ai-for-interactive-applications.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [rule-engine](<https://devfeed.tech/topics/rule-engine.md>)

Tags: [3d](<https://devfeed.tech/tags/3d.md>), [actions](<https://devfeed.tech/tags/actions.md>), [annotation](<https://devfeed.tech/tags/annotation.md>), [annotations](<https://devfeed.tech/tags/annotations.md>), [appearance](<https://devfeed.tech/tags/appearance.md>), [automated](<https://devfeed.tech/tags/automated.md>), [behaviors](<https://devfeed.tech/tags/behaviors.md>), [challenges](<https://devfeed.tech/tags/challenges.md>), [change](<https://devfeed.tech/tags/change.md>), [changes](<https://devfeed.tech/tags/changes.md>), [collection](<https://devfeed.tech/tags/collection.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [content](<https://devfeed.tech/tags/content.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [customizable](<https://devfeed.tech/tags/customizable.md>), [data](<https://devfeed.tech/tags/data.md>), [interactive](<https://devfeed.tech/tags/interactive.md>), [modeling](<https://devfeed.tech/tags/modeling.md>), [rules](<https://devfeed.tech/tags/rules.md>), [world](<https://devfeed.tech/tags/world.md>)

### AI overview

WorldRuler aims to make generative interactive worlds follow configurable rules while preserving them across save and reload. It represents rules, dependencies, and object states in an executable 3D rule graph, then links rule execution, video generation, and perceptual feedback in a loop. Dynamic object binding grounds rule-defined entities in visual instances and uses stored states to constrain generation from their first appearance, including when an affected object was initially unseen. Object attribute state replacement and counterfactual reconstruction training target designated changes while limiting effects on unrelated content. The authors also introduce RuleVerse, built using GTA-V and Habitat, with detailed state and event annotations, including rules that take effect before affected objects become visible. Their abstract reports improved rule-state persistence and realization of specified outcomes.

### Source excerpt

arXiv:2611.24869v1 Announce Type: new Abstract: Rules govern how player actions change interactive worlds. Different rules create distinct gameplay experiences within similar scenes and action spaces. For customizable gameplay, generative world models must follow user-defined rules beyond real-world behaviors. However, realizing user-defined rules presents two challenges: Rule-State Persistence and Rule-to-Visual Realization. The former requires preserving rule definitions, dependencies, and world states across saving and loading, so that restored worlds follow the established rules. The latter requires accurately realizing rule-induced changes on designated objects and attributes, particularly when affected objects are unseen. In this paper, we propose WorldRuler, which organizes rule definitions, dependencies, and object states in an executable 3D rule graph and connects rule execution, video generation, and perceptual feedback in a closed loop. The graph supports joint saving and restoration of rules and world states, while dynamic object binding grounds rule-specified entities in visual instances and uses stored states to constrain generation from their first appearance. We further introduce object attribute state replacement and counterfactual reconstruction training, enabling accurate attribute control and compatible appearance and motion changes while limiting interference with unrelated content. We develop an automated data collection and annotation pipeline using GTA-V and Habitat to construct RuleVerse, a dataset of rule-driven interactions with detailed state and event annotations, including cases where rules take effect before affected objects become visible. Extensive experiments demonstrate the effectiveness of WorldRuler in preserving rule-state consistency across saving and loading and accurately realizing rule outcomes on designated objects and attributes.

## Autonomous Video Generation for Self-Evolving World Models

DevFeed: [Autonomous Video Generation for Self-Evolving World Models](<https://devfeed.tech/articles/autonomous-video-generation-for-self-evolving-world-models-143281.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.24152>)

Author: Xin Wang, Wenxuan Liu, Tongtong Feng, Wenwu Zhu

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [arxiv](<https://devfeed.tech/tags/arxiv.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [bind](<https://devfeed.tech/tags/bind.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [decision-making](<https://devfeed.tech/tags/decision-making.md>), [delay](<https://devfeed.tech/tags/delay.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [drones](<https://devfeed.tech/tags/drones.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [environments](<https://devfeed.tech/tags/environments.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generation](<https://devfeed.tech/tags/generation.md>), [models](<https://devfeed.tech/tags/models.md>), [video](<https://devfeed.tech/tags/video.md>), [video-generation](<https://devfeed.tech/tags/video-generation.md>), [world](<https://devfeed.tech/tags/world.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

