World Models & Planning - Weekly Roundup

Human trajectory logs are no longer the performance ceiling for autonomous driving: closed-loop reinforcement learning paired with distilled foundation models outperforms human demonstration baselines across major open and closed-loop benchmarks.

Monolithic 3D scenes with hundreds of heavily occluded objects can be cleanly parsed into individual editable meshes without ever training on multi-object data.

Robotic world-action models do not need pretrained internet video backbones to scale: training native planning and dynamics modules from scratch across 30,000 hours unlocks strong zero-shot manipulation and drives a 17.7-point gain on an embodiment comprising under 2% of the data.

Numerical reasoning in HTN planning just got a major upgrade with a new SMT-based encoding that sets a competitive baseline for future advancements.

The framework not only guarantees the discovery of an approximate Nash equilibrium but also certifies when no exact equilibrium exists, redefining our approach to equilibrium analysis in stochastic games.

Human trajectory logs are no longer the performance ceiling for autonomous driving: closed-loop reinforcement learning paired with distilled foundation models outperforms human demonstration baselines across major open and closed-loop benchmarks.

Monolithic 3D scenes with hundreds of heavily occluded objects can be cleanly parsed into individual editable meshes without ever training on multi-object data.

Robotic world-action models do not need pretrained internet video backbones to scale: training native planning and dynamics modules from scratch across 30,000 hours unlocks strong zero-shot manipulation and drives a 17.7-point gain on an embodiment comprising under 2% of the data.

Numerical reasoning in HTN planning just got a major upgrade with a new SMT-based encoding that sets a competitive baseline for future advancements.

The framework not only guarantees the discovery of an approximate Nash equilibrium but also certifies when no exact equilibrium exists, redefining our approach to equilibrium analysis in stochastic games.

GBTs can outperform neural networks in game-playing scenarios, revealing a potential misalignment in the community's reliance on deep learning methods.

Sequential refinements using MPI can enhance offline RL performance beyond traditional behavior constraints, outperforming established baselines.

State alignment significantly boosts planning success in robotic tasks, achieving 100% success on TwoRoom and 98% on PushT, outperforming traditional models.

A single scalar from a learned operator can predict convergence in coupled dynamical systems, revealing the limits of traditional backpropagation methods.

TIGPO redefines how long-horizon LLM agents leverage historical transitions, leading to superior performance in complex environments.

Removing unnecessary ordering constraints in HTN plans can significantly streamline planning processes without sacrificing validity.

Bioinfoysis boosts bioinformatics accuracy from 27.81% to 64.13% by ensuring that every step of analysis is grounded in real-time evidence and planning.

Current world models may prioritize likelihood over safety, but the new Risk-Informed World Model shifts the focus to what truly matters for decision-making in safety-critical environments.

Proactive service agents can significantly enhance user experience by inferring needs and acting before explicit instructions are given, transforming the interaction paradigm.

A novel deadline-aware metric and a hybrid reinforcement learning architecture dramatically improve network control efficiency, enabling timely delivery of critical information in dynamic environments.

Uncovering causal relationships through symmetries could revolutionize how we approach complex machine-learning tasks beyond IID constraints.

A single character in a memory directive can shift an agent's retrieval success by over 78 points, underscoring the critical role of memory management in AI decision-making.

StyleDrive achieves a remarkable 17.08-point increase in driving score over existing methods, showcasing its superior long-horizon consistency and adaptability to diverse driving styles.

SV-WAM achieves high-performance autonomous driving planning without the computational burden of generating future videos during inference.

Overcoming the 2D-to-4D spatial bottleneck in VLMs does not require native 3D architectures; factorizing visual projections into verifiable planar, depth, and temporal RL objectives delivers immediate 4-6% benchmark gains.

A novel hierarchical framework enables autonomous vehicles to anticipate long-term traffic dynamics while making real-time decisions, achieving superior driving performance.

Agile Mode can dynamically adjust obstacle avoidance strategies, achieving impressive real-time performance in cluttered environments despite inherent limitations.

UPG reveals that nominally identical pitch sequences can have vastly different physical executions, transforming our understanding of pitching strategy dynamics.

Achieving higher novel-view quality in 3D scene generation without requiring dense volumetric representations or ground-truth supervision is a game changer for practical applications.

Agents equipped with Semantic Bayesian World Models can make nuanced decisions by treating knowledge as a dynamic web of beliefs rather than static facts.

CloudCast v2 not only extends cloud-cover forecasting from hours to 12 hours but also enhances spatial accuracy, outperforming previous models in real-time applications.

Scaling task diversity, not just dataset size, is the key to achieving robust zero-shot generalization in offline multi-agent reinforcement learning.

LeanGRPO achieves up to 1.83x speedup in diffusion RL without sacrificing optimization quality by eliminating redundant computations.

