Papers by Weijun Li
Cut the Deadwood Out: Backdoor Purification via Guided Module Substitution (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Model NLP models are often trained on datasets from untrusted platforms, posing significant risks of data poisoning attacks. |
| Approach: | They propose a retraining-free method that selectively replaces modules in the victim model based on a trade-off signal between utility and backdoor. |
| Outcome: | The proposed method outperforms even the strongest defense baseline against challenging attacks like LWS. |
Personalizing LLMs with Binary Feedback: A Preference-Calibrated Optimization Framework (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods focus on isolated user histories, neglecting the essential role of inter-user differences. |
| Approach: | They propose a framework that personalizes Large Language Models via preference-calibrated binary signals. |
| Outcome: | The proposed framework outperforms baselines in a variety of personalization tasks and backbone LLMs. |
SwapMoE: Serving Off-the-shelf MoE-based Large Language Models with Tunable Memory Budget (2024.acl-long)
Copied to clipboard
| Challenge: | Mixture of experts (MoE) is a popular technique to improve capacity of Large Language Models (LLMs) but memory-constrained devices are a major concern in edge AI training and serving. |
| Approach: | They propose a framework for efficient serving of MoE-based large language models with tunable memory budgets. |
| Outcome: | Experiments show that SwapMoE can reduce memory consumption while maintaining reasonable accuracy. |
Skill Weaving: Efficient LLM Improvement via Modular Skillpacks (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can specialize under fixed memory and inference budgets, but they struggle to achieve high performance across heterogeneous domains. |
| Approach: | They propose a modular improvement framework that partitions full capabilities of a general-purpose model into domain-specific delta modules that reorganize and refine the model's internal knowledge. |
| Outcome: | The proposed framework outperforms monolithic models on multi-task and agentic benchmarks and achieves up to 4 speedup. |
Vocabulary Hijacking in LVLMs: Unveiling Critical Attention Heads by Excluding Inert Tokens to Mitigate Hallucination (2026.acl-long)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) are capable of processing visual inputs, but are susceptible to hallucinations. |
| Approach: | They propose a method to localize and localize specific visual tokens, which are defined as **Inert Tokens**, across layers, revealing a rigid semantic collapse. |
| Outcome: | The proposed approach reduces the likelihood of LVLMs being hijacked by visual inputs while maintaining general capabilities. |
Seeing the Forest through the Trees: Data Leakage from Partial Transformer Gradients (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies show that distributed machine learning is vulnerable to gradient inversion attacks . a recent study demonstrated the possibility of reconstructing private textual training data using partial gradients . |
| Approach: | They propose to use partial gradients to reconstruct training data using a Transformer model. |
| Outcome: | The proposed method is vulnerable to gradient inversion attacks, the authors show . they show that applying differential privacy on gradients during training offers limited protection . |