World Models & Planning - Weekly Roundup

By making environment design a learnable process, SPADE unlocks a new frontier in self-improvement for language agents, leading to substantial performance gains across diverse tasks.

Training on environments synthesized from real-world business scenarios boosts agent performance on both enterprise tasks and diverse benchmarks, revealing a new frontier in scalable RL training.

RecVerse outperforms traditional simulators by maintaining a nuanced understanding of user intent and memory, leading to more realistic shopping behaviors.

Orthogonal JEPA achieves superior predictive performance by factorizing latent states, allowing for more efficient learning in complex systems.

Nearly half of New York's predictive service line classifications show alarming inconsistencies, with significant implications for public health and safety regarding lead exposure.

Training on environments synthesized from real-world business scenarios boosts agent performance on both enterprise tasks and diverse benchmarks, revealing a new frontier in scalable RL training.

RecVerse outperforms traditional simulators by maintaining a nuanced understanding of user intent and memory, leading to more realistic shopping behaviors.

Orthogonal JEPA achieves superior predictive performance by factorizing latent states, allowing for more efficient learning in complex systems.

Nearly half of New York's predictive service line classifications show alarming inconsistencies, with significant implications for public health and safety regarding lead exposure.

Identifying piecewise-smooth dynamical systems could revolutionize our understanding of complex systems in climate dynamics and mechanics.

A simple Gaussian counterexample reveals that existing concentration inequalities for discounted least-squares estimators are fundamentally flawed, necessitating significant corrections.

HVAC systems can be optimized for energy efficiency and occupant comfort with a model that adapts seamlessly to unseen environments, achieving significant performance gains.

Achieving collision-free navigation for AUVs and ASVs near complex offshore wind infrastructures, this approach cuts goal-distance error by up to 93% through physics-informed planning.

Turning circle-based control barrier functions enable autonomous surface vehicles to navigate complex environments without predefined paths, achieving superior safety and efficiency.

Jointly forecasting surgical instrument trajectories and visual states could revolutionize surgical motion planning by providing a more coherent understanding of action-scene dynamics.

End-to-end learning can drastically improve early classification accuracy in non-stationary environments, outperforming traditional methods that treat classification and triggering as separate tasks.

CLaST achieves unprecedented accuracy in time series forecasting by preserving contextual relationships, outperforming existing models by up to 48.6%.

Pre-trained classifier weights can act as robust prototypes for adapting to shifting label distributions in open-world scenarios, significantly enhancing model reliability.

Milestone inference can significantly enhance credit assignment in long-horizon reinforcement learning, leading to superior agent performance without extra model complexity.

Classical world models can fail to align with reality, but a quantum approach using a single qutrit achieves perfect alignment and optimal policy matching.

GFHNNs can outperform conventional neural networks in learning Hamiltonian dynamics while requiring significantly less training data.

Conversational surveys combined with multimodal LLMs can enhance travel behavior predictions, achieving over 71% accuracy by leveraging visual context.

RuleMaze reveals that separating perception, execution, and rule verification can dramatically enhance MLLMs' ability to follow complex natural-language instructions in spatial planning tasks.

DECOWAM achieves a 21.7% reduction in action prediction error while maintaining robust task performance, showcasing the power of embodiment-aware factorization in mobile manipulation.

DARS achieves superior performance in instruction-based image editing by transforming outcome-level feedback into actionable, localized supervision for both planning and rendering stages.

Fine-tuning a language model for crowd simulation can significantly improve accuracy using only aggregate mobility data, achieving a 25% reduction in destination-share error.

Accurate trajectory forecasting doesn't guarantee understanding of the underlying physical properties, as shown by ExPhy's insights into model performance.

StateMem improves current-state accuracy by up to 1.8x, revealing that existing memory systems are ill-equipped to handle evolving contexts in LLM interactions.

Landmark-guided optimal transport reveals hidden geometric transformations, enhancing distribution matching beyond mere cost minimization.

Modern autonomous driving systems hinge on learned representations that prioritize safety and compliance, not just raw performance metrics.

Hierarchical tactile modeling boosts robot manipulation success rates by over 40% in contact-rich tasks, revealing the power of structured tactile forecasting.

Static environments can be transformed on-the-fly to better suit agent learning, resulting in up to a 9.0-point performance boost with fewer execution steps.

Adaptive probabilistic shielding can evolve in real-time, enhancing safety in reinforcement learning without sacrificing exploration.

TabPFN-TS achieves near-parity with state-of-the-art models in heat load forecasting while eliminating the need for retraining, paving the way for more adaptable energy management solutions.

