Reproduce “PSRT: Accelerating LRM-based Guard Models via Prefilled Safe Reasoning Traces” 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
Jiawei Zhao, Yuang Qi, Weiming Zhang, Nenghai Yu, Kejiang Chen
Researchers from the University of Science and Technology of China developed Prefilled Safe Reasoning Traces (PSRT) to accelerate Large Reasoning Model (LRM)-based guard models, reducing inference overhead by eliminating explicit reasoning trace generation while largely preserving detection performance, with a minor average F1 score drop of 0.015 across various models and datasets.
Key insights
- The complex reasoning capabilities of LRMs can be condensed into optimized, prefilled virtual embeddings rather than requiring explicit, token-by-token generation for classification tasks.
- The initial Supervised Fine-Tuning (SFT) stage and the average initialization of the safe reasoning trace are critical for the method's effectiveness and performance.
- Different base LRM architectures (e.g., Llama vs. Qwen) exhibit varied intrinsic cognitive behaviors when adapting to prefilled traces, leading to contrasting performance shifts on jailbreak datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.