Method Drift›Speculative decoding
SpecReason
SpecReason: Fast and Accurate Inference-Time Compute via Speculative ReasoningSpeculative decoding · first seen Apr 10, 2025
superseded — cited as a baseline and beaten by newer methods
1 papers critique it · 2 beat it on benchmarks
What papers say
Verbatim critique sentences, each from a paper that cites SpecReason as a baseline.
“SpecReason, which exhibited more noticeable accuracy reductions on several tasks (e.g., dropping from $91.8\%$ to $85.9\%$ on GSM8K with Deepseek-R1, a $6\%$ decrease)”
— Scaling Speculative Decoding with Lookahead Reasoning
Beaten on benchmarks
Head-to-head results where a newer method reports beating SpecReason. Values are copied from the source paper's tables — verify against the cited paper.
- Scaling Speculative Decoding with Lookahead Reasoning
LR(ours) beats SpecReason · Accuracy [Draft: Deepseek-R1-Distill 1.5B / Target: Deepseek-R1-Distill 32B]
92.8 vs 85.9
- Scaling Speculative Decoding with Lookahead Reasoning
LR(ours) beats SpecReason · Accuracy [Draft: Qwen3 1.5B / Target: Qwen3 32B]
96.4 vs 94.5
- Beyond Tokens: Semantic-Aware Speculative Decoding for Efficient Inference by Probing Internal States
SemanticSpec beats SpecReason · Pass@1 [Draft: DeepseekR1-1.5B Target: DeepseekR1-32B Math-500]
0.910 vs 0.840
- Beyond Tokens: Semantic-Aware Speculative Decoding for Efficient Inference by Probing Internal States
SemanticSpec beats SpecReason · Pass@1 [Draft: DeepseekR1-1.5B Target: QwQ-32B Math-500]
0.9217 vs 0.770
Newer alternatives
Recent methods in the same sub-problem, not yet superseded in the knowledge base.