The article argues that world models should be judged by whether their generated futures support valid action for a particular embodied agent, not by visual plausibility alone. It proposes autonomous video generation evaluated through counterfactual controllability: generating intervention-conditioned futures, binding them to embodiment constraints, checking them under distribution shifts, and distilling surviving branches into compact decision variables. Its closed loop has four stages--Generation, Binding, Verification, and Distillation--with corresponding metrics of novelty, consistency, out-of-distribution performance, and efficiency. Drones and manipulators are proposed as testbeds, where factors such as wind, sensing limits, actuation delay, contact dynamics, and recovery constraints can be perturbed and verified. The supplied text presents this as a framework and examples, not evidence of demonstrated performance.

### Source excerpt

arXiv:2606.24152v3 Announce Type: replace-cross Abstract: Large-scale video generation models are increasingly described as world models because they can learn rich spatiotemporal regularities from visual data. We argue that an ideal world model should benefit in a self-evolving generative character. As such, a self-evolving world model can be realized via a generative model conditioned by various constraints, whose future predictions are refined through feedback. However, traditional visually plausible predictions alone are not enough to establish whether an imagined future is physically actionable for a particular embodied agent, failing to provide informative feedback from environments for self-evolving improvement. To realize self-evolving world models, this article proposes the concept of autonomous video generation, which is evaluated through counterfactual controllability, i.e., the ability to i) generate intervention-conditioned futures, ii) bind these future frames to embodiment constraints, iii) verify them under distribution shifts, and iv) distil surviving branches into compact variables for decision-making. We formalize a four-stage closed-loop optimization of Generation, Binding, Verification and Distillation, together with four corresponding evaluation metrics: novelty, consistency, out-of-distribution (OOD) and efficiency. We further discuss two examples, i.e., drones and manipulators, as early embodied testbeds where wind, sensing limits, actuation delay, contact dynamics and recovery constraints can be systematically perturbed and verified. The central claim is that the framework of autonomous video generation for self-evolving world models should not be judged by video fidelity alone, but by whether the generated frames improve valid action under counterfactual interventions and various embodiment constraints.

## Measuring the Conservatism of Latent Safety Filters for World Models

DevFeed: [Measuring the Conservatism of Latent Safety Filters for World Models](<https://devfeed.tech/articles/measuring-the-conservatism-of-latent-safety-filters-for-world-models-143434.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23685>)

Author: Johannes Mootz, Mani Amani, Reza Akhavian

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [responsible-ai](<https://devfeed.tech/topics/responsible-ai.md>)

Tags: [classifier](<https://devfeed.tech/tags/classifier.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [cs-sy](<https://devfeed.tech/tags/cs-sy.md>), [eess-sy](<https://devfeed.tech/tags/eess-sy.md>), [encoder](<https://devfeed.tech/tags/encoder.md>), [navigation](<https://devfeed.tech/tags/navigation.md>), [overlap](<https://devfeed.tech/tags/overlap.md>), [planner](<https://devfeed.tech/tags/planner.md>), [probability](<https://devfeed.tech/tags/probability.md>), [safety](<https://devfeed.tech/tags/safety.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

Conformal calibration limits the probability that a latent safety filter accepts an unsafe action, but does not prevent it from rejecting safe actions. The paper measures this conservatism with "overlap": the share of safe transitions in cells where safe and unsafe transitions coexist. It proves this is the minimum share of safe transitions a one-decision-per-cell filter must refuse to accept no unsafe transition in the data. The authors compare safety-label training on a classifier atop the world model, the world model itself, or a separate encoder. In a visual navigation task, the separate encoder has the lowest overlap; at the same conformal level, its filter rejects the fewest safe actions and enables the planner to reach the most goals among the three designs.

