LLM world models pass mechanical checks but miss intended structure
Of 134 valid game-track models, mean structural F1 is 0.366 and only 28 exactly match the reference.
Stories per day · 30 days
2 RSS
Relatedworld-models7robotics3benchmark2embodied-ai2reinforcement-learning23d1agents1ai-tool1benchmarks1causal-reasoning1cheminformatics1compositional-generalization1compositionality1configwm1embodied-agents1generalization1google-deepmind1hybrid-planning1
30 stories, newest first · one card per story
Stories tagged #planning
Quasipolynomial Approximate Planning for Decodable POMDPs
Introduces quasipolynomial-time approximate planning algorithms for ρ-decodable POMDPs, which are weaker than observability assumptions.
PlanAssist AI cuts planning officer time by 58%
PlanAssist enters 12-month beta with 15 UK councils, full rollout planned for 2027.
Reaction verification during search improves retrosynthesis routes
On Retro190, feeding verifier results into search lifts verified solvability from 51.1% to 62.6%.
SERA self-evolving harness lifts long-horizon robot planning on RoboMME
With a fixed VLA executor, autonomous harness selection reached 51.38% success versus 44.38% without self-evolution.
CausGround grounds action preconditions for embodied robot planning
Without parameter updates, reported VirtualHome success rises from 24.9% to 53.3% and RoboCasa from 43.0% to 57.0%.
WorldStream streams open-ended 3D worlds with persistent editable state
Hierarchical planners compile terrain ahead of the player while one world record restores prior visits and edits.
ConfigWM uses task language to filter distractors in world-model states
Authors report ConfigWM beats the strongest baseline by 22.3% on average as visual noise rises.
Slot-JEPA adds object slots and masks to latent world models
Frozen slot features stay transferable on seven benchmarks and improve planning on six planar and four robot tasks.
Agima trains one model to imagine outcomes before acting
After SFT on two-box Sokoban, Agima solves 40% of four-box puzzles versus 2% for an action-only planner.
AnyGoal tests whether world models generalize past expert rollouts
The benchmark recombines in-support states and actions that expert trajectories rarely pair, and current models fail those cases.
AnyGoal tests world-model generalization beyond expert rollouts
Current models fail when familiar state and action factors are recombined into underrepresented combinations.
Paper defines solver-sufficient symbols for hybrid robot planning
A counterexample-guided method invents symbols that cut solving cost on simulated and real robots while keeping solution quality.
Towards Safe World Models: Risks, Defenses, and Evaluation
A survey maps world-model safety risks, defenses, and evaluation across six lifecycle stages using 163 works.
Learning Commute-Time-Preserving World Models for Planning
CTWMs learn commute-time-preserving latents for planning and match LeWM with half the parameters.
Belief-Aware Multi-Agent Path Finding under Map Uncertainty
MAGIC updates shared traversability beliefs so multi-agent pathfinding avoids costly reroutes under unknown map changes.
DAGent: Evaluate-then-Grow Planning for Deep Research Agents
DAGent grows research-agent plans from evidence and beats open baselines on BrowseComp-Plus, GAIA, and xbench-DeepSearch.
JEPA-TTT: Persistent Test-Time Training of Latent World Models for Planning under Dynamics Shifts
JEPA-TTT adapts a pretrained latent world model at test time when environment dynamics shift.
ATLAS: Aligned Transport of Latent Structure for Reliable World Model Planning
ATLAS preserves relational latent geometry so world-model planners keep novelty structure for reliable goal reaching.
Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution
Pattern-specific executors close LLM agents' plan-execution gap, lifting ALFWorld success from 0.48 to 0.92.
FONDANT: Strong and Best-Effort Planning via Antichains
FONDANT computes strong and best-effort FOND policies with antichains and improves coverage on larger instances.
FlexiWorld: Learning and Planning via Flexible Action Chunks Across Multiple Time Scales
FlexiWorld uses variable-length action chunks and mixed-span goals, reaching 89.29 percent mean success.
Adaptive Latent Capacity for World Models
ALeWM learns compact latent prefixes for JEPA world models and beats fixed-width LeWM.
Omni-Decision: Evidence-Ledger Planning for Omni-Modal Agents
Omni-Decision plans omni-modal agent tasks from a compact evidence ledger instead of noisy history.