Harnessing LLMs with Direction-Scope Control for Combinatorial Solver Refinement
Abstract
Harnesses are increasingly unlocking the potential of large language model (LLM) agents by making their actions more controllable. In automatic heuristic design (AHD) for combinatorial optimization (CO), LLMs iteratively revise solvers using execution and verification feedback. However, these revisions often break previously validated solver behavior, making refinement unstable and costly. We characterize this pattern as feasibility oscillation and document it across CO benchmarks, AHD methods, and backbone LLMs. Tracing this oscillation, our analysis reveals that feasibility degradation often follows large solver revisions, while substantial feedback redundancy makes it harder to identify which failures deserve attention. Motivated by these findings, we introduce DiSH, a harness that jointly controls revision direction and scope. It selects representative failure evidence to guide each update and progressively narrows the revision scope as feasible behavior accumulates. Across nine CO benchmarks, DiSH consistently improves QYI scores and reduces feasibility oscillation, while roughly halving feedback token consumption relative to prior methods that use full feedback. These results establish DiSH as a strong baseline for solver refinement under harness control.
Related
- DAOCP: a dual active set solver for optimal control problems
- Adaptive Grid Refinement for Optimal Feedback Control
- ElastiQP: An Always-Feasible QP Solver for Constrained Robot Control
- Mutual Information Optimal Density Control of Linear Systems and Generalized Schrödinger Bridges with Reference Refinement
- CuPIQP: A Batched and Differentiable QP Solver on GPUs for Robot Learning and Control
BibTeX
@misc{cao2026harnessing,
title = {{Harnessing LLMs with Direction-Scope Control for Combinatorial Solver Refinement}},
author = {Yili Cao and Anqi Li and Haoran Li and Yongtang Shi and Zicheng Zhang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.05221},
eprint = {2610.05221},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.05221},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Harnessing LLMs with Direction-Scope Control for Combinatorial Solver Refinement
AU - Yili Cao
AU - Anqi Li
AU - Haoran Li
AU - Yongtang Shi
AU - Zicheng Zhang
PY - 2026
DA - 2026/10/04
PB - arXiv
DO - 10.48550/arxiv.2610.05221
UR - https://arxiv.org/abs/2610.05221
AB - Harnesses are increasingly unlocking the potential of large language model (LLM) agents by making their actions more controllable. In automatic heuristic design (AHD) for combinatorial optimization (CO), LLMs iteratively revise solvers using execution and verification feedback. However, these revisions often break previously validated solver behavior, making refinement unstable and costly. We characterize this pattern as feasibility oscillation and document it across CO benchmarks, AHD methods, and backbone LLMs. Tracing this oscillation, our analysis reveals that feasibility degradation often follows large solver revisions, while substantial feedback redundancy makes it harder to identify which failures deserve attention. Motivated by these findings, we introduce DiSH, a harness that jointly controls revision direction and scope. It selects representative failure evidence to guide each update and progressively narrows the revision scope as feasible behavior accumulates. Across nine CO benchmarks, DiSH consistently improves QYI scores and reduces feasibility oscillation, while roughly halving feedback token consumption relative to prior methods that use full feedback. These results establish DiSH as a strong baseline for solver refinement under harness control.
ER -