### Source excerpt

arXiv:2611.23685v1 Announce Type: cross Abstract: A latent safety filter decides from the latent of a world model whether a planned action is safe. Conformal calibration sets the threshold of the filter so that it accepts an unsafe action only with a small probability. However, the filter can still be conservative and refuse many safe actions, depending on the latent. To measure the conservatism that a latent forces on the filter, we group similar latents and actions into cells. A filter that gives one decision per cell can refuse the unsafe transitions of a cell only by refusing the safe ones in it as well. We call the share of safe transitions that lie in such cells the overlap. We prove that the overlap is the minimum share of safe transitions that such a filter has to refuse in order to accept no unsafe transition in the data. Both the overlap and the conformal calibration can be applied to any latent. We use this to compare three existing designs, in which the safety labels train a classifier on top of the world model, the world model itself, or a separate encoder beside it. In a visual navigation task, the separate encoder has the lowest overlap, and at the same conformal level its filter refuses the fewest safe actions and lets the planner reach the most goals of the three.

## Think Offline, Act in Real Time: Self-Improving Visual Game Agents through Code

DevFeed: [Think Offline, Act in Real Time: Self-Improving Visual Game Agents through Code](<https://devfeed.tech/articles/think-offline-act-in-real-time-self-improving-visual-game-agents-through-code-142811.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24590>)

Author: Zipeng Qiu, Shusheng Xu, Jiale Li, Yukun Zhou, Chenyue Li, Wenshuang Li, Tongkai Yang, Binhang Yuan

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>), [real-time](<https://devfeed.tech/topics/real-time.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [code](<https://devfeed.tech/tags/code.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [game](<https://devfeed.tech/tags/game.md>), [improvement](<https://devfeed.tech/tags/improvement.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [play](<https://devfeed.tech/tags/play.md>), [policy](<https://devfeed.tech/tags/policy.md>), [program](<https://devfeed.tech/tags/program.md>), [progress](<https://devfeed.tech/tags/progress.md>), [real-time](<https://devfeed.tech/tags/real-time.md>), [tasks](<https://devfeed.tech/tags/tasks.md>)

### AI overview

The study compares two ways for agents to improve at unfamiliar real-time visual games from a few attempts. In "model as policy," a trained 2B controller selects actions during play, with committed action chunks continuing while it reasons. In "code as policy," a model writes, tests, and revises a program between attempts, and that program acts during play. Across 170 GameWorld tasks, limited to at most eight one-minute executions per run, code as policy achieved 47.80% mean progress versus 30.90% for the controller and led in all five game categories. Improving the controller's weights on one game brought only modest gains. The authors conclude that learning offline into a program can work better in this setting.

### Source excerpt

arXiv:2611.24590v1 Announce Type: new Abstract: Pre-trained models can read screens and reason about games, yet on an unfamiliar game, an agent still has to learn to play. We ask how an agent can self-improve on such a game from a few of its own attempts, in real-time visual games where the agent sees only the screen and the game keeps running while the agent thinks. We study two kinds of policy that can carry this improvement. In model as policy, a model selects actions during play; we adapt real-time chunking so that committed actions keep executing while the model reasons, and train a 2B controller for this interface. In code as policy, a model writes, tests, and revises a program between attempts, and the program acts during play. Across 170 GameWorld tasks, with at most eight one-minute executions per run, code as policy reaches 47.80% mean progress, against 30.90% for our trained controller, and leads in all five game categories. Improving the controller's weights on a single game, by contrast, yields only modest gains. These results favor letting a model think offline and storing what it learns in a program that acts in real time. Project repository: https://github.com/TabibitoQZP/SIGame

## Worth Another Look:Constrained World-Model Planning for Visual Inspection

DevFeed: [Worth Another Look:Constrained World-Model Planning for Visual Inspection](<https://devfeed.tech/articles/worth-another-look-constrained-world-model-planning-for-visual-inspection-142764.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24187>)

Author: Ziyue Kang, Zixi Qi, Yuchen Dai, Nana Lv, Xirong Cheng, Liang Liu, Ying Li, Haoyan Wei, Haiyang Yao

