Papers by Zhixuan Chu
Perplexity-Aware Data Scaling Law: Perplexity Landscapes Predict Performance for Continual Pre-training (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have impressive capabilities across a wide range of domains, but their generalpurpose pre-training objectives often leave them illsuited for specialized applications such as healthcare. |
| Approach: | They propose a perplexity-aware data scaling law that establishes a predictive relationship between the perplexities of domain-specific data and the test loss. |
| Outcome: | Experiments on medical and general-domain benchmarks show that the proposed scaling law consistently identifies near-optimal training subsets with significantly reduced data consumption. |
Self-Para-Consistency: Improving Reasoning Tasks at Low Cost for Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent studies have shown that self-consistency decoding can improve performance for complex reasoning tasks with large language models. |
| Approach: | They propose a self-consistency decoding strategy that generates multiple paraphrases for each test question and then generates reasoning paths for the original and all the paraphrased questions based on greedy decoding. |
| Outcome: | The proposed strategy reduces the sampling number and improves performance on complex reasoning tasks. |
Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision Boundary (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) often refuse to answer legitimate queries, causing models to treat many reasonable prompts as potentially risky. |
| Approach: | They propose a framework that automatically generates and selects overrefusal prompts near the safety boundary. |
| Outcome: | The proposed framework identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. |
Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework (2026.findings-acl)
Copied to clipboard
Jiaqi Weng, Han Zheng, Hanyu Zhang, Ej Zhou, Qinqin He, Jialing Tao, Hui Xue, Zhixuan Chu, Xiting Wang
| Challenge: | Existing studies on how SAEs derive most fine-grained latent features for safety remain unexplored. |
| Approach: | They propose a framework for interpreting SAE features in safety-critical domains . they train a suite of SAEs with human-readable explanations and systematic evaluations based on pornography, politics, violence, and terror . |
| Outcome: | The proposed framework reduces interpretation cost by 55% and improves safety-critical features. |
Incorporating Causal Analysis into Diversified and Logical Response Generation (2022.coling-1)
Copied to clipboard
| Challenge: | Existing generation-based models generate generic and safe responses such as "So am I" or "I don't know" |
| Approach: | They propose to predict the mediators to preserve relevant information and auto-regressively incorporate the mediator into generating process. |
| Outcome: | The proposed model generates relevant and informative responses and outperforms the state-of-the-art in terms of automatic metrics and human evaluations. |
Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics (2026.acl-long)
Copied to clipboard
Ziwen Xu, Chenyan WU, Hengyu Sun, Haiwen Hong, Mengru Wang, Yunzhi Yao, Longtao Huang, Hui Xue, Shumin Deng, Zhixuan Chu, Huajun Chen, Ningyu Zhang
| Challenge: | Methods for controlling large language models (LLMs) are often studied in isolation, obscuring connections and making comparison difficult. |
| Approach: | They propose a preference-utility analysis that separates control effects into preference and utility, and measures both on a shared log-odds scale using polarity-paired contrastive examples. |
| Outcome: | The proposed approach improves preference while preserving utility. |