Papers by Yuting Zhang
Personalized Text Generation with Contrastive Activation Steering (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to personalized text generation rely on retrieval-augmented generation and parameter-efficient fine-tuning. |
| Approach: | They propose a training-free framework that disentangles and represents personalized writing style as a vector in LLM’s activation-space. |
| Outcome: | The proposed framework achieves 8% relative improvement in personalized generation while reducing storage requirements by 1700 over PEFT method. |
DCP: Dual-Cue Pruning for Efficient Large Vision-Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing pruning methods for large vision language models use visual tokens to prune . existing methods fail to balance efficiency and semantic alignment due to large number of visual token. |
| Approach: | They propose a cross-modal pruning framework that considers textual semantics and visual self-attention to combine them to achieve efficient inference acceleration. |
| Outcome: | The proposed pruning framework can retain only 25% of the visual tokens, with a minimal performance degradation of only 0.063% on LLaVA-1.5-13B. |
Sinkhorn Distance Minimization for Knowledge Distillation (2024.lrec-main)
Copied to clipboard
Xiao Cui, Yulei Qin, Yuting Gao, Enwei Zhang, Zihan Xu, Tong Wu, Ke Li, Xing Sun, Wengang Zhou, Houqiang Li
| Challenge: | Existing knowledge distillation methods investigate divergence measures but fail to deliver effective supervision when few distribution overlap exists between teacher and student. |
| Approach: | They propose a knowledge distillation method that exploits the Sinkhorn distance to ensure a nuanced assessment of the disparity between teacher and student distributions. |
| Outcome: | The proposed method outperforms state-of-the-art methods on all kinds of LLMs with encoder-only, encoder decoder, and decoded architectures. |
GlimpRouter: Efficient Collaborative Inference by Glimpsing One Token of Thoughts (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing routing strategies rely on local token probabilities or post-hoc verification, introducing significant inference overhead. |
| Approach: | They propose a step-wise collaboration framework that generates only the first token of each reasoning step and routes it to a larger model only when initial token entropy exceeds a threshold. |
| Outcome: | The proposed approach reduces inference latency while preserving accuracy. |
PsyMem: Fine-grained Psychological Alignment and Explicit Memory Control for Advanced Role-Playing LLMs (2026.tacl-1)
Copied to clipboard
| Challenge: | Existing role-playing models rely on superficial textual descriptions or simplistic metrics, inadequately modeling both intrinsic and extrinsic character dimensions. |
| Approach: | They propose a framework that integrates fine-grained psychological attributes and explicit memory control for role-playing. |
| Outcome: | The proposed framework outperforms baseline models in human-likeness and character fidelity. |
SplitThenMerge: Token-Level Skill-Compositional Sparse Mixture-of-Experts for Complex Domain-Specific Tasks (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing domain adaptation methods train heterogeneous skills together, making it difficult to reliably coordinate multiple skills when solving complex tasks. |
| Approach: | They propose a framework that decomposes domain competence into atomic skills and composes them dynamically during generation. |
| Outcome: | The proposed framework decomposes domain competence into atomic skills, trains them independently, and composes them dynamically during generation. |
Stealthy Attack on Large Language Model based Recommendation (2024.acl-long)
Copied to clipboard
| Challenge: | Recent advances in recommender systems have been overlooked due to their emphasis on textual content. |
| Approach: | They propose to introduce large language models into recommendation models to exploit the semantic understanding and strong transferability of LLMs. |
| Outcome: | The proposed approach significantly boosts an item’s exposure by altering its textual content during the testing phase, without requiring direct interference with the model’s training process. |
Agent-based Substructure Counting under Local Differential Privacy (2026.acl-long)
Copied to clipboard
| Challenge: | Recent studies have demonstrated the ability of Large Language Models (LLMs) to process graph problems. |
| Approach: | They propose to decompose substructure counting into node-level tasks distributed among node agents and embed the knowledge of distributed algorithms and DP frameworks in the curator agent and privacy controller. |
| Outcome: | Extensive experiments on 6 real-world datasets validate the effectiveness of the proposed framework for substructure counting tasks under edge local differential privacy (LDP). |
From Verbatim to Gist: Distilling Pyramidal Multimodal Memory via Semantic Information Bottleneck for Long-Horizon Video Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Existing multimodal large language models struggle with long-horizon video understanding due to limited context windows and static memory mechanisms that fail to mirror human cognitive efficiency. |
| Approach: | They propose a pyramidal multimodal memory architecture grounded in Fuzzy-Trace Theory that structures memory hierarchically into a *Sensory Buffer*, *Episodic Stream*, and *Symbolic Schema*. |
| Outcome: | The proposed architecture achieves state-of-the-art on both offline and streaming tasks, demonstrating robust generalization and validating the effectiveness of cognition-inspired memory organization. |