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [beam-search](<https://devfeed.tech/tags/beam-search.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [budget](<https://devfeed.tech/tags/budget.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-mm](<https://devfeed.tech/tags/cs-mm.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [environmental-monitoring](<https://devfeed.tech/tags/environmental-monitoring.md>), [fallback](<https://devfeed.tech/tags/fallback.md>), [inspection](<https://devfeed.tech/tags/inspection.md>), [lightweight](<https://devfeed.tech/tags/lightweight.md>), [lookahead](<https://devfeed.tech/tags/lookahead.md>), [observability](<https://devfeed.tech/tags/observability.md>), [offline](<https://devfeed.tech/tags/offline.md>), [online](<https://devfeed.tech/tags/online.md>), [optical](<https://devfeed.tech/tags/optical.md>), [pairing](<https://devfeed.tech/tags/pairing.md>), [performance](<https://devfeed.tech/tags/performance.md>), [planning](<https://devfeed.tech/tags/planning.md>), [policies](<https://devfeed.tech/tags/policies.md>), [pricing](<https://devfeed.tech/tags/pricing.md>), [visual-inspection](<https://devfeed.tech/tags/visual-inspection.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

### AI overview

The paper proposes budget-aware visual inspection that combines a compact recurrent world model with exact symbolic feasibility and tariff rules. Receding-horizon beam search replans schedules for class queries, crops, image enhancements, and stopping; slack-aware shadow pricing and a legal fallback support larger budgets. Across eight terrestrial, optical, and sonar benchmarks, multi-step lookahead beats static schedules and search-free offline RL under tight budgets, with gains of up to 28.4 AP points over static priority on RUOD and 15.1 on ARIS. Dynamic pricing improves large-budget results across all eight domains, at an 11 ms planning cost. The approach uses frozen detectors without retraining.

### Source excerpt

arXiv:2611.24187v1 Announce Type: cross Abstract: Visual inspection in field robotics and environmental monitoring must localize multiple targets under strict compute and sensing caps. Exhaustive sweeps waste calls on absent classes, while myopic and search-free policies stumble over heterogeneous tariff cliffs or discard viewport edits whose value emerges only after re-querying. We formulate inspection over frozen detectors as neuro-symbolic observation planning under partial observability, decoupling a compact recurrent world model (WM) from exact symbolic feasibility and tariff rules. Receding-horizon beam search with online replanning jointly schedules class queries, crops, enhancements, and termination (stop), paired with slack-aware shadow pricing and legal fallback at larger budgets. Across eight terrestrial, optical, and sonar benchmarks, multi-step lookahead outperforms static schedules and search-free offline RL under tight budgets -- gaining up to 28.4 AP percentage points over static priority on RUOD and 15.1 on ARIS -- while dynamic pricing lifts large-budget performance on all eight domains at 11 ms planning cost. Ultimately, pairing lightweight learned dynamics with exact symbolic constraints turns frozen detectors into adaptive, budget-aware inspection agents without retraining.

## Feasibility Ceilings Before Learning: Sizing and Timing Bounds for Drone-Mesh Interception

DevFeed: [Feasibility Ceilings Before Learning: Sizing and Timing Bounds for Drone-Mesh Interception](<https://devfeed.tech/articles/feasibility-ceilings-before-learning-sizing-and-timing-bounds-for-drone-mesh-interception-143333.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.20823>)

Author: Dmytro Babarytskyi

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Amazon robotics research](<https://devfeed.tech/topics/amazon-robotics-research.md>), [World models](<https://devfeed.tech/topics/world-models.md>)

Tags: [bias](<https://devfeed.tech/tags/bias.md>), [control](<https://devfeed.tech/tags/control.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [drone](<https://devfeed.tech/tags/drone.md>), [interception](<https://devfeed.tech/tags/interception.md>), [mesh](<https://devfeed.tech/tags/mesh.md>), [monte-carlo](<https://devfeed.tech/tags/monte-carlo.md>), [multi-agent](<https://devfeed.tech/tags/multi-agent.md>), [packing](<https://devfeed.tech/tags/packing.md>), [ppo](<https://devfeed.tech/tags/ppo.md>), [probability](<https://devfeed.tech/tags/probability.md>), [robotics](<https://devfeed.tech/tags/robotics.md>), [timing](<https://devfeed.tech/tags/timing.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [variance](<https://devfeed.tech/tags/variance.md>)

