ADSD
Adversarial Prompts for Acceptance Collapse in Speculative Decoding
Suffix search, inference and evaluation for prompt-induced acceptance collapse. Run benign and fixed-suffix speculative decoding with Qwen2.5-14B and Qwen2.5-0.5B on GSM8K.
Recorded GSM8K results
| Metric | Benign | ADSD |
|---|---|---|
| Mean latency (s) | 26.38 | 42.77 |
| Block efficiency | 6.021 | 3.585 |
| Accuracy | 82.13% | 87.07% |
62.11% mean latency increase over 263 questions.
Suffix search
Using the same prepared environment and runtime, run bash scripts/run_search.sh. The fixed configuration performs five sweeps over a five-token suffix; the completed search recovered the default suffix with identical text and token IDs.
Project status
- ✓ GSM8K data preparation and verification.
- ✓ Runtime preparation and benign inference.
- ✓ ADSD inference with the provided suffix.
- ✓ Paired benign/ADSD evaluation and metric reporting.
- ✓ CPU tests for runtime preparation, launching and metrics.
- ✓ Discrete suffix search with the provided configuration.
Citation
@misc{wang2026adversarialpromptsacceptancecollapse,
title = {Adversarial Prompts for Acceptance Collapse in Speculative Decoding},
author = {Run Wang and Chaoyi Zhou and Xi Liu and Yi Zhu and Amir Salarpour and Pedram MohajerAnsari and Zhi-Qi Cheng and Feng Luo and Siyu Huang and Mert D. Pes{\'e}},
year = {2026},
eprint = {2607.21804},
archivePrefix = {arXiv},
primaryClass = {cs.CR},
url = {https://arxiv.org/abs/2607.21804}
}