Recommender systems are critical for user engagement, but optimizing them is labor-intensive. Kuaishou's RecHarness framework tackles this by integrating bandit algorithms with LLM-based agents. The core idea is to treat each candidate modification to the recommendation model as an 'arm' in a multi-armed bandit problem. The system evaluates which modifications are most promising and allocates resources accordingly, while LLM agents generate and refine these candidates. This approach reduces the trial-and-error cost of traditional MLE workflows and allows the system to continuously improve itself. RecHarness addresses a key challenge: efficiently routing limited testing resources to the most impactful changes. Early results suggest significant gains in optimization efficiency. For engineers building large-scale recommendation systems, this offers a glimpse into the future of automated ML operations.
Kuaishou's RecHarness framework combines bandit algorithms with LLM agents to automate recommender system optimization, enabling self-evolving models.