### AI overview

The abstract derives pre-policy bounds for drone-mesh interception and validates them by Monte Carlo. For noisy linear-track extrapolation, least-squares variance propagation plus a delta-method arrival-time term predicts aim-point uncertainty within 1.8% of simulation. A closed-form packed-mesh coverage estimate is accurate to 0.007 absolute. With agents separated by at least 2r, coverage is capped at 90.69% in the plane and 74.05% in a volume; allowing overlap removes this separation-driven ceiling above crossover p=0.80. Among five arrangements, hexagonal packing is best, while centroidal-Voronoi coverage control reaches half its coverage at small budgets. Gaussian uncertainty is conservative for heavy tails and anisotropy but sensitive to bias: a one-sigma aim-point offset costs 14 coverage points. The repositioning commit-time window may be empty. Relative to the applicable ceiling, analytic placement reaches 85% of permitted coverage; centralized PPO averages 15% across three trainer seeds, with paired-bootstrap comparisons against a null policy inconclusive at every seed.

### Source excerpt

arXiv:2610.20823v2 Announce Type: replace-cross Abstract: A mesh of interceptors defending a fixed site must spread over the predicted arrival point of an approaching object. This is a recurring task in multi-agent robotics: N agents, each covering a disc of radius r, spread over a probability density so that a point drawn from it falls within some agent's disc. We give closed forms for the quantities that bound performance before any policy is chosen, and validate each by Monte Carlo. The uncertainty of a point predicted by extrapolating a noisy linear track follows from least-squares variance propagation plus a delta-method term for the uncertain arrival time, agreeing with simulation to within 1.8%. The coverage probability of a packed mesh admits a closed form accurate to 0.007 absolute, with a volumetric analogue. A naive area ratio misses a ceiling: while agents stay at least 2r apart, no mesh exceeds 90.69% coverage in the plane or 74.05% in a volume. This ceiling belongs to the separation constraint, not to the problem - permitting overlap substitutes Kershner's thinnest-covering density and removes it, the better regime above a crossover at p=0.80. Hexagonal packing is optimal among five arrangements; the centroidal-Voronoi configuration coverage control supplies reaches only half its coverage at small budgets. The Gaussian assumption is conservative under heavy tails and anisotropy but fragile under bias: a one-sigma aim-point offset costs 14 points of coverage. A repositioning team must also choose a commit time, and the admissible window can be empty. Against the applicable ceiling, analytic placement attains 85% of what the configuration permits while centralised PPO averages 15% over three trainer seeds, with a paired bootstrap against a null policy containing zero at every seed. We argue such ceilings should be computed before any learning experiment and quoted alongside every result.

## Counterfactual World Models: JEPAs That Reason About "What If"

DevFeed: [Counterfactual World Models: JEPAs That Reason About "What If"](<https://devfeed.tech/articles/counterfactual-world-models-jepas-that-reason-about-what-if-142830.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24778>)

Author: Randall Balestriero, Yann LeCun, Kerda Varaku

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [clip](<https://devfeed.tech/tags/clip.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [minimax](<https://devfeed.tech/tags/minimax.md>), [pixels](<https://devfeed.tech/tags/pixels.md>), [planner](<https://devfeed.tech/tags/planner.md>), [planning](<https://devfeed.tech/tags/planning.md>), [simulator](<https://devfeed.tech/tags/simulator.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>), [vehicle](<https://devfeed.tech/tags/vehicle.md>), [world-model](<https://devfeed.tech/tags/world-model.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