RISE not only boosts planning efficiency but also redefines how imagination budgets can be adaptively allocated in complex environments.

Stream4D transforms video generation by enabling coherent motion and dynamic scene representation, outperforming static critics that hinder realism.

Usability testing reveals that DesCartes Builder empowers domain experts to create real-time digital twins with unprecedented ease and reliability.

Continuous-time reinforcement learning can effectively tackle non-Markovian Hawkes processes, outperforming discrete methods in optimization tasks.

The Wasserstein entropic value-at-risk reveals hidden catastrophic risks that traditional measures overlook, adapting agent behavior dynamically as confidence grows.

Multi-agent DDPG achieves full coverage and a remarkable fairness index of 0.94, outperforming single-agent methods in complex campus environments.

Linear masking can lead to critical components being entirely omitted from dynamical models, but a new algebra-based scoring method recovers these components with significantly fewer coordinates.

Constant exploration in time-varying Gaussian process bandits can yield sharper regret bounds, challenging the need for increasing exploration parameters with horizon length.

Achieving a 32% reduction in energy consumption while cutting waiting times by 30% could redefine operational efficiency in data centers powered by LLMs.

RP1 achieves near-perfect planning success with 1,000 times fewer world-model rollouts and is up to 67 times faster than traditional methods.

A novel computational framework reveals how team communication dynamics in VR can be dissected into coherent phases that align with specific actions, enhancing our understanding of collaboration.

PU-HNO outperforms traditional methods by learning stable propagation structures from noisy labels, revolutionizing indoor radio map generation.

SparsePR cuts attention-reconstruction error while speeding up video generation by over 2.5x without sacrificing quality.

Future scene predictions can now directly influence trajectory selection in autonomous driving, enhancing decision-making accuracy.

AlphaClifford consistently outperforms state-of-the-art synthesis heuristics by reducing gate counts while using a less expressive gate set, revolutionizing Clifford circuit optimization.

Existing video quality metrics fall short for camera-controlled generation, but CWQA sets a new standard by accurately predicting perceptual quality with a tailored approach.

Audio and acceleration modalities can predict physical interactions with surprising accuracy, but the executor's architecture is the true bottleneck in achieving effective multimodal execution.

SCAPE reduces scenario-level prediction error by up to 34.7%, enabling safer and more efficient deployment of robot-learning policies in real-world environments.

Grid cells can reduce spatial aliasing by up to 99%, transforming how we understand place representation in complex environments.

SAM-TD enables stream-based robotics planning to adhere to complex temporal constraints, a capability previously unattainable in this domain.

FS-MPC achieves superior sample efficiency and stability in controlling high-dimensional robotic systems, outperforming traditional methods in challenging tasks.

Long-horizon planning with online adaptation can significantly enhance service robots' efficiency in environments with unpredictable, time-varying rewards.

PEF not only boosts navigation success rates in complex vascular environments but also adapts seamlessly to patient-specific anatomies, paving the way for improved clinical outcomes.

BLS cuts assembly planning time in half while ensuring collision-free execution, transforming how robots approach complex assembly tasks.

Runtime for complex project-scheduling simulations can be slashed from over 1,200 seconds to under 200 seconds using agentic AI optimizations, saving researchers significant computational resources.

By making environment design a learnable process, SPADE unlocks a new frontier in self-improvement for language agents, leading to substantial performance gains across diverse tasks.

Action-conditioned objectives can significantly enhance the effectiveness of Euclidean-cost MPC by aligning latent representations with real task progress.

SkillGate achieves a 30% boost in trial success for long-horizon agents by fundamentally rethinking how skill selection is rewarded during execution.

Managerial behavior, not model size or vendor, dictates success in long-term decision-making tasks, as evidenced by the surprising performance of claude-fable-5 in FM-Bench.

Acceptance in Code World Models certifies sample consistency but often overlooks critical events, leading to substantial planning failures.

Understanding how to recover lost distinctions in model performance could revolutionize the way we approach architecture design and deployment in AI systems.

Contrastive inverse dynamics can boost performance by over 20 points on complex tasks without the need for Gaussian regularization or pretraining.

Achieving high-fidelity causal network reconstruction and forecasting accuracy without relying on a single global hyperparameter could revolutionize our understanding of complex dynamical systems.

Integrating general-purpose knowledge graphs with traffic sensor data can dramatically enhance forecasting accuracy, revealing insights that traditional methods overlook.

SPACE recalibrates multivariate forecasting by leveraging current sample clouds, resulting in improved coverage efficiency that outperforms traditional methods.

