AI Paper Atlas

Research map / Large language models / Agents, code and math

Hierarchical Representations for Cross-task Automated Heuristic Design using LLMs

Fei Liu, Rui Zhang, Shunyu Yao, Qinglong Hu, kefeng zheng, Zhichao Lu, Qingfu Zhang

ICML 2026 · Large language models › Agents, code and math · approach: LLMs for optimization and heuristics

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add multi-task knowledge transfer and Pareto population management (MTHS vs STHS) (ablation)MTHS · CVRPLib set Xgap to best-known (%): 5.36 -> 4.29Improved
add multi-task knowledge transfer (MTHS vs STHS) (ablation)MTHS · TSPLib 500-1000gap to best-known (%): 3.59 -> 2.64Improved
add multi-task knowledge transfer (MTHS vs STHS) (ablation)MTHS · TSPLib 50-99gap to best-known (%): 0.87 -> 0.72Mixed
use hierarchical MTHS evolved metaheuristic vs single-task LLM AHD baselines on FSSPFSSP heuristic design · Taillard FSSP averagegap to upper bound (%): 0.68 (ILS) -> 0.24Improved
use structured template metaheuristic representation instead of abstract, pseudocode or thought (ablation)MTHS · CVRP and FSSP convergencebest score vs samplesImproved
guide LLM with the MTHS metaheuristic (+MH) on unseen BPPGemini-2.5-pro · BPP n500, c150gap to lower bound: 1.32% -> 0.34%Improved
guide GPT-5-mini with the MTHS metaheuristic on unseen BPPGPT-5-mini · BPP n1000, c150gap to lower bound: 0.98% -> 0.65%Improved
guide LLM with the MTHS metaheuristic on black-box optimizationGPT-5-mini · BBOP average of top-10score (lower is better): 2.86 -> 0.32Improved
guide GPT-5-mini with the MTHS metaheuristic on admissible set problemGPT-5-mini · ASP average of top-10score (lower is better): 631.5 -> 899.1Hurt
guide Gemini-2.5-pro with the MTHS metaheuristic on admissible set problemGemini-2.5-pro · ASP average of top-10score (lower is better): 824.1 -> 742.6Mixed
multi-task search over three tasks instead of single-task searchesMTHS · TSP, CVRP, FSSPnumber of code evaluations: about 67% fewer evaluationsImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy