Reproduce “Secure Speculative Decoding for Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yichi Zhang, Zhiqi Wang, Neil Gong, Yuchen Yang

SecureSD addresses security losses from lossy speculative decoding by applying stricter target-model verification to early draft tokens and relaxing it later. For Qwen3-0.6B/8B, security performance retention averaged 0.744, versus 0.905 on utility benchmarks, across six methods and sweeps. SecureSD raised Jailbreak-SD retention by 0.270, with a 0.026 speedup change, but does not eliminate target-model vulnerabilities or establish universal robustness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.