Cactus: Accelerating Auto-Regressive Decoding with Constrained Acceptance Speculative Sampling
Open paper
Benchmark
Model
Rank
Results
arithmetic-reasoning-on-gsm8k
Cactus (δ=1.0; Qwen 3 32B verifier + Qwen 3 1.7B drafter)
#31
Accuracy: 94.4
Rank counts only results with a code link.