Property-Guided LLM Program Synthesis for Planning
The paper discusses a novel approach to program synthesis using property-guided LLMs. This method improves efficiency by providing concrete counterexamples when a program fails to meet defined properties. The results show significant reductions in program generation and evaluation costs while enhancing the quality of synthesized programs.
- ▪Property-guided LLM program synthesis checks if a candidate satisfies a formally defined property.
- ▪When a property is violated, the system provides a counterexample, reducing the need for extensive program generation.
- ▪The approach was evaluated on ten planning domains, generating seven times fewer programs on average compared to prior methods.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.16142 (cs) [Submitted on 15 May 2026] Title:Property-Guided LLM Program Synthesis for Planning Authors:Augusto B. Corrêa, André G. Pereira, Jendrik Seipp View a PDF of the paper titled Property-Guided LLM Program Synthesis for Planning, by Augusto B. Corr\^ea and 2 other authors View PDF HTML (experimental) Abstract:LLMs have shown impressive success in program synthesis, discovering programs that surpass prior solutions. However, these approaches rely on simple numeric scores to signal program quality, such as the value of the solution or the number of passed tests.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
