Language models design heuristics for logistics-style problems, if you give them structure
The gain came from the framework, not a bigger model, the authors say.
Của The desk, News · · 1 phút đọc
Chịu trách nhiệm: Dimitri Gärtner
Thẻ: optimisation · algorithms · language models · nature
Bài này chưa có bản Tiếng Việt, nên bạn đang đọc tiếng Anh.
- Tin
- The gain came from the framework, not a bigger model, the authors say.
- Vì sao đáng kể
- A Nature Machine Intelligence paper presents LACE, which scored 0.945 across 36 classic optimisation problems versus 0.571 for directly prompting an LLM.
- Điều đã được xác nhận
- The photograph is generated for this story.
Gắn với một nguồn được nêu tên. Chưa đánh dấu là đã kiểm chứng. Nature Machine Intelligence: Large language models discover complementary heuristics for combinatorial optimization
The news
A paper published in Nature Machine Intelligence on October 1 introduces LACE, short for LLM-driven Algorithm Construction via Complementary Evolution. It breaks algorithm design into a fixed contract of inputs, outputs, a tool library and a portfolio of heuristics, then evolves specialist heuristics that cover different kinds of problem instances within strict runtime limits.
Why it matters
Combinatorial optimisation runs scheduling, routing, hospital rosters and energy grids. Automating heuristic design could cut months of expert work per problem.
The details
Across 36 classic CO-Bench problems, LACE reached an average score of 0.945, against 0.870 for the strongest existing LLM-based method and 0.571 for direct LLM prompting without a framework, the authors report. They attribute the gain to the framework rather than the model.
On four structurally new problems, LACE scored 0.97 to 0.99, while five existing LLM-based baselines failed to produce any feasible algorithm.
What we could not check
- Scores are on the authors' chosen benchmark suite.
- We could not check real-world deployment results.
Nguồn
The photograph is generated for this story. It does not show the people named in it.