Beyond a scalar: distributional serving interfaces let multiple task heads reuse watch-time distributions
_reachsumit · x · 2026-09-24
A paper proposes the Distributional Serving Interface (DSI) for watch-time prediction in short-video feeds.
- Problem: existing methods expose only a single watch-time estimate at serving time, giving downstream models no probabilities for completion, overplay, or other event regions.
- Method: DSI comprises a distribution provider, a compact low-dimensional summary, and lightweight per-task readouts. The provider learns a joint distribution over four watch states and their event times; duration-based rules remove incompatible combinations while a restoration loss preserves second-level accuracy. The summary outputs event probabilities, duration-relative time scales, and uncertainty statistics; after training, the provider is frozen and value/ranking readouts are trained to reuse the summary.
- Results: lowest MAE on KuaiRec, KuaiRand-1K, and WeChat21, beating the strongest of nine baselines by 1.9%–8.5%, with best XAUC on two datasets.
More from Research
- OpenAI publishes internal pretraining perplexity curve, proposes perplexity as shared safety bar — merettm · 2026-11-23
- AI math swarm proves a 962-variable quartic is minimal; 4,973-variable version spans 1.62 MB of text — gleech · 2026-11-23
- Even OpenAI's quasi-Riemann result already spawned vast expository work, argues AI-proof skeptic — gleech · 2026-11-23
- CMU's ExploitBench ranks LLM cyber agents on a capability ladder, not pass/fail — goyal__pramod · 2026-11-23
- Harvard paper: higher SWE-Bench scores didn't translate into more merged pull requests — rohanpaul_ai · 2026-11-23
- Princeton's SSoN Benchmark: Best Vision Model Scores 39.7% vs Human 83.3% on Spatial Consistency — jiqizhixin · 2026-11-23