Draft Models for Speculative Decoding
Architectures aimed at improving speculative token acceptance length.
- TypeResearch
- PeriodJan 2026 — Present
- AffiliationTencent · Research Internship
- StatusOngoing
Overview
Research into draft-model architecture designs that aim to increase the accepted token length in speculative decoding.
- Research focus: draft-model architectures.
- Objective: improve accepted token length in speculative decoding.
- This project is ongoing; detailed results are not published here.
This public overview does not include private code, unpublished measurements, or confidential details from industry collaborations.