Model Space Reasoning as Search in Feedback Space for Planning Domain Generation

James Oswald; Daniel Obolensky; Volodymyr Varha; Vasilije Dragovic; Kavitha Srinivas; Harsha Kokel; Michael Katz; Shirin Sohrabi

cs.AIApr 9, 2026

Model Space Reasoning as Search in Feedback Space for Planning Domain Generation

Authors: James Oswald, Daniel Obolensky, Volodymyr Varha, Vasilije Dragovic, Kavitha Srinivas, Harsha Kokel, Michael Katz, Shirin Sohrabi

Organizations: Rensselaer Polytechnic Institute, Troy, NY, USA · IBM Research, Yorktown Heights, NY, USA

Abstract

The generation of planning domains from natural language descriptions remains an open problem even with the advent of large language models and reasoning models. Recent work suggests that while LLMs have the ability to assist with domain generation, they are still far from producing high quality domains that can be deployed in practice. To this end, we investigate the ability of an agentic language model feedback framework to generate planning domains from natural language descriptions that have been augmented with a minimal amount of symbolic information. In particular, we evaluate the quality of the generated domains under various forms of symbolic feedback, including landmarks, and output from the VAL plan validator. Using these feedback mechanisms, we experiment using heuristic search over model space to optimize domain quality.

Figures & tables

Figure 1: An overview of our model space reasoning as search in feedback space pipeline for planning domain generation from natural language descriptions.

Figure 2: HDE score averaged across all domains for each model and feedback pipeline. Error bars indicade 95% confidence intervals over trials, i.e., 95% of trials scored within the error bars.

		Average HDE %								# Perfect @ 20
	Domain	N	LR	LS	VR	VS	LVR	LVS	Avg	N	LR	LS	VR	VS	LVR	LVS
deepseek-chat	blocks	59±5	54±5	66±5	52±4	51±5	59±5	59±4	57±5	1	1	2	0	1	0	0
	bloxorz	31±7	42±9	41±9	28±7	39±8	33±7	36±8	36±8	1	3	3	0	2	1	1
	checkers	1±1	14±5	8±4	15±5	5±3	15±5	3±3	9±4	0	0	0	0	0	0	0
	flow	33±8	40±10	55±9	43±7	42±8	29±8	37±9	40±8	1	4	4	1	1	1	3
	hiking	68±5	71±5	73±5	79±6	85±4	70±5	76±5	75±5	4	4	7	9	8	5	6

Table 1: HDE results in percentages across all domains and models, averaged over twenty trials varying model seed and description class, reported with standard error on the mean. # perfect 20 is the number of perfect HDE scores of our 20 trials, i.e the count of HDE scores = 100%. The “Aggregate” row is an average for Average HDE columns and Sum for # perfect 20 Columns. Columns from left to right: Model name; Domain name; N: No feedback; LR: Landmark feedback with random single feedback; LS: Landmark feedback with search; VR: Plan validation feedback with random single feedback; VS: Plan validation feedback with search; LVR: Landmark and plan validation feedback with random single feedback; LVS: Landmark and plan validation feedback with search; Average: HDE averages across all feedback types and the total number of perfects for that domain across all feedback types. Bold values indicate the best average performing feedback type for each domain and model combination.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Domain	$\|\mathcal{A}\|$	$\|\mathcal{F}\|$	$\|\mathcal{T}\|$	$\mathcal{A}$ arity		$\mathcal{F}$ arity		Del.
				$<$	$>$	$<$	$>$
blocks	4	5	1	1	2	0	2	yes
bloxorz ∗	3	4	3	5	7	2	3	yes
checkers ∗	4	6	2	3	5	1	2	yes
flow	3	8	2	2	3	0	2	yes
hiking	2	6	1	2	2	1	2	yes

Appendix

Table 2: List of domains with some basic properties. Columns from left to right: domain name; number of lifted actions and lifted predicates; number of types; the min. and max. arity of the actions; the min. and max. arity of the predicates; and whether a literal can be a delete effect. A

\cdot^{*}

in the name indicates that the domain is novel , not present in the training data of any of the LLMs experimented on.

Symbolic Planning

Symbolic planning represents goals, states, and actions explicitly, then searches for action sequences that satisfy constraints.

See all 89 papers

LeanPlan: Optimal Planning with LLM-Generated Heuristics and Admissibility Proofs
Oct 6, 2026André G. Pereira, Augusto B. Corrêa, Felipe Meneguzzi +1Automated Heuristic Design LLM Planning
Federal University of Rio Grande do Sul Brazil · University of Oxford United Kingdom · University of Aberdeen United Kingdom +1
Self-Improvement for Fast, High-Quality Plan Generation
May 5, 2026Robert Gieselmann, Henrike von Huelsen, Mihai Samson +9Symbolic Planning Generative Modeling
Amazon
Successor-Generator Planning with LLM-generated Heuristics
Jan 30, 2025Alexander Tuisov, Yonatan Vernik, Alexander ShleyfmanAutomated Heuristic Design LLM Planning
Independent Researcher · Computer Science Department, Bar-Ilan University

Oct 6, 2026J/K move · Enter open · S save

LeanPlan: Optimal Planning with LLM-Generated Heuristics and Admissibility Proofs

André G. Pereira, Augusto B. Corrêa, Felipe Meneguzzi, Jendrik Seipp

Federal University of Rio Grande do Sul Brazil · University of Oxford United Kingdom · University of Aberdeen United Kingdom+1

Automated Heuristic Design LLM Planning Sequential Decision Making Symbolic Planning Formal Verification

Open paper AskPDF arXiv

Frontier large language models (LLMs) can generate heuristic functions that guide search to achieve state-of-the-art performance in satisficing planning, where any plan is acceptable. However, these heuristics are not guaranteed to be admissible and can lead to suboptimal plans. We introduce LeanPlan, the first planning system that finds optimal plans with LLM-generated heuristics whose admissibility is machine-checked. Given a domain description and training tasks, an agentic loop uses planner feedback to iteratively improve a reusable domain-specific heuristic, its admissibility proof and the required domain assumptions. LeanPlan implements the heuristic, its proof and an efficient planner with machine-checked grounding and search in Lean 4. We evaluate LeanPlan on ten domains from the International Planning Competition and three new domains, using test tasks with up to 57 times as many objects as the training tasks. With GPT-5.6 Sol in the agentic loop, we successfully generate heuristics and admissibility proofs for all these domains. With the resulting heuristics, LeanPlan usually expands fewer states than the state-of-the-art Scorpion planner and solves more tasks overall.

Abstract

Symbolic Planning

Explore similar work

LeanPlan: Optimal Planning with LLM-Generated Heuristics and Admissibility Proofs

Self-Improvement for Fast, High-Quality Plan Generation

Successor-Generator Planning with LLM-generated Heuristics