Early-Signal Detector

method · first seen 4 Oct 2026

speculative decoding

Also: SD, speculative drafting, parallel action decoding

Inference acceleration where a drafter proposes multiple tokens that the target model verifies in parallel.

Speculative decoding speeds up autoregressive generation by having a cheap drafter propose several tokens that the target model checks in one forward pass, keeping the output distribution unchanged. The idea dates to 2022-2023 and is now standard in serving stacks. The sources show ongoing refinement rather than a new idea: papers adapt it to diffusion language models, looped transformers, long-context attention dilution, tree verification scheduling, session-local drafters, streaming speech recognition and robot action decoding. On the code side, an official EAGLE-4 release and several local inference engines for CPUs and Apple Silicon list it as a supported feature.

It is already a default technique for cutting LLM latency, so the activity here reflects how widely it is used rather than early adoption. The most interesting threads are its extension to non-autoregressive and looped architectures and to domains outside text generation.

Analyst note (established), 17 Nov 2026.

Mentions per week

02619 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 36 on 14 Nov 2026 · 120 mentions in 28 days · 118 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration -0.24

Earliest sources

Latest sources

Mentioned together with

mixture of expertsdiffusion language modelmulti-token predictionquantizationKV cachevLLMtree-based speculative decodingblock diffusion draftingknowledge distillationsparse attention