MEMOIR: Memory-Guided Tree Search for LLM Solver Synthesis
A novel framework named MEMOIR advances the synthesis of large language model (LLM) solvers for combinatorial optimization (CO). CO is crucial for decision-making in areas such as logistics and chip design, where infeasible solutions are ineffective, and even minor improvements can lead to significant economic benefits. Recent advancements have utilized LLMs to automate the creation of executable solver programs from natural-language descriptions. Current tree-search and evolutionary agents enhance candidate pathways in parallel, lacking explicit knowledge transfer, which can lead to constraint violations and convergence on similar algorithm types. MEMOIR features a two-tier memory system: branch-local memory maintains execution-based refinement details during iterations of a single algorithm design, while global memory retains compressed summaries of algorithms and failure modes across branches. Insights are distilled during a reflection phase at the end of each branch. The framework is elaborated in arXiv:2605.17539.
Key facts
- MEMOIR is a memory-guided tree-search framework for LLM solver synthesis.
- It targets combinatorial optimization problems in logistics and chip design.
- Existing methods lack explicit knowledge transfer between branches.
- MEMOIR uses branch-local and global memory levels.
- Branch-local memory stores refinement details within a branch.
- Global memory stores algorithmic and failure-mode summaries across branches.
- A reflection step at branch termination distills insights.
- The paper is available on arXiv with ID 2605.17539.
Entities
Institutions
- arXiv