World Models & Planning - Weekly Roundup

LLMs struggle with social forecasting, achieving only 75% accuracy on a benchmark that reveals critical gaps in their understanding of temporal dynamics and probability calibration.

AtlasVLA outperforms multi-view baselines by over 17% in long-horizon tasks, showcasing the power of proactive reasoning in embodied AI.

SimWAM achieves state-of-the-art performance in autonomous driving while eliminating the need for future video generation at inference, drastically reducing latency.

WorldTrace redefines memory management in video world models, achieving up to 19.5% better episodic recall without any retraining.

EffectLearner achieves unprecedented video object removal quality by effectively reasoning about complex object-induced effects in dynamic scenes.

AtlasVLA outperforms multi-view baselines by over 17% in long-horizon tasks, showcasing the power of proactive reasoning in embodied AI.

SimWAM achieves state-of-the-art performance in autonomous driving while eliminating the need for future video generation at inference, drastically reducing latency.

WorldTrace redefines memory management in video world models, achieving up to 19.5% better episodic recall without any retraining.

EffectLearner achieves unprecedented video object removal quality by effectively reasoning about complex object-induced effects in dynamic scenes.

AutoThread reduces simulation execution time by up to 83.8% while boosting throughput to 1.8x that of existing RL inference methods.

Leveraging large language models to generate synthetic transitions, ProDVI boosts sample efficiency in deep reinforcement learning without the need for extensive pre-collected data or simulators.

Token-level entropy emerges as a powerful predictor of task difficulty, revealing hidden flaws in environment design that traditional metrics overlook.

Learning to Rank models can significantly enhance the efficiency of tensor-network contraction plans, outperforming traditional methods and adapting across different GPU architectures.

SR-JEPA reveals that a predictive pathway can infer missing entity representations in 3D scenes with remarkable accuracy, transforming how we approach latent state learning.

Zero-loss exactness in optimal transport dynamics could revolutionize how we approach matching problems in high-dimensional spaces.

Langevin correction can transform the way flow-based generative models optimize policies, leading to sharper and more accurate sample generation during reinforcement learning.

Agentic LLM-guided feature selection boosts mortality prediction accuracy in cardiac arrest cases, achieving state-of-the-art results with a fraction of the parameters.

GSBF achieves beamforming without the need for instantaneous channel state information, dramatically reducing latency and complexity in MIMO systems.

Hierarchical Latent Prediction reduces error accumulation in language models, enabling coherent long-horizon reasoning and more efficient decoding.

No physics engine is uniformly faithful, with critical failures in simulating impulsive contact and rapid textile motion revealed by the GAUGE benchmark.

DreamGuard achieves a groundbreaking safety-utility balance by predicting long-term risks, outperforming traditional guardrails that only react to immediate threats.

HoloWorld achieves a groundbreaking integration of indoor and outdoor urban generation, enhancing spatial coherence and visual identity across entire cityscapes.

Noise-aware residual correction boosts the realism of autoregressive audio-visual generation, tackling issues of identity drift and desynchronization head-on.

Current action-conditioned world models are limited by their reliance on visual patterns, failing to generalize physical dynamics across different robot embodiments.

Simply adding multi-view videos to latent action models doesn't guarantee 3D awareness; LAWM-3D reveals the critical design choices needed for success.

Early trajectory decoding from video diffusion models can cut planning latency by nearly 50% without sacrificing decision quality.

Achieving a balance between local detail and global coverage, this method significantly enhances UAV photogrammetry accuracy and completeness.

A single model can now adaptively balance short-term precision and long-term accuracy in weather forecasts, revolutionizing how we approach atmospheric predictions.

Kastor reduces forecasting error by nearly 43% while enhancing the physical fidelity of simulations, outperforming traditional methods in both accuracy and efficiency.

Ignoring future cross-variable dependencies can lead to significant forecasting errors, but a new structural regularizer, CvLoss, bridges this gap and boosts model accuracy.

Staging social interactions in trajectory prediction leads to remarkable gains in accuracy and consistency, reshaping how we approach agent behavior modeling.

Active reasoning in holonic digital twins can revolutionize how AI interacts with the physical world, enabling real-time coordination and decision-making in uncertain environments.

iARCS transforms 3D scene generation by ensuring that synthetic environments meet essential functional constraints while maintaining diversity and realism.

The cascade hybrid model outperformed traditional forecasting methods, achieving an impressive R² of 0.889 in predicting cattle weight gain despite irregular data collection.

Achieving the highest fidelity in mobile GUI interaction, AppDeltaWorld redefines how agents predict and interact with app interfaces by focusing on code updates instead of images.

