Scholar Feed

Speculative decoding

RSD

Reward-Guided Speculative Decoding for Efficient LLM Reasoning

Superseded baseline#17 of 151 most-superseded · first seen Jan 31, 2025

Superseded — cited as a baseline and beaten by newer methods

3 papers critique it · 2 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites RSD as a baseline.

Existing approximation algorithms such as K-SEQ sun2023spectr, MSS miao2024specinfer, RSD jeon2024recursive, and LP and vocabulary truncation khisti2025multi cannot guarantee near-optimal speedups: K-SEQ has a $(1-1/e)$-approximation guarantee, but the others have no formal guarantees.
— Global Resolution: Optimal Multi-Draft Speculative Sampling via Convex Minimization
“However, reward models are neural networks that require additional forward passes, substantially increasing the computational cost of SD.”
— Entropy-Aware Speculative Decoding Toward Improved LLM Reasoning
“Although effective in improving reliability, it incurs substantial drawbacks. First, reliance on external verifiers significantly increases latency and compute overhead. Second, pre-trained reward models are often specialized to specific tasks, making them difficult to generalize across diverse reasoning tasks.”
— From Tokens to Steps: Verification-Aware Speculative Decoding for Efficient Multi-Step Reasoning

Beaten on benchmarks

Head-to-head results where a newer method reports beating RSD. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.