ASPIRE, submitted on 16 Sep 2026, proposes non-synchronized batched self-speculative decoding for long-context LLM inference
arXiv submission 2609.17943, submitted on 16 Sep 2026, proposes ASPIRE, a non synchronized batched self speculative decoding framework for long context LLM inf…
Published on MyPrivateClaw
Sep 18, 2026, 4:25 PM UTC
Coverage date
Sep 17, 2026
Last updated
Sep 18, 2026, 4:25 PM UTC
News summary
The preprint describes ASPIRE as a non synchronized batched self speculative decoding framework for long context LLM inference. arXiv submission 2609.17943 was submitted on 16 Sep 2026, and the framework combines a unified mixed forward that drafts and verifies requests in the same batched forward pass, an online speculation scheduler with a batch aware cost model, and an intra draft refresh layer that runs full attention at a single designated layer to update the sparse KV cache during drafting. The authors report that across three models and five reasoning and long context benchmarks ASPIRE achieves 1.70 4.58× decoding throughput over autoregressive baselines, and improves average speedup by approximately 27% over the strongest prior self speculative baselines. The reported figures are the authors' own measurements on three models across five benchmarks, so per request performance wil…