Deft model fixes LLM formulaic writing with Distribution Fine-Tuning
HamelHusain · x · 2026-08-20
Deft is a new model and lab focused on improving AI writing. Its research paper highlights that standard SFT leads to formulaic outputs, overusing certain tokens or phrases ("slop signs").
Deft introduces Distribution Fine-Tuning (DFT), a post-training step that aligns the model's output distribution more closely with the training data. Experiments show DFT improves MMD by 49% and JMQ by 63%. The model significantly boosts creativity, coherence, and clarity, achieving a 100% human-written score on the Pangram detector.
More from Research
- Five pharma giants federally train OpenFold3 on 20,000 private structures, +11 pts — jiqizhixin · 2026-10-16
- SWAN: a semantic-level text watermark that survives paraphrasing — burkov · 2026-10-16
- Interp researchers push back: 'networks are already clean' claim rests on dropout artifact — aryaman2020 · 2026-10-16
- LayerTrim: training-free pruning keeps reasoning LLMs intact on AIME — SeoulNatlUniv · 2026-10-16
- Simular presents neuro-symbolic computer use agents at BayLearn — xwang_lk · 2026-10-16
- Google's Hartmut Neven disputes chip test, outlines plan to link a brain to a quantum computer — Chris_Armstrong · 2026-10-16