RING internalizes retrieval into LLMs and runs 3–19× faster than RAG

_reachsumit · x · 2026-08-04

• The paper proposes RING (Retrieval-Internalized Generation), a new way to inject large-scale knowledge directly into model parameters.
• Instead of using an external retriever at serving time, RING learns a parametric search policy over an internal memory via reinforcement learning.
• Training happens in three stages: continued pretraining to inject new corpora, supervised fine-tuning for a search-then-answer pattern, and RL with hierarchical rewards.
• The authors also introduce News-2025, a benchmark built from news published after the base model’s pretraining cutoff, to evaluate genuinely new knowledge without leakage.
• Reportedly, RING can match or exceed RAG while running 3–19× faster.

Original post →

More from Research

• Xiaomi's CodeMidas turns source code into RL environments, doubling DeepSWE to 21.7% — maier_ak · 2026-09-22
• HeyGen and Kaggle launch Code2Video Bench for motion graphics code generation — MeganRisdal · 2026-09-22
• Alibaba's Qwen team launches RecreationBench to test hybrid computer-use agents by app recreation — TianbaoX · 2026-09-22
• CVPR 2025 organizer: 90%+ of first authors had one submission, so authorship caps won't cut volume — CSProfKGD · 2026-09-22
• Toby Ord on swarm scaling: 10,000-agent run cost ~$20M, solved Navier-Stokes in 88 hours — tobyordoxford · 2026-09-22
• Toby Ord: Agent Swarm Estimates Put Intelligence Explosion Parameter λ at 0.5-0.6 — tobyordoxford · 2026-09-22

Research channel →