SPEED-Bench Speculative Decoding Replication

Replicating paper OpenReview: Rl2uQlCoQX (arXiv: 2604.09557)

Claim 4: Synthetic Overestimation

Synthetic Input Speedup: 1.28x (35.17 vs 27.46 TPS)

Natural Input Speedup: 0.78x (slowdown! 23.19 vs 29.8 TPS)

Synthetic Speedup Overestimation Gap: 64.1%

Claim 5: Optimal Draft Length

Draft length 1 yields highest TPS of 15.31, with performance decreasing monotonically to 9.69 at length 5 due to acceptance rate decay on natural text.