Papers by Xiao Pu
LoRAMoE: Alleviating World Knowledge Forgetting in Large Language Models via MoE-Style Plugin (2024.acl-long)
Copied to clipboard
Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Wei Shen, Limao Xiong, Yuhao Zhou, Xiao Wang, Zhiheng Xi, Xiaoran Fan, Shiliang Pu, Jiang Zhu, Rui Zheng, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Experimental results show that, as the instruction data increases, LoRAMoE can significantly improve the ability to process downstream tasks, while maintaining the world knowledge stored in the LLM. |
| Approach: | They propose a framework that introduces several low-rank adapters and integrates them by using a router network to freeze the backbone model and force a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks. |
| Outcome: | The proposed framework freezes the backbone model and forces a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks, to alleviate world knowledge forgetting. |
Stumbling Blocks: Stress Testing the Robustness of Machine-Generated Text Detectors Under Attacks (2024.acl-long)
Copied to clipboard
| Challenge: | Existing studies on this topic focus on the robustness of specific detectors or particular attack methods. |
| Approach: | They stress test the detectors’ robustness to malicious attacks under realistic scenarios using LLMs and metric-based detectors. |
| Outcome: | The proposed methods are based on a set of LLM-based models and their performance is compared under different budget levels. |
Is Summary Useful or Not? An Extrinsic Human Evaluation of Text Summaries on Downstream Tasks (2024.lrec-main)
Copied to clipboard
| Challenge: | a recent study focused on intrinsic evaluation, which assesses the quality of summaries, e.g. coherence, fluency, and informativeness, but it focused on task-based extrinsic evaluation to determine the usefulness of summarizations. |
| Approach: | They incorporate three downstream tasks to measure the usefulness of summaries . they find that fine-tuned models produce more useful summary across all three tasks . |
| Outcome: | The proposed model produces more useful summaries across all three tasks compared to zero-shot models . human evaluation provides more reliable performance assessment compared with automatic methods . |
Video-MMMU: Evaluating Knowledge Acquisition from Multidisciplinary Professional Videos (2026.acl-long)
Copied to clipboard
| Challenge: | Existing video benchmarks do not evaluate the knowledge acquisition capabilities of Large Multimodal Models (LMMs) existing video benchmark focuses on static, general visual understanding tasks, without evaluating whether models can acquire knowledge dynamically. |
| Approach: | They propose a multi-modal, multi-discipline, multitrack benchmark that evaluates Large Multimodal Models’ ability to acquire knowledge from college-level, educational videos. |
| Outcome: | The proposed benchmark reveals a substantial gap between human learners and current Large Multimodal Models (LMMs) and focuses on improving their learning efficiency. |
Style-Compress: An LLM-Based Prompt Compression Framework Considering Task-Specific Styles (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Prompt compression reduces inference time and costs while maintaining informativeness for different usage scenarios. |
| Approach: | They propose a framework that adapts a smaller language model to compress prompts for a larger model on a new task without additional training. |
| Outcome: | The proposed framework outperforms two baseline models in four tasks . iteratively generates and selects effective compressed prompts as task-specific demonstrations . |
B4: A Black-Box Scrubbing Attack on LLM Watermarks (2025.naacl-long)
Copied to clipboard
| Challenge: | Experimental results demonstrate superior performance of black-box scrubbing attack on watermarks compared with other baselines. |
| Approach: | They propose a black-box scrubbing attack on watermarks that embeds a hidden pattern invisible to human into generated content of a specific LLM. |
| Outcome: | The proposed method outperforms baselines in 12 different environments. |
Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection (2026.acl-long)
Copied to clipboard
| Challenge: | a rapid proliferation of large language models (LLMs) generate text that increasingly resembles human writing . this makes it difficult to capture subtle cues that distinguish AI-generated content from human-written content . |
| Approach: | They propose a framework that disentangles AI-detection semantics from generator-aware artifacts by latent encoding and perturbation-based regularization. |
| Outcome: | The proposed framework disentangles AI-detection semantics from generator-aware artifacts on 20 representative LLMs across 7 categories. |
On the Zero-Shot Generalization of Machine-Generated Text Detectors (2023.findings-emnlp)
Copied to clipboard
| Challenge: | rampant proliferation of large language models generates text indistinguishable from human-written language. |
| Approach: | They train neural detectors on outputs of a new generator and test their performance on held-out generators. |
| Outcome: | The proposed detectors can be built on training data from medium-sized models. |