Background
Recent papers have shown that planning is one of the tasks that remains very challenging for LLMs. Specifically LLMs struggle at generating long plans to achieve a goal without making mistakes such as prescribing the execution of inapplicable actions [1,2]. For simple planning benchmark domains, better results have been obtained by using LLMs to generate domain-specific programs (e.g. python code) that produces plans, policies, or heuristics to guide conventional state space search [3,4,5]. However, even in this case, performance drops significantly when the vocabulary used to describe the planning domain diverges from the standard one, suggesting that LLMs have not grasped the general reasoning mechanisms that underly planning.
Aim
In this research, we will use LLMs to help planning in novel ways. There are two distinct and largely independent subprojects. Both of them will use plan space planning [6,7], which is a different planning paradigm than state space search and we postulate is better suited to exploit the capabilities of LLMs. Interestingly, whilst this paradigm closely captures the reasoning required to build plans, it lost popularity over state-space search due to the difficulty of finding high-performance domain-independent heuristics to navigate the more complex search space. LLMs provide an opportunity to remedy this and restore partial-order planning to its rightful place. The plan space planning paradigm is also well suited to using LLMs as a source of commonsense knowledge during planning, to determine which real-world actions may be useful to fulfill a given purpose in a plan without resorting to a pre-define domain model given a priori (open world planning).
- In the first project, we aim at using the LLM to generate heuristics to guide plan-space planning and improve the scalability of the latter.
- In the second project, we aim at replacing the domain model planners start from, with the commonsense-knowledge of the LLM, leading to an open-world partial-order planning system.
The main research questions are whether LLMs are capable of fulfilling those roles and if so how, and to design, implement, and evaluate possible approaches.
Ideal Student for This Project
- computer science student
- HD average
- excited by AI research and AI planning
- enjoys programming
- has taken ANU AI classes, with a good balance between search, reasoning, planning, and machine learning aspects
- experience with LLMs (in-context learning, fine-tuning) would be a plus
Other particularities
This project is in collaboration with CNRS in France
Contact
If you are interested, please email Sylvie Thiebaux (Sylvie.Thiebaux@anu.edu.au)
- Which of these two projects interest you most
- The type of research project you are seeking (12-unit, or 24-unit)
- A copy of all your transcripts
- CV
- Any questions you may have
References
[1] Karthik Valmeekam, Matthew Marquez, Alberto Olmo Hernandez, Sarath Sreedharan, Subbarao Kambhampati (2023): PlanBench: An Extensible Benchmark for Evaluating Large Language Models on Planning and Reasoning about Change.
[2] Karthik Valmeekam, Kaya Stechly, Subbarao Kambhampati (2024): LLMs Still Can’t Plan; Can LRMs? A Preliminary Evaluation of OpenAI’s o1 on PlanBench. CoRR abs/2409.13373
[3] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, Michael Katz (2024): Generalized Planning in PDDL Domains with Pretrained Large Language Models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI’24), pages 20256-20264.
[4] Dillon Z. Chen, Johannes Zenn, Tristan Cinquin, Sheila A. McIlraith (2025): Language Models For Generalised PDDL Planning: Synthesising Sound and Programmatic Policies. In Proceedings of the 18th European Workshop on Reinforcement Learning (EWRL’25).
[5] Augusto B. Corrêa, André Grahl Pereira, Jendrik Seipp (2025): The 2025 Planning Performance of Frontier Large Language Models. CoRR abs/2511.09378
[6] Malik Ghallab, Dana Nau, Paolo Traverso (2025): Plan Space Planning (Section 3.4). Acting, Planning, and Learning. Cambridge University Press.
[7] Arthur Bit-Monnot, Roland Godet (2025): Towards Canonical and Minimal Solutions in a Constraint-Based Plan-Space Planner. Proceedings of the European Conference on Artificial Intelligence (ECAI’25), pages 4678 - 4685.