Researcher proposes DPPs to boost output diversity: reward quality and dispersion over sets
nanjiang_cs · x · 2026-11-21
Addressing the problem of homogeneous model outputs, the author proposes bringing back determinantal point processes (DPPs).
- A DPP is a distribution over sets of outputs, not individual strings: a set is more likely when its members are high-quality and their embeddings point in different directions.
- This jointly rewards per-item quality and within-set diversity, making it a natural fit for combating output homogenization.
- Commenter @nanjiangcs notes he immediately thought of DPPs from the main tweet's visual, having overlapped with DPP researcher Alex Kulesza at Michigan.
More from Research
- CT-Bench, a benchmark for LLM agent cooperation via self-negotiated contracts, accepted at COLM 2026 — maxhkw · 2026-11-21
- Frank Nielsen demos geodesic right-angle triangles in dually flat spaces — FrnkNlsn · 2026-11-21
- Google's Pro² Agent hits 76% medal rate on MLE-Bench, a new record — dair_ai · 2026-11-21
- Phillip Isola: language is a camera — what's its color space and resolution? — phillip_isola · 2026-11-21
- Alessandro Masullo announces a purely human proof of the Riemann hypothesis — davilagrau · 2026-11-21
- Open-source PlaceBench tests 6 voice assistants on 150 real-place questions: Dots 92, Ask Maps 49 — whowasntright · 2026-11-21