Papers by Zhangchen Xu
Small Models Struggle to Learn from Strong Reasoners (2025.findings-acl)
Copied to clipboard
Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubramanian, Radha Poovendran
| Challenge: | a small learning gap exists between large and small language models . long CoT data and large model responses are not beneficial for small models - a problem that may be due to the small student model's ability to handle distribution shifts. |
| Approach: | They propose a mix distillation strategy that balances reasoning complexity by combining long and short CoT examples or reasoning from both larger and smaller models. |
| Outcome: | The proposed strategy outperforms training on large and small models on short CoT and small model CoT. |
Understanding the Language Model to Solve the Symbolic Multi-Step Reasoning Problem from the Perspective of Buffer Mechanism (2025.findings-emnlp)
Copied to clipboard
Zhiwei Wang, Yunji Wang, Zhongwang Zhang, Zhangchen Zhou, Hui Jin, Tianyang Hu, Jiacheng Sun, Zhenguo Li, Yaoyu Zhang, Zhi-Qin John Xu
| Challenge: | Large language models struggle with complex reasoning tasks, such as mathematical problem-solving. |
| Approach: | They constructed a symbolic multi-step reasoning task to investigate the information propagation mechanisms in Transformer models when solving the task through direct answering and Chain-of-Thought (CoT) reasoning. |
| Outcome: | The proposed algorithm improves on 7 multi-step reasoning datasets, while introducing only 132 trainable parameters. |
Stronger Models are Not Always Stronger Teachers for Instruction Tuning (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing methods to optimize instruction-following capabilities of large language models (LLMs) assume that larger or stronger models are stronger teachers and therefore adopt smaller models as response generators. |
| Approach: | They propose to use large-scale instruction datasets to tune large language models to align with specific tasks and user intents. |
| Outcome: | The proposed metric outperforms most baselines in identifying the effectiveness of response generators. |
SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decoding (2024.acl-long)
Copied to clipboard
| Challenge: | Despite advances in large language models, they face substantial challenges in terms of safety. |
| Approach: | They develop a safety-aware decoding strategy for large language models to defend against jailbreak attacks. |
| Outcome: | The proposed strategy outperforms six defense methods against jailbreak attacks on five LLMs. |
Temporal Sampling for Forgotten Reasoning in LLMs (2026.acl-long)
Copied to clipboard
Yuetai Li, Zhangchen Xu, Fengqing Jiang, Bhaskar Ramasubramanian, Luyao Niu, Bill Yuchen Lin, Xiang Yue, Radha Poovendran
| Challenge: | a new metric measures the percentage of questions that were answered incorrectly during fine-tuning . |
| Approach: | They propose a decoding strategy that draws outputs from multiple checkpoints along the training trajectory. |
| Outcome: | The proposed method improves reasoning performance and consistency across benchmarks. |
SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities (2025.findings-acl)
Copied to clipboard
Fengqing Jiang, Zhangchen Xu, Yuetai Li, Luyao Niu, Zhen Xiang, Bo Li, Bill Yuchen Lin, Radha Poovendran
| Challenge: | Emerging large reasoning models (LRMs) leverage long chain-of-thought (CoT) reasoning to enhance their reasoning capabilities. |
| Approach: | They conduct a systematic study of LRM safety using human annotations to assess their safety. |
| Outcome: | The proposed safety measures are compared to state-of-the-art models on strong and wildjailbreak datasets. |
KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing code-focused resources typically fail to ensure either the breadth of coverage or verifiable correctness. |
| Approach: | They propose a synthetic dataset that provides high-quality, verifiable training data for Large Language Models for coding. |
| Outcome: | The proposed dataset surpasses Qwen2.5-Coder-32B-Instruct and DeepSeek-R1-Distill-Llama-70B in performance on coding benchmarks. |
CleanGen: Mitigating Backdoor Attacks for Generation Tasks in Large Language Models (2024.emnlp-main)
Copied to clipboard
Yuetai Li, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Dinuka Sahabandu, Bhaskar Ramasubramanian, Radha Poovendran
| Challenge: | Generative large language models (LLMs) have remarkable performance in generation tasks, but datasets used to train or fine-tune these models are often not disclosed to users. |
| Approach: | They develop an inference time defense called CleanGen to mitigate backdoor attacks for generation tasks in large language models. |
| Outcome: | The proposed inference time defense achieves lower attack success rates (ASR) compared to baseline defenses for all five backdoor attacks. |