.png){kind=icon}
Return
Parametric Knowledge Cannot Overcome Instance Bias in LLM-Based Automated Algorithm Design
.png){kind=icon}
PRE
AI
DOI：10.1109/tevc.2026.3735709.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Leveraging the advantage of the internal parametric knowledge, large language model (LLM)-based automated algorithm design (AAD) is advancing rapidly. Most existing methods combine LLMs with evolutionary frameworks, where candidate algorithms are evaluated on problem instances and improved accordingly. Consequently, both the internal parametric knowledge of LLMs and the external guidance from evaluation instances jointly shape the final algorithm. However, it remains unclear how the internal knowledge of LLMs interacts with external instance guidance. In this work, we empirically investigate the impact of instances on LLM-based AAD across both seen and unseen problems. For seen cases, LLMs possess sufficient parametric knowledge to directly generate competitive algorithms, making the guidance from instances largely negligible. While for unseen cases where LLMs lack adequate prior knowledge, LLM-based AAD relies more on external instance guidance, and biased instance distributions significantly degrade the performance. To further investigate this dependency, we propose an evolutionary method to automatically generate adversarial instances. The results show that, on unseen problems, the vulnerability to instance bias widely exists in instance-incorporated LLM-based AAD methods. Our findings demonstrate that while the rich parametric knowledge of LLMs mitigates instance bias in familiar tasks, overcoming instance dependency is crucial for unseen problems.
Keywords:
automated algorithm design
large language models
instances
Journal
IF:
12
Papers:
1.9K
Citations:
2.4W
Organization
Cited Papers
No cited papers available