Kuaishou Team Proposes GR4AD-X, Upgrading Generative Ad Recommendation with Multi-Path Semantic IDs
_reachsumit · x · 2026-10-30
Kuaishou researchers published an arXiv paper introducing GR4AD-X, an upgrade to their industrial generative recommendation (GR) framework GR4AD for large-scale advertising.
GR4AD reformulates item retrieval as autoregressive generation of discrete semantic identifiers (SIDs), but deployment revealed three limitations: deterministic SID assignment is sensitive to semantic boundaries; conventional backbone scaling yields diminishing revenue gains under strict serving constraints; and sequence-level value optimization gives limited credit assignment to intermediate SID decisions, causing likelihood-based beam search to prune high-value candidates too early.
GR4AD-X addresses these with:
- Multi-Path SID assignment for more reliable semantic ID construction;
- Multi-dimensional capacity scaling for a better effectiveness-efficiency trade-off under serving constraints;
- Unified Value-Guided Generation and Discrimination, with value-guided beam pruning and reranking to preserve potentially high-value ad candidates.
More from Venture
- Marc Lou: 2 Years to Earn $1, 8 Months to 1K Followers — Real Indie Startup Timeline — marclou · 2026-10-30
- Does Louis Vuitton Even Need AI SEO? Big Brands and Generative Search Visibility — gaganghotra_ · 2026-10-30
- Claude + PeekPanda MCP turns natural language into competitor ad intelligence across 5M+ ads — fekdaoui · 2026-10-30
- Guidewire buys agentic underwriting startup Hyperexponential for $1.7B in cash — emmanuelvivier · 2026-10-30
- Amazon invests another $5B in Anthropic, securing $100B+ decade-long AWS commitment — emmanuelvivier · 2026-10-30
- Amazon and Intel earnings misses deepen AI bubble fears, tech stocks slide — nordicinst · 2026-10-30