ACL Anthology paper page 2025.emnlp-main.334 - TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text

PUBLISHER'S OWN DEPOSIT for TurboRAG, and the record that CHANGES the wave-4 answer. Wave 4 recorded papers/w4p-turborag-precomputed-chunk-kv as arxiv-preprint on an explicitly negative search: the only two OpenReview records for the title were 'ICLR 2025 Conference Withdrawn Submission' and 'ACL ARR 2025 May Submission', and an ARR batch was correctly refused as a venue. The ARR submission was in fact a real submission that was later ACCEPTED and published, which is the whole point of an ARR batch - a negative ARR search is not a negative venue finding, only a statement about the review pipeline's state at the time. Anthology ID 2025.emnlp-main.334; volume 'Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing'; pages 6588-6601; Suzhou, China; November 2025; Venue EMNLP; SIG SIGDAT; publisher ACL; DOI 10.18653/v1/2025.emnlp-main.334; bibkey lu-etal-2025-turborag. IDENTIFICATION IS EXACT ON BOTH AXES, which is the standard this repo requires: the title is character-for-character the arXiv title, and the FIVE author names on the Anthology page (Songshuo Lu, Hua Wang, Yutian Rong, Zhi Chen, Yaohua Tang) are the same five, in the same order, that the arXiv API returns for 2410.07590. No author-overlap inference was needed here. THE CAMERA-READY ABSTRACT DIFFERS FROM THE ARXIV ABSTRACT and the difference is a mechanism claim, not wording: the arXiv v1 abstract says the system 'fine-tune[s] a pretrained language model to maintain model accuracy', while the Anthology abstract describes the method as a hybrid offline-online paradigm that pre-computes chunk-level KV caches, stitches them at inference with INDEPENDENT-ATTENTION and REORDERED-RoPE techniques, and 'preserves answer quality WITHOUT CHANGING THE MODEL ARCHITECTURE'. The 9.4x max / 8.6x mean TTFT reduction against conventional RAG is identical in both. A reader holding the arXiv PDF alone will get the weaker mechanism story and will not see the two named techniques.