Achieving a staggering 90.4% reduction in robot-motion error, Hydra-0 redefines how we model and control robotic actions across varied environments.

HLSR achieves real-time congestion avoidance by intelligently rerouting only the vehicles that need it, enhancing urban mobility without overwhelming computational resources.

Teams can achieve solutions through iterative coordination that individual planning cannot unlock, but some goals may remain fundamentally unverifiable due to representational constraints.

Certifying and synthesizing policy portfolios for robust MDPs is computationally harder than previously thought, with implications for both theory and practice.

Counting policies instead of agents transforms DecPOMDPs from intractable to efficiently solvable, paving the way for scalable multi-agent systems.

RATTL allows agents to dynamically balance caution and reward maximization, adapting their decision-making as they learn about their environment.

Surrogates can now handle over 3 million controllable variables, achieving up to 26.5x speedups over traditional simulation methods.

LLMs can significantly outperform traditional symbolic methods in PDDL model repair, but their reliability still falters in complex domains.

Static benchmarks mislead model performance assessments, as LiveHouse-TS reveals dramatic shifts in rankings when evaluated in real-time environments.

Zero-shot adaptation to new tasks is now feasible with neurosymbolic world models that leverage structured symbolic components for reward prediction.

GS-Voxel revolutionizes 3D scene generation by allowing scalable, fitting-free structured latents that adapt to the complexity of the scene without the need for per-scene optimization.

Short trajectories in Kac's walk can be indistinguishable from Haar measure after just $O(n(k+\log n)\log n)$ steps, challenging previous assumptions about mixing times.

Dijkstra's algorithm can be transformed into a powerful exact planner for stochastic navigation, achieving efficiency gains of up to 19 times less computational work than traditional methods.

A unified policy enables humanoid robots to robustly interact with complex environments, achieving record-breaking success rates in dynamic terrain and command scenarios.

Group Decentralized RHCR achieves high throughput for multi-agent pathfinding while slashing computational costs, making it viable for larger agent counts.

GAPL achieves a remarkable reduction in collision rates and displacement errors, showcasing the potential of LLMs in trajectory planning for autonomous driving.

Coupling predictive safety with action-conditioned modeling allows heterogeneous robots to navigate complex environments with unprecedented reliability and reduced risk.

Trust assessments in vehicular networks can be effectively aggregated into a continuous field, revealing critical low-trust patterns that traditional methods miss.

Quantum simulations could revolutionize our understanding of molecular reaction dynamics, but they face critical theoretical challenges that could limit their effectiveness.

Counterfactual recourse in education becomes truly actionable when recommendations are not just model-valid but also semantically feasible and machine-checkable.

Traditional risk assessment fails for AI, but a new capability-based planning framework reveals actionable insights for crisis preparedness.

WorldMind achieves a breakthrough in NPC behavior by decoupling state understanding from action generation, leading to significantly more coherent and context-aware interactions in gameplay.

Combining visual exploration with symbolic planning enables agents to achieve over 90% success in executing complex household tasks, far surpassing traditional methods.

Agents can significantly reduce communication needs by leveraging memory more effectively, revealing a critical balance between remembering and signaling.

OnGameLearn not only navigates the complexities of strategic interactions but also adapts to evolving contextual signals, achieving superior performance in competitive pricing scenarios.

QWM achieves unprecedented sample efficiency in reinforcement learning by leveraging world models without succumbing to compounding bias, outperforming prior methods on challenging benchmarks.

HarnessEval-W transforms world model evaluation from mere scoring to a transparent reasoning process that mirrors human judgment.

Analytical prior information can cut prediction errors by over 89% compared to direct learning methods when simulation data is limited.

Video world models are significantly miscalibrated, often collapsing to a single outcome and failing to reproduce expected distributions.

Real-time interaction with physics-governed simulations from monocular video is now possible, even with unseen forces.

The learned dosing policy for sepsis not only surpasses clinician performance but also suggests a shift towards less intravenous fluid, challenging traditional practices in critical care.

TAILS resolves cross-task ambiguities in continual learning by directly correcting feature representations, leading to significant performance boosts without changing the underlying model.

Planning performance can be dramatically improved by aligning latent embedding geometry with task-relevant state representations, as shown by SCALE's consistent outperformance of LeWM.

Identical delay summaries can yield vastly different regret outcomes, highlighting the crucial impact of timing in bandit optimization.

Achieving a 91.7% success rate in obstacle avoidance, Orbit-Planner redefines how satellite agents navigate dynamic environments without relying on fixed maps.

Semantically informative labels can skew LLM decision-making, leading to either enhanced performance or catastrophic failures depending on alignment with reward structures.