Meta's Agentic Feature Optimization Lifts Ad Ranking Models 0.20% in Online A/B Tests
_reachsumit · x · 2026-11-20
Meta published an end-to-end agentic system that optimizes feature selection across a fleet of ads ranking models, under fleet-wide constraints (feature pipeline cost) and per-model constraints (serving QPS), drawing from a shared pool of tens of thousands of features.
- LLM agents propose hypotheses and generate/shortlist cross features; deterministic tools run experiments and verify. Multiple importance methods rank features; feature blending uses Pareto fronts for multi-task models
- An LLM Researcher agent tunes selection in 10–25 asynchronous trials under resource caps; a trimming step keeps pipeline cost in budget
- On production models: normalized entropy improved 0.31% on average, pipeline cost cut 2.3% with no regression; launched models improved a primary online metric by 0.20% in A/B tests
More from coding & agent
- OtterKit turns local MCP servers into public URLs with one command for Claude web testing — Conscious-Return-183 · 2026-11-20
- Andrew Carr blocks Antigravity's mobile PWA — agentic coding in bed is too addictive — andrew_n_carr · 2026-11-20
- AutoRAGBench: LLM Agents Build RAG Systems From Scratch, GPT-6.1 Scores 56.1 Beating Human SOTA — _reachsumit · 2026-11-20
- Procedural Gojira Built with Gemini and Three.js TSL — majidmanzarpour · 2026-11-20
- User finds their AI agent quietly ate 54GB of disk with unused git worktrees — itsOmSarraf_ · 2026-11-20
- Harness engineering gap: Provider harness 39.0% vs Mini-SWE 12.8% on same model — andrew_n_carr · 2026-11-20