How to de-slop AI writing: score multiple generations against each other, skip SFT
ivan_bezdomny · x · 2026-08-25
ivanbezdomny shares hands-on lessons from building non-slop news writing: he narrows to one task, and since SFT doesn't really work, he scores multiple generations of the same news against each other. He also counts word/phrase frequencies from tweets to compare language distributions, admitting this resource is underused. General-purpose models do a poor job removing slop beyond banning certain vocabulary, he says; a longer post is coming.
More from Research
- Nonprofit Sophron Research Launches to Develop AI Model Evaluations — ryan_t_lowe · 2026-08-26
- NeurIPS workshop CFP: Interpreting Agent Behavior — mdredze · 2026-08-26
- Reasoning Models Outperform via Higher Recovery Rates, Not Just "More Thinking" — Jeande_d · 2026-08-26
- Study Finds Reasoning Models' Amplified Behaviors Weakly Linked to Correctness — Jeande_d · 2026-08-26
- HOMIE Gen2 launched: 360° vision and spatial audio rig targets the Experience Scaling Law for Physical AI — liuziwei7 · 2026-08-26
- Research Finds Subliminal Learning More Powerful, Applies to Standard Finetuning and SGD — johnhewtt · 2026-08-26