Apple's CoGR Co-Evolves Query and Item Generators with RL, Boosts Retrieval F1 by Up to 36.1%
antoine_chaffin · x · 2026-09-06
Apple researchers propose CoGR (arXiv:2609.00638), a generative retrieval framework that trains LLM keyword generators for both the query and item sides instead of only augmenting queries. A two-stage pipeline — SFT to align the keyword space, then alternating GRPO-based RL — co-evolves both sides against each other's frozen inverted index, with the item side rewarded via counterfactual marginal retrieval-F1 gains. Keywords match through a standard inverted index, preserving existing keyword infrastructure. Across 10 sparse, dense, and generative baselines, CoGR achieves the best results, with +10.9% / +36.1% F1 over the strongest baselines on an internal App Store marketplace dataset.
More from Research
- Frank Nielsen maintains a comprehensive research hub on Bregman divergences for ML — FrnkNlsn · 2026-09-06
- Manifold market at 36% odds that an MF group proof will appear this month — Sauers_ · 2026-09-06
- Benchmark proposal: testing AI reasoning across classic onemorelevel.com Flash games — BeingBudget8847 · 2026-09-06
- Model auto-optimizes ALOHA robot hardware design, validated on real hardware — SongShuran · 2026-09-06
- Rumor: Anthropic Solved a Millennium Prize Problem, Terence Tao Responds — littmath · 2026-09-06
- Harvard-led arXiv paper models LLM adoption as a 'cognitive virus' with dependence tipping points — rohanpaul_ai · 2026-09-06