On Sample-Efficient Generalized Planning via Learned Transition Models
#Generalized Planning #Transition Models #AI Research #Sample Efficiency #Neural Networks #ICAPS 2026 #arXiv
📌 Key Takeaways
- Researchers developed a new approach to generalized planning using learned transition models
- Their method predicts intermediate world states rather than actions directly
- The approach achieved higher success rates with fewer training instances than previous methods
- Their research was accepted at ICAPS 2026 after being submitted to arXiv
📖 Full Retelling
Researchers Nitin Gupta, Vishal Pallagani, John A. Aydin, and Biplav Srivastava introduced a novel approach to generalized planning in artificial intelligence through their paper submitted to arXiv on February 26, 2026, which was later accepted at ICAPS 2026, addressing the limitations of current Transformer-based planners that require large datasets and suffer from state drift in complex scenarios. The research presents a fundamental shift from existing methods that directly predict action sequences to a transition-model learning approach where neural models explicitly approximate the successor-state function and generate plans by rolling out symbolic state trajectories rather than predicting actions directly. This methodology learns domain dynamics as an implicit world model by autoregressively predicting intermediate world states, offering improved generalization across planning problems sharing common domain models. The researchers systematically evaluated multiple state representations and neural architectures, including relational graph encodings, to demonstrate that their approach achieves higher out-of-distribution satisficing-plan success compared to direct action-sequence prediction while requiring significantly fewer training instances and smaller model sizes.
🏷️ Themes
Artificial Intelligence, Machine Learning Efficiency, Planning Systems
📚 Related People & Topics
Neural network
Structure in biology and artificial intelligence
A neural network is a group of interconnected units called neurons that send signals to one another. Neurons can be either biological cells or mathematical models. While individual neurons are simple, many of them together in a network can perform complex tasks.
Entity Intersection Graph
Connections for Neural network:
🌐
Large language model
2 shared
🌐
Mechanistic interpretability
2 shared
🌐
Machine learning
1 shared
🌐
Transformers
1 shared
Original Source
--> Computer Science > Artificial Intelligence arXiv:2602.23148 [Submitted on 26 Feb 2026] Title: On Sample-Efficient Generalized Planning via Learned Transition Models Authors: Nitin Gupta , Vishal Pallagani , John A. Aydin , Biplav Srivastava View a PDF of the paper titled On Sample-Efficient Generalized Planning via Learned Transition Models, by Nitin Gupta and 3 other authors View PDF HTML Abstract: Generalized planning studies the construction of solution strategies that generalize across families of planning problems sharing a common domain model, formally defined by a transition function $\gamma : S \times A \rightarrow S$. Classical approaches achieve such generalization through symbolic abstractions and explicit reasoning over $\gamma$. In contrast, recent Transformer-based planners, such as PlanGPT and Plansformer, largely cast generalized planning as direct action-sequence prediction, bypassing explicit transition modeling. While effective on in-distribution instances, these approaches typically require large datasets and model sizes, and often suffer from state drift in long-horizon settings due to the absence of explicit world-state evolution. In this work, we formulate generalized planning as a transition-model learning problem, in which a neural model explicitly approximates the successor-state function $\hat{\gamma} \approx \gamma$ and generates plans by rolling out symbolic state trajectories. Instead of predicting actions directly, the model autoregressively predicts intermediate world states, thereby learning the domain dynamics as an implicit world model. To study size-invariant generalization and sample efficiency, we systematically evaluate multiple state representations and neural architectures, including relational graph encodings. Our results show that learning explicit transition models yields higher out-of-distribution satisficing-plan success than direct action-sequence prediction in multiple domains, while achieving these gains with sign...
Read full article at source