TraveL captures the nuances of traveler behavior and regional correlations, leading to a 14.7% improvement in travel time distribution estimation over existing methods.

SurgeGen can generate diverse and realistic storm surge scenarios, offering a computationally efficient alternative to traditional physics-based models.

SimSkill transforms traffic simulation mastery by autonomously evolving its capabilities and consolidating knowledge without modifying its core model.

Achievable visitation measures in reinforcement learning form a dually flat statistical manifold, transforming our understanding of planning-as-inference.

Fast and effective counterfactual explanations for routing decisions can be achieved through innovative integer programming techniques, outperforming traditional methods by a significant margin.

A single-agent CAE simulation harness can outperform complex multi-agent systems, achieving a 96.4% success rate by leveraging execution feedback and domain knowledge.

Stateagent transforms video generation by effectively tracking and inferring world states, leading to a 54% improvement in coherence across segments.

OctWorld achieves unprecedented long-range video generation with spatial consistency, outperforming traditional methods by leveraging an innovative 3D memory architecture.

A groundbreaking pipeline has generated over 8,000 hours of action-conditioned video, overcoming the limitations of real-world data collection.

Scheduling imagination in VLA models can cut GPU costs by 80% while boosting performance and robustness in real-world tasks.

Reversing turns in headland coverage can dramatically enhance autonomous driving efficiency in arable farming, especially at tricky corners.

Bridging the action-sufficiency gap in robotic manipulation, GIFT achieves up to 12.6 points improvement over existing models by integrating structured intermediate features.

Fragmentation in robot learning systems limits their effectiveness, but a unified framework could enable robots to reason and act more reliably in complex environments.

Eliminating the need for per-scene optimization, this method boosts 3D model performance while enhancing visual fidelity and spatial consistency.

Credibility in virtual testing can now be quantitatively assessed, transforming how automated driving systems are validated for safety.

LaPla reduces quantization errors in autonomous driving by transforming high-dimensional semantics into continuous actions, achieving significant performance gains over existing methods.

Rather than treating agent trajectories as dead post-training demonstrations, researchers can now resurrect thousands of fully executable, verifiable terminal environments directly from tool-execution logs.

LLM agents trained without any programmatic verifiers can actually outperform models trained on ground-truth reward signals when trajectory-level rubric judgments are dynamically decomposed into step-level advantages.

Monolithic video evaluation dilutes critical action cues, but chunking interactive rollouts into action-aligned visual evidence allows targeted reward models to outperform GPT-5.5 at scoring world model dynamics.

World models no longer require fragile offline reconstruction pipelines—baking native physics, depth, and camera pose directly into a unified multimodal generative process unlocks self-calibrating, closed-loop 3D spatial simulation at scale.

Decoupling task synthesis from on-policy rollouts solves the learning-signal saturation bottleneck in terminal agents, boosting long-horizon RL performance by up to 18 percentage points on Terminal-Bench 2.1.

Today's top video generators score near 0.8 on standard benchmarks like VBench, but none surpass 0.42 when tested on whether two co-occurring objects obey the same basic laws of physics.

LLMs struggle to generate robust travel itineraries, with substantial gaps in handling delays and crowd dynamics revealed by UTP-Bench.

LLMs systematically prioritize atmospheric elements over character-driven action in storytelling, revealing a fundamental divergence from human authorship.

By harmonizing diverse video datasets and architectures, SolarWM enables real-time interactions in video world models, pushing the boundaries of long-horizon inference.

Predicted-state matching enables web agents to achieve superior action selection by ensuring that their state predictions are discriminative, leading to better task performance.

Reinforcement learning can cut weather forecast errors by up to 45% while maintaining numerical stability in operational models.

Matching graph complexity to task granularity can dramatically enhance performance in power grid control, challenging the notion that more complex representations are always better.

RepEmp reveals that the future capacity to model and plan is more critical than immediate fidelity in representation selection, reshaping our understanding of model construction.

A pseudo-video spatial-memory Transformer achieves perfect validation accuracy on complex tasks, revealing that structured working memory can outperform mere latent capacity in predictive modeling.

DCRL reduces worst-case bootstrap depth from linear to logarithmic, leading to a dramatic improvement in long-horizon offline goal-conditioned RL performance.

Bridging the gap between graph topology inference and generative modeling could unlock new avenues for innovation in graph learning.

RideSkill revolutionizes ride-sharing by enabling real-time adaptive dispatch without the overhead of constant LLM calls, enhancing both efficiency and scalability.

CoSPOT achieves superior online time series forecasting by leveraging frequency-domain insights, allowing it to adapt to unseen patterns with minimal parameter updates.

DynG-Diff achieves superior probabilistic forecasting by dynamically adjusting guidance based on real-time variable reliability, outperforming traditional methods in noisy environments.

Variable-length action chunks can boost LLM agent performance by up to 31% while slashing decision-making time by nearly 79%.

