Papers by Ruqi Zhang
CoT-UQ: Improving Response-wise Uncertainty Quantification in LLMs with Chain-of-Thought (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing uncertainty quantification methods for Large language models are primarily prompt-wise rather than response-wise, which leads to inefficiency. |
| Approach: | They propose a new approach to quantify response-wise uncertainty by integrating LLMs’ inherent reasoning capabilities through Chain-of-Thought (CoT) into the UQ process. |
| Outcome: | The proposed framework outperforms existing uncertainty quantification methods and achieves an average improvement of 5.9% AUROC compared to existing methods. |
Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent research has focused on test-time alignment, where additional compute is allocated during inference to enhance LLM safety and reasoning capabilities. |
| Approach: | They propose a reward-shifted speculative sampling algorithm that aligns a draft model with human preferences while the target model remains unchanged. |
| Outcome: | The proposed algorithm achieves superior gold reward scores at a significantly reduced inference cost in test-time weak-to-strong alignment experiments. |