The proposed counterfactual world model estimates what an alternative action would have done under the same hidden disturbance, without counterfactual training data, added parameters, or retraining. It combines the observed outcome with the difference between the existing model's predictions for the alternative and factual actions. The result relies on actions and disturbances acting as separate forces in smooth dynamics; the authors identify discontinuities, coordinate folds, and disturbances that scale commands as failure cases. They prove an O(h²) error bound relative to disturbance variance and a matching minimax lower bound for unpaired factual data. In reported image-based JEPA experiments, counterfactual R² rose from 38-46% to 86-96% on four rendered simulators; an underwater vehicle reached 90-94% with sixteen times more factual data. Hindsight planning reached the goal in 35-69% of events, versus 14-23% for an oracle planner that knew the simulator but not the outcome.

### Source excerpt

arXiv:2611.24778v1 Announce Type: cross Abstract: World models operate in a world they cannot fully observe. When a hidden disturbance, such as a current pushing a boat, changes the outcome of an action, a counterfactual question follows: what would have happened under a different action, subject to the same disturbance? Its answer separates bad decisions from bad luck. We prove that world models trained on factual transitions cannot answer it on their own, whatever their data or capacity. For many physical systems, the observed outcome and a simple property of the dynamics suffice: natura non facit saltus, nature makes no leaps. When actions and disturbances act as separate forces, their interaction is second-order in time, so the counterfactual outcome is, to first order, the observed outcome plus the difference between the world model's predictions under the alternative and factual actions. This yields a counterfactual world model with no counterfactual data, extra parameters, or retraining. For transitions of duration $h$, we prove an $O(h^2)$ bound on the counterfactual error relative to the disturbance variance, establish a matching minimax lower bound from unpaired factual data, and characterize failures at discontinuities, coordinate folds, and disturbances that scale the command. We validate these predictions from pixels with JEPA world models on frozen DINOv2 and masked autoencoder (MAE) encoders. On four rendered simulators, our approach improves counterfactual $R^2$ from $38$--$46\%$ to $86$--$96\%$ with the same world models; on a fifth, an underwater vehicle, it reaches $90$--$94\%$ with sixteen times more factual data. Four further frozen encoders, DINOv3, SigLIP~2, CLIP and EVA-02, give $81$--$95\%$ on the same simulators. In hindsight planning, it finds the action that would have reached the goal in $35$--$69\%$ of events, against $14$--$23\%$ for an oracle planner that knows the simulator but not the outcome.

## Considering Context: When World Models Need Context Encoders

DevFeed: [Considering Context: When World Models Need Context Encoders](<https://devfeed.tech/articles/considering-context-when-world-models-need-context-encoders-143326.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.06651>)

Author: Oleg Smirnov, Sofiane Ennadir, John Pertoft, Bjartur Hjaltason, Sara Karimi

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [context](<https://devfeed.tech/topics/context.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [environment](<https://devfeed.tech/tags/environment.md>), [environments](<https://devfeed.tech/tags/environments.md>), [framework](<https://devfeed.tech/tags/framework.md>), [headroom](<https://devfeed.tech/tags/headroom.md>), [history](<https://devfeed.tech/tags/history.md>), [information](<https://devfeed.tech/tags/information.md>), [mechanism](<https://devfeed.tech/tags/mechanism.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [residual](<https://devfeed.tech/tags/residual.md>), [risk](<https://devfeed.tech/tags/risk.md>), [state](<https://devfeed.tech/tags/state.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

The study proposes predictive sufficiency to measure how much access to an environment's latent context improves next-step prediction under an agent's own visitation distribution. It separates information recoverable from experience from the residual requiring true context, distinguishing both from a finite model's prediction deficit. Across environments with varying identification difficulty, predictive headroom depends on the prior and agent behavior rather than simply on the MDP class: in some settings behavior implicitly identifies context, leaving no measurable one-step predictive benefit from adding it; in others, learned state reveals context only partially, and true context further reduces risk. The authors offer a criterion, estimated from an ordinarily trained agent, for matching contextual mechanisms to available predictive information.