Multi-turn jailbreak attacks reveal that framing requests as actionable can significantly reduce model vulnerability, highlighting a critical leverage point in AI safety.

A significant reasoning-generation gap exposes the limitations of current generative models, revealing that they often produce outputs that are locally plausible yet globally illogical.

VIPS reveals that integrating vehicle and infrastructure observations can significantly enhance the robustness of autonomous driving evaluations without the pitfalls of traditional simulation methods.

Test-time planning for robots can be dramatically improved by focusing on selecting reliable future-action hypotheses rather than merely generating more of them.

Trajectory replay metrics can provide a more accurate assessment of robot control performance than traditional rollout errors, challenging existing evaluation practices.

Intermittent measurements no longer spell disaster for nonlinear control systems—this framework guarantees stability and performance even in the face of uncertainty.

Target reactivity fundamentally reshapes the optimization landscape of stochastic resetting, revealing that the best search strategies depend on the targets' chemical kinetics.

Securely managing virtual card games just got a boost with protocols that can efficiently select and sort cards while preserving game integrity.

Language models exhibit substantial incoherence in probabilistic forecasts, with irrelevant details amplifying errors by an order of magnitude.

Gradient-based optimization can transform how data centers allocate load by leveraging differentiable electricity market clearing, achieving near-optimal solutions efficiently.

The PDT Framework uncovers critical gaps in decision-making reliability for future-aware autonomous driving planners, revealing that many touted improvements may not withstand rigorous scrutiny.

Achieving a 53% docking success rate with a world model approach, this research outperforms traditional reinforcement learning methods while demonstrating remarkable generalization capabilities.

Achieving state-of-the-art performance in trajectory planning, DiffuSearch reduces collisions and enhances comfort by aligning objectives across generation and refinement stages.

A predictive world model enables humanoid robots to navigate challenging terrains with a staggering 93.3% success rate, outperforming traditional methods.

SA-WAM not only integrates 3D awareness into action models but also sets new performance benchmarks in robotic policy learning.

By focusing on what changes rather than the entire scene, this model achieves unprecedented accuracy and efficiency in object manipulation tasks.

Instead of trusting black-box scalar scores to evaluate world models, video verification can match end-to-end VLM recall while producing fully auditable, spatiotemporally grounded physical proofs of failure.

Traditional causal inference requires hand-crafted pipelines and bespoke estimator training for every dataset, but causal foundation models can now infer treatment effects entirely zero-shot via in-context learning.

Time-varying transition dynamics in Poisson-Gamma models lead to a marked improvement in predictive accuracy for count time series.

Statistical priors may mimic real-world activity patterns but risk oversimplifying human behavior by fragmenting routines and limiting variability.

Solaris redefines user interfaces by generating them on-the-fly, enabling a level of adaptability and interactivity previously unattainable with traditional coding methods.

NashDreamer achieves unprecedented sample efficiency in zero-sum imperfect-information games by centralizing model learning, challenging the status quo of decentralized approaches.

Evolving causal models allows scientific agents to refine their beliefs and predictions more effectively, outperforming traditional methods in experimental settings.

Bridging robust reasoning with fast learning, this dual-process architecture boosts robotic planning efficiency and accuracy across diverse tasks.

By leveraging structured high-dimensional representations, this method reduces sampling variance and boosts decision-making efficiency in robotic systems.

A/B test simulations using data-driven personas can achieve up to 90% accuracy, drastically cutting down on the time and resources needed for real-world testing.

Elevating quantum gates to first-class values transforms how hybrid quantum-classical programs are represented and optimized.

HitMem revolutionizes 3D memory for embodied agents by effectively managing dynamic changes in environments, achieving remarkable improvements in task execution accuracy and efficiency.

RESELF achieves unprecedented accuracy in 3D scene reconstruction and motion estimation from egocentric video, outperforming traditional methods that tackle these challenges separately.

CoSMO achieves an 18.6% to 21.2% improvement in task completion rates by rethinking how edge nodes manage and offload tasks based on semantic state rather than mere freshness.

A compact 1.7B parameter policy can outperform larger models in recommendation tasks by leveraging simulated user feedback for training.

PLANET redefines multi-object tracking by embedding 3D scene geometry into query formation, leading to unprecedented accuracy in challenging scenarios.

Language is becoming a powerful interface for precise control in interactive video generation, enabling effective manipulation with minimal training.

A low-complexity LLM can effectively reduce energy poverty without compromising performance, achieving significant equity improvements while minimizing carbon footprint.

Preserving spatial topology in quantum many-body simulations can drastically reduce compilation overhead, unlocking new efficiencies in quantum computing.

Trajectory scoring can be dramatically improved by leveraging a dataset that better captures decision boundaries, leading to enhanced performance in autonomous driving systems.