Google's FLVM fixes measurement bias in short-video recommendation signals
_reachsumit · x · 2026-09-29
Google proposes the Factorized Latent Value Model (FLVM) to address measurement bias in recommender feedback: the same observed behavior can stem from genuine enjoyment, passive consumption, or inattention, and in short-form video, watch-based signals are strongly confounded by video duration — identical watch time implies different preference for different lengths, while ratio metrics systematically favor short videos. Optimizing raw engagement thus amplifies measurement artifacts rather than user value.
FLVM treats observed behaviors as noisy measurements of a low-dimensional factorized latent value state with structured output heads for heterogeneous feedback. A restricted baseline path absorbs predictable variation from confounders (duration, user propensity, session context), while a routed latent path estimates preference-relevant value advantage, yielding a latent value score for ranking.
More from Research
- Amazon's GEB grounds entity biographies for long-video memory, hits 72% on EgoLifeQA — amazon · 2026-09-30
- NVIDIA's LongLive-Plug: Distill Once, Deploy Training-Free Across 54 Downstream Video Models — nvidia · 2026-09-30
- CrossBFM Distills a Shared Behavior Space Across Humanoid Robots in Under One GPU-Hour — Tan-Dzung Do · 2026-09-30
- AnisoWM: anisotropic representations improve planning in JEPA world models — SeoulNatlUniv · 2026-09-30
- Meituan's SAKI: Coupling-Routed Teacher Supervision Speeds Up On-Policy Distillation 4.22x — meituan · 2026-09-30
- Real2Gym Turns Videos Into Interactive Robot Gyms, Beating GPT-6 Direct Mode by 33% on Real Robots — Shanghai-AI-Laboratory · 2026-09-30