### Source excerpt

arXiv:2610.06651v2 Announce Type: replace Abstract: Methods for generalization in model-based reinforcement learning typically assume that an agent cannot recover the latent context governing the environment dynamics from its own experience, and therefore supplies it externally. We formalize and test this assumption with \emph{predictive sufficiency}, which quantifies what access to the context adds to next-step prediction under the visitation distribution an agent induces, and separates that quantity into a history-recoverable part and a residual requiring the true context, both distinct from a finite model's deficit. We classify context-aware algorithms by the predictive risk their conditioning set can target and demonstrate across environments of increasing identification difficulty that the headroom does not follow the MDP class. The same task under different priors leaves predictive headroom in one setting and nothing distinguishable from zero in another, where the agent's behavior implicitly identifies the context and any benefit of such a mechanism cannot be attributed to missing one-step predictive information. Where headroom persists, the learned state exposes it only partially, and adding the true context still lowers the risk. Our contribution is an analytical framework, instantiated by a practical criterion for matching contextual mechanisms to the predictive information available to them, estimated from the ordinary trained agent.

## Does the Action Matter? Probing Action-Conditioned Semantic Forecasting as a World Model for Autonomous Driving

DevFeed: [Does the Action Matter? Probing Action-Conditioned Semantic Forecasting as a World Model for Autonomous Driving](<https://devfeed.tech/articles/does-the-action-matter-probing-action-conditioned-semantic-forecasting-as-a-world-model-for-autonomous-driving-142837.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24840>)

Author: Binay Kumar Singh, Niels Da Vitoria Lobo

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [autonomous vehicles](<https://devfeed.tech/topics/autonomous-vehicles.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [driving](<https://devfeed.tech/tags/driving.md>), [forecasting](<https://devfeed.tech/tags/forecasting.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [planning](<https://devfeed.tech/tags/planning.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [predictions](<https://devfeed.tech/tags/predictions.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

### AI overview

The study tests whether autonomous-driving semantic forecasts actually depend on the ego action, using test-time action interventions scored with segmentation metrics. The executed action improves mid-term mIoU by at most 2.5 points on Cityscapes and 0.6 on BDD100K; supplying a wrong action lowers it by up to 23.2 and 7.0 points, respectively, below a model without action input. On real data, feature-wise action conditioning is 1.6 times more action-dependent than input-level conditioning at similar accuracy; in a synthetic setting where history does not determine action, their dependence is equal. The findings indicate that accuracy gains show action use but do not establish how strongly forecasts rely on actions.

### Source excerpt

arXiv:2611.24840v1 Announce Type: cross Abstract: World models are increasingly used in autonomous driving to predict how a scene evolves under the ego vehicle's actions, supporting counterfactual reasoning and model-based planning. However, existing evaluations focus on prediction fidelity and planning performance, and rarely verify that the predictions depend on the conditioning action. In this paper, we study action dependence in action-conditioned semantic forecasting, which predicts future semantic segmentation from past segmentations and the ego action. We compare a forecaster with feature-wise and input-level action conditioning against an identical model without action input, and measure action dependence with test-time action interventions scored by standard segmentation metrics. The evaluation is validated in a synthetic environment with known action-dependent dynamics and applied to Cityscapes and BDD100K, both within each dataset and across the two vehicle platforms. Experiments show that the executed action improves mid-term mIoU by at most 2.5 points on Cityscapes and 0.6 points on BDD100K, whereas a wrong action reduces it by up to 23.2 and 7.0 points, below the model without action input. On real data, feature-wise conditioning depends 1.6 times more on the action than input-level conditioning at similar accuracy, whereas both depend on it equally in the synthetic environment, where the history does not determine the action. These results show that an accuracy gain indicates that a driving world model uses its action, but not how strongly its forecasts depend on it. Code is available at https://github.com/binayksingh/does-action-matter.

[Next page](<https://devfeed.tech/topics/world-models.md?cursor=WyIyMDI2LTExLTE5VDA1OjAwOjAwKzAwOjAwIiwgIjM4YTk0OWQwLTRmYmMtNDI4MS05ZGIzLTVlM2VjOTI0NWVlZSJd>)