Medusa: multihead drafts with tree attention

What changed

Added future-token prediction heads to a language-model backbone, with tree-attention verification, frozen-backbone or joint training and a typical-acceptance option.

Scientific scope

Added future-token prediction heads to a language-model backbone, with tree-attention verification, frozen-backbone or joint training and a typical-acceptance option. Inherits Stern-style multihead prediction and speculative verification. A frozen backbone alone does not ensure exact output-distribution preservation: the proposed typical-acceptance rule deliberately relaxes distribution matching; Medusa2 also updates the backbone. Empirical generation quality is not equality of distributions or universal acceleration.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits Stern-style multihead prediction and speculative verification. A frozen backbone alone does not ensure exact output-distribution preservation: the proposed typical-acceptance rule deliberately relaxes distribution matching; Medusa2 also updates the backbone. Empirical generation quality is not equality of distributions or universal acceleration.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction