PyOhio 2026
30 Minute Talk
LLMs break text into chunks called tokens before they do anything with it. The way that splitting happens can quietly hurt your results, and most people never notice. In this talk, I'll cover two simple techniques that can improve LLM output quality without any training or fine-tuning. The first, token healing, fixes a common problem where the boundary between your prompt and the model's response lands in an awkward spot, confusing the model. The second is choosing a better splitting pattern for how text gets chunked into tokens in the first place. I'll show both techniques in Python, explain why they work, and share benchmark results across several open-weight LLMs.
