A new paper introduces loss-conditioned state execution, a model-agnostic method for deciding whether a learned world model should adopt its proposed feasible next state or stay at the current one. The authors argue that predictive informativeness on its own does not tell you whether switching states actually improves the model, and use a loss-based criterion instead.
A new arXiv paper argues that adding inference-time compute helps video world models only when the extra samples are actually better and can be reliably picked out. The authors separate the benefit of a larger candidate pool from the ability to select the best one, an effect they call sampling headroom versus selection gain. They propose auditing test-time scaling along this distinction to judge whether the added compute is worth its cost.
DreamQAS targets reinforcement-learning-based quantum architecture search, where a variational quantum eigensolver is run again after every added gate even though the circuit transitions and legal actions are already known. The method keeps those exact dynamics and instead learns a decision-useful world model, reducing the number of VQE evaluations needed during search. It is a research preprint posted to arXiv.
A new arXiv paper proposes jointly training a language model agent's policy alongside a world model, so the agent learns both which actions earn rewards and how those actions change the environment. The authors argue that standard reinforcement learning gives sparse guidance about environmental consequences, and that combining policy learning with world modeling addresses this gap. The work targets improved decision-making for LLM-based agents in interactive settings.
A new arXiv paper introduces Valerant, a system that automatically creates explorable game maps using action-conditioned world models. The approach builds on World Action Models, which combine predictive modeling with action generation so that anticipated future states can steer agent behavior. The work aims to address the limited exploration of general-purpose applications of such models in embodied AI.
A new arXiv paper examines world models, systems that predict what happens next from current conditions, and how they behave when generating several modalities such as visual simulations at once. The authors propose an auditing approach to detect misalignment across these outputs, arguing that a single model can encode conflicting physical accounts. The work frames such inconsistency as a safety concern for multi-modal generation.
Researchers describe an Immune World Model that aims to capture immune processes across cellular, tissue, and patient-level scales rather than treating them in isolation. The model was assembled using a governed evolutionary AI Scientist framework, according to the preprint. It is positioned as a tool for forecasting and for generating hypotheses about immune therapies.
A new arXiv preprint examines the widely held assumption that learned dynamics models which conform to underlying physics produce more accurate predictions. The authors recover exact polynomial invariants from trajectories and canonicalise them as a method for diagnosing faults in reinforcement learning simulators and world models. The work frames these invariants as a diagnostic tool for checking whether a model's learned dynamics actually respect physical structure.
A new arXiv paper examines how pretrained world models, which learn latent representations of observations and predict their evolution under actions, can be compromised by backdoor attacks when reused as general-purpose dynamics backbones for control tasks. The authors study the security risks this reuse creates for downstream control systems.
This arXiv paper introduces LPA-CWM, a learned adjudicator that weighs the outputs of counterfactual world models when extracting motion from pretrained video predictors. The authors note that predictions produced under different target-frame masks differ in reliability, making uniform weighting suboptimal. Their approach learns to score and combine those predictions rather than treating them equally.
Researchers introduce LePlanner, an iterative amortized controller designed to plan within the latent spaces of world models built on joint-embedding predictive architectures. The work targets the high computational cost of existing planning methods, which typically rely on either search-based or gradient-based optimization. The paper is posted on arXiv as a cross-listing.
A new arXiv preprint examines how automatic research agents, which use large language models to write code and iterate on experiments, can be scaled further. The authors propose incorporating world models so these agents can better predict and plan within their research environments. The work targets the long-standing goal of automating empirical research end to end.
A new arXiv preprint proposes Semigroup-JEPA, a joint-embedding predictive architecture augmented with a latent dynamics consistency objective. The approach is designed to help world models capture physical behavior so that predicted dynamics remain plausible in unseen scenarios without retraining. The paper was posted to arXiv's cs.AI and cs.LG listings.
A new arXiv paper in cs.AI argues that AI world models can capture core aspects of physical scenes, such as depth, motion, and how objects hold together and interact, without task-specific training. The authors draw an analogy to early childhood cognition, where humans build intuitive physics from relatively little experience. The work suggests developmental principles could guide the design of more sample-efficient artificial learners.
A new arXiv paper examines why LLM-based agents often stop improving when placed in unfamiliar settings, identifying a failure mode the authors call exploration collapse. Under reinforcement learning, the researchers argue, the agent's world model fails to align with the new environment's state distribution, causing exploration to break down. The work characterizes this phenomenon from a world-modeling perspective to explain when and why such collapse occurs.
A new paper examines how large language models can generalize in domains that lack abundant real, labeled examples. By expressing a domain's dynamics as code, the authors show a single template can instantiate many simulated world models whose executions yield verified training data. The goal is to manufacture generalization examples cheaply where real-world annotation is scarce.
A new paper introduces Hi-FLoop, a hierarchical architecture that uses state-feedback loops to model traffic scenes across multiple decision timescales. The method targets long-horizon, closed-loop multi-agent simulation, aiming to produce coordinated and physically plausible rollouts from road maps and past agent behavior.
A new arXiv paper introduces GameWAM, a world action model intended for agents that play video games featuring first-person viewpoints, fast visual changes, persistent world state, and varied native controls. The authors argue that current game agents translate visual and task context straight into actions without modeling world dynamics explicitly, and their approach adds that missing component. The submission is a replacement cross-list entry on arXiv cs.LG.
Researchers posted a paper describing Motus2, a general world model intended to let embodied agents perceive, predict, act, evaluate, and improve inside one system. The authors argue that prior world models typically bolt an action head onto a simulator rather than unifying those capabilities. The work focuses on dexterous manipulation as its test setting.
A new arXiv paper examines whether contact predictions from a compact, randomly initialized visuotactile world model can actually support better decisions in robotic lifting. The authors combine trajectory-level uncertainty calibration with reward alignment and force constraints to guide manipulation behavior.