How Much Dense Attention is Necessary? Oracle-Guided Sparse Prefill for Full/GQA Layers in Hybrid Long-Context Models (arXiv:2606.07703) - the authors' own report behind the SOTA2/WizWand TTFT leaderboard

FETCHED AND READ 2026-10-08: arXiv abs page for 2606.07703, plus the SOTA2 paper page https://www.sota2.com/research/paper/6a279e1290966ade919bf6de that carries the same abstract and the abstract text quoted below. 26 pages, 2 figures, 11 tables, cs.LG/cs.AI/cs.CL, submitted by Zhen Zhang. THIS IS THE SAME PAPER the leaderboard page bench-sota2-wizwand-vllm-ttft-32k-256k points at - the sota2 page's own 'Related benchmarks' table lists the TTFT 32K-256K task at 2.944 s for Qwen3.5-9B (distilled indexer), and the paper page is the document the leaderboard row links to. So the Tier-D aggregator URL and this Tier-C author preprint are the same measurement, and citing both is what lets a reader audit the leaderboard number against its source. Abstract (verbatim, the parts that bear on the record): 'Preliminary single-card TTFT measurements show distilled-indexer sparse serving speedups of 1.71x for Qwen3.5-0.8B on NPU and 1.93x for Qwen3.5-9B on GPU against its dense FlashAttention-2 baseline.' and 'a Qwen3.5-9B RULER-style sweep from 4K to 100K stays within 0.48 points'. Also: 'We introduce an attention-mass top-k oracle ... The oracle is a diagnostic reference, not a deployable accelerator'. AND THE HONEST LIMIT, taken from the paper's own words because it bounds how much this raises the record: the abstract labels the TTFT numbers PRELIMINARY and explicitly says the first release 'separates oracle feasibility, distilled-indexer quality, and runtime headroom, leaving a fully matched quality-latency frontier to future work'. The paper does not state the GPU model ('GPU 80G' is a leaderboard label, not a part), and the leaderboard does not state concurrency or cache state. So the PUBLISHING PARTY is now a primary author-preprint source (Tier C by the brief: a preprint by its own authors), but the hardware identity and the load conditions remain unpublished, and those cannot be invented.