Papers by Zhenbo Xu
Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work has shown that safety behaviors are governed by low-rank structures . Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks . |
| Approach: | They propose a safety alignment system that disentangles safety-relevant directions into monosemantic features and constructs an interpretable safety subspace from SAE directions. |
| Outcome: | Empirically, the proposed model achieves 99.6% safety rates across multiple model families and scales . low-rank Adaptation consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks compared with previous methods . |
SCVQ: Sparse-Compensated Vector Quantization for Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing vector quantization methods incur inference overhead due to massive codebook storage and intensive index lookups. |
| Approach: | They propose a framework for vector quantization that incorporates a salience-aware weighted K-means clustering scheme with symmetry constraints to reduce codebook size and indexing costs. |
| Outcome: | The proposed framework achieves a perplexity of 5.78 on WikiText-2 for LLaMA-2-7B at 2-bit quantization while delivering a 1.4 speedup over existing baselines. |