AI-powered Student Digital Twins could revolutionize higher education by transforming how institutions prevent student failure and enhance career alignment.

Integrating historical predictions into geospatial models boosts crop classification accuracy by over 1.6 percentage points, correcting significant recall biases.

PhyLatent slashes failure rates in JEPA world models while boosting predictive control success from 70% to 78%—and even up to 98% in complex environments.

Robust-WAM achieves superior out-of-distribution generalization in robot control by seamlessly integrating semantic foresight into action predictions while leveraging extensive VGM pretraining.

$\omega$-0 enables humanoid robots to seamlessly integrate movement and manipulation, outperforming traditional models by predicting coordinated actions directly from sensory inputs.

TrajDebug uncovers the root causes of failures in long-horizon agent trajectories, enabling targeted improvements that could significantly boost agent performance.

SJRL not only overcomes collision challenges in multi-agent navigation but also adapts dynamically to real-world constraints, outperforming traditional methods in complex environments.

GeniWorld achieves robust zero-shot generalization in robotic manipulation, outperforming traditional models even with minimal training data.

Existing economic simulations are stuck in the past—this blueprint reveals how to build adaptive, self-evolving economic agents that could transform decision-making in AI.

World rehearsal enables LLM agents to internalize environment dynamics, achieving superior performance without costly external interactions.

Explicitly disentangling world dynamics from view rendering allows MAS to achieve unprecedented scalability and consistency in multiplayer simulations.

Trajectory scoring in aerial navigation can be revolutionized by focusing on unexplainable prediction discrepancies, leading to more robust and efficient UAV navigation.

Training LLM agents without expert supervision can lead to better performance and generalization across diverse environments.

Training under stationary ambiguity allows control policies to maintain robustness against shifting latent factors, crucial for applications like financial hedging.

Relying on stale spatial memory can more than double an agent's failure rate, revealing a hidden safety risk in memory-augmented VLMs.

Latent video flow-matching enables atmospheric data assimilation that not only propagates information seamlessly but also competes with state-of-the-art forecasting models using sparse data.

HelloWorld allows users to engage in seamless social interactions with video characters, achieving superior quality without sacrificing visual aesthetics.

Shared classical randomness can unlock a vast array of distributions in shallow quantum generative models, outperforming purely unitary approaches that struggle with long-range correlations.

Exploration bonuses can either amplify or neutralize memory architecture differences, depending on how memory content is acquired and supervised.

Myopic planners can fail spectacularly when faced with the need to acquire capabilities for future experiments, leading to unbounded approximation ratios in goal-directed discovery.

NBDM outperforms traditional models by effectively capturing nonlinear dynamics, achieving superior long-horizon forecasting accuracy even with missing control inputs.

Argus achieves a 78% success rate on long-horizon reasoning tasks while using 21% fewer tokens in mature workflows, showcasing a revolutionary approach to agentic autonomy.

NodeJEPA reveals that masking structural information can significantly improve node-level representation learning without relying on traditional reconstruction methods.

Football-aware simulations can boost exact-score forecasting accuracy by over 4% while revealing critical limitations in LLM integration.

AI-generated hazard scenarios can now be traced back to real-world ASRS reports, enhancing operational safety analysis in aviation.

Explicit language memory boosts VLA model performance in long-horizon tasks, enhancing both success rates and interpretability.

Shared rollouts can distort compliance scores, leading to misleading evaluations of driving policies, with implications for safety and performance assessments.

PhysMind achieves a remarkable 38.23-point accuracy boost in physical reasoning tasks by transforming videos into executable worlds without the need for training.

DreamWAM achieves up to 75.47% accuracy in unseen scenarios, showcasing that structured future state representations can dramatically enhance action model performance beyond traditional RGB methods.

MobileWAM achieves superior mobile manipulation performance by seamlessly integrating foresight into action planning, outpacing state-of-the-art methods.

EgoAfford reveals that effective task-oriented affordance grounding can significantly enhance multi-step planning in complex environments.

Mimir achieves an impressive 86.0% success rate on long-horizon tasks, outperforming leading closed-source models by a significant margin.

muSync-GS achieves unprecedented realism in driving video synthesis by synchronizing vehicle dynamics with environmental changes, ensuring that every scene edit reflects true physical interactions.

Achieving a 49% increase in success rate on out-of-distribution tasks, Faster-WAM redefines efficiency in future-aware robot manipulation models.

SCOPE guarantees safe navigation in unknown 3D environments by transforming uncertified paths into actionable observation tasks, achieving near-zero risk of entering unsafe areas.

