September 2026 · Draft
Quantifying and Fixing “Idea Slop”
A working research draft on measuring repeated ideas in LLM outputs, pairwise concept overlap, distinct relevant ideas, and an experimental Deft Brainstorm Mode.
Research
May 2026
Fixing LLM Writing with Distribution Fine-Tuning
SFT-trained models fail to match their training distribution at any sampler setting. DFT, a new post-training step, closes the gap: 49% better MMD, 63% better judge quality, and output with no overused-token fingerprint beyond human-to-human variation.