Research
NAIRL’s Professor Seong-Whan Lee of Korea University to Present at NeurIPS 2026
A research team led by Professor Seong-Whan Lee of the Department of Artificial Intelligence at Korea University, affiliated with the National AI Research Lab (NAIRL), has developed GDPruner, a method for shrinking large language models that keeps their mathematical reasoning and code generation ability largely intact. The study, “Generation-Drift-Guided Block Pruning for Large Language Models,” will be presented at NeurIPS 2026.
Large language models deliver strong performance, but their size drives up compute and operating costs. One remedy is block pruning, which removes some of a model’s computational blocks outright. Existing pruning methods, however, tend to degrade sharply on complex tasks such as math reasoning and code generation.
The team’s insight is that small prediction errors compound as a model generates text. GDPruner has the original model generate its own sample text, then measures how far a pruned candidate’s predictions drift from the original on that text, a quantity the team calls generation drift. It focuses on the later positions of each sequence, where accumulated drift is most visible, and evaluates the combined effect of removing several blocks at once, widening its search when drift rises sharply. The whole process runs without any external calibration data.
Across multiple language models and benchmarks, GDPruner outperformed existing pruning methods under most conditions and held up notably well on math and code tasks even at high pruning ratios. The pruned models generated text faster than the originals under the same measurement setup, and the method also applied to large-scale and mixture-of-experts (MoE) models.
The work points to a way of cutting inference costs by keeping only the computation that matters for generation, which could help organizations run generative AI on limited resources. NAIRL will continue to introduce research that advances the efficiency and reliability of AI to the broader domestic and international AI community.