LLMs autonomously evolve better video encoding heuristics for VVenC

A new paper from researchers Liqiang He, Yingwen Zhang, Riyu Lu, Meng Wang, and Shiqi Wang explores using large language models (LLMs) to automatically design heuristics for frame-level quantization parameter (QP) adaptation in video encoding. The challenge: optimal QP allocation is content-dependent, and classical Lagrangian rate-distortion optimization (RDO) often suffers from inaccurate multiplier settings. Fixed-QP schemes, while simple, are content-agnostic. The team built a closed-loop evolutionary framework where an LLM iteratively proposes RDO heuristics as algorithmic ideas with executable code. Each candidate is evaluated directly by encoding with the Fraunhofer Versatile Video Encoder (VVenC), using a scoring function that compares QP choices based on past frame statistics and current candidates.

Experimental results across multiple test sets show the evolved heuristic achieves promising rate-distortion improvements over both the fixed-QP scheme and the Lagrangian baseline. Notably, the LLM autonomously discovered an adaptive heuristic that penalizes QP fluctuations via entropy-based terms, a novel insight for RDO algorithm design. This approach could reduce the manual effort in tuning video codec parameters and lead to more efficient compression—especially valuable for streaming services and high-resolution content. The paper, submitted to arXiv on June 18, 2026, is available at arXiv:2606.20847.