Reversible action cycles enable self-verification in long-horizon planning, cutting state returning drift by 44% and boosting accuracy nearly 4x.

Long-horizon LLM agents can achieve 96.9% task success by learning to adapt their external execution support through trainable harness policies.

SpecRoll achieves up to 2.15x faster generation in RL rollouts by cleverly balancing fast and slow adaptation strategies.

Compiling recursive logic programs into quantum annealers could revolutionize how we solve complex computational problems by ensuring optimal solutions are reached efficiently.

A novel hybrid planning-learning architecture enables UUVs to navigate dynamically changing underwater environments with unprecedented robustness and safety.

Switching policies derived from stationary policies can achieve optimality in constrained sc-LTL planning, balancing objectives and safety like never before.

A new two-step method for model-free reinforcement learning guarantees optimal policy convergence for co-safe LTL objectives, overcoming significant challenges in traditional approaches.

PRIMAL3 achieves unprecedented scalability in multi-agent pathfinding, successfully coordinating 100,000 agents while navigating complex environments.

GASP achieves analytical-level success in collision-aware motion generation while cutting inference time to near milliseconds, revolutionizing real-time planning for robotics.

Action-conditioned world models can be misled by statistical biases, but a new framework shows how to enforce true action dynamics for superior performance.

Chord recovery accuracy jumps from 18% to 54% with minimal supervision, showcasing a new frontier in music co-creation agents.

ABSeeker's innovative credit assignment method allows it to achieve performance levels comparable to much larger models, redefining expectations for long-horizon search agents.

Recursive task synthesis not only slashes generation costs to $0.05 per task but also produces increasingly complex challenges that boost model performance by up to 10 points on key benchmarks.

WorldClaw can generate expansive, editable 3D worlds from text prompts while maintaining both global coherence and intricate local details.

The intrinsic geometry of GFlowNets reveals surprising insights into when temporal interactions can be ignored, fundamentally changing how we approach forward-policy training.

Hybrid agents that combine LLM-driven planning with RL optimization achieve superior performance in complex decision-making tasks, outperforming traditional methods.

POEM's innovative use of $\mathrm{SO}(2)$ rotations reveals how to effectively manage periodicity drift, transforming time series forecasting.

Local correction mechanisms in offline RL can significantly enhance value estimation stability, reducing the impact of out-of-distribution errors.

Score matching isn't essential for diffusion model training—it's a byproduct of how we reverse the reference process.

Achieving over 90% success in robotic tasks with a model that can be trained on a single GPU challenges the notion that high performance requires massive computational resources.

TARL revolutionizes memory management for long-term agents by enabling nuanced updates that significantly enhance state recovery and reduce corruption.

Planning failures in multilingual systems can be systematically diagnosed and mitigated, leading to significant performance gains in low-resource languages.

Pivot-Centric Trajectory Prediction achieves state-of-the-art accuracy in long-horizon forecasting by transforming the prediction process into manageable sub-tasks, effectively minimizing errors.

LLMs may excel at predicting outcomes in sports, but they often converge on incorrect answers, revealing a critical flaw in their forecasting abilities.

Achieving faster convergence to terminal constraints, this method drastically reduces sample inefficiency in trajectory optimization for complex robotic tasks.

Achieving a 2dB PSNR improvement, EmbodiedVAE transforms how robots learn and execute manipulation tasks by providing compact and controllable latent representations.

SHM-POMDP achieves 2.5x higher information gain in geologic exploration while maintaining principled uncertainty quantification.

Passively safe guidance strategies could revolutionize autonomous space operations by ensuring reliable rendezvous even in the face of navigation uncertainties.

Interleaving reasoning and execution allows PACE to hide 66.8% of thinking time within action execution, drastically improving planning efficiency.

Role-aware evidence routing in surgical video prediction leads to superior performance, demonstrating that asymmetric information exchange can enhance multi-observer models.

OneDayAgent achieves a groundbreaking 0.821 score on long-horizon tasks, proving that a single harness can effectively manage execution across diverse LLM backends without tuning.

Certifying action safety in memory-grounded agents can drastically reduce the risk of unsafe commitments, ensuring reliable decision-making in complex environments.

LLMs struggle with social forecasting, achieving only 75% accuracy on a benchmark that reveals critical gaps in their understanding of temporal dynamics and probability calibration.

CUDA MPC achieves real-time optimization for high-dimensional and fast-dynamic systems, outperforming traditional solvers by up to 965 times.

FORTUNE redefines UAV path planning by effectively managing uncertain PoI demands while prioritizing public safety and operational efficiency.