Apple's CoGR: Co-Evolving Generative Retriever with Reinforcement Learning
_reachsumit · x · 2026-09-02
Apple researchers propose CoGR, a framework training LLMs to directly construct retrieval representations (keywords) for both query and item sides, rather than just augmenting queries. It matches keywords via inverted index for infrastructure compatibility. The training involves SFT and co-evolving RL, alternately optimizing both sides to maximize retrieval F1. Experiments show CoGR outperforms 10 baselines.
More from Research
- HydroGym RL platform for fluid dynamics published in Nature with 60+ environments — ricardovinuesa · 2026-09-03
- Paper proposes agents that outlive their model, harness, and host by splitting identity from plumbing — omarsar0 · 2026-09-03
- Stanford paper: steering directions causally shift context-vs-memory choice, but barely transfer across tasks — niloofar_mire · 2026-09-03
- Meta's Muse Spark jumps 1.1 to 1.3 in 55 days, photo-to-3D simulation at $0.60 — alexandr_wang · 2026-09-03
- Developer once tried building AI benchmark from Puzzlescript, similar to ARC-AGI-3 — Darpinian · 2026-09-03
- Do induction heads and attention sinks count? Debate over interpretability's missed milestone — aryaman2020 · 2026-09-03