PPaperPicks

Xinmei Tian

University of Science and Technology of China, Department of Electronic Engineering and Information Science, Hefei, China

24 papers at tracked venues · 22 at CORE A* · active 20242026

Venues

Frequent coauthors

Papers

  1. Bridging the Language Gap: Uncovering and Aligning Shared Circuits for Multi-Hop Reasoning in Multilingual LLMs
  2. Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality
  3. A Similarity Paradigm Through Textual Regularization Without Forgetting
  4. A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops
  5. An Effective Levelling Paradigm for Unlabeled Scenarios
  6. Detecting Generated Images by Fitting Natural Image Distributions
  7. Enhancing Target-unspecific Tasks through a Features Matrix
  8. Epistemic Uncertainty for Generated Image Detection
  9. Interpret and Improve In-Context Learning via the Lens of Input-Label Mappings
  10. Leveraging Submodule Linearity Enhances Task Arithmetic Performance in LLMs
  11. Towards Generalizable Detector for Generated Image
  12. Visual Evidence Prompting Mitigates Hallucinations in Large Vision-Language Models
  13. Adaptive Time-Stepping Schedules for Diffusion Models
  14. Advancing Prompt Learning through an External Layer
  15. Convergence of Bayesian Bilevel Optimization
  16. Enhanced Motion-Text Alignment for Image-to-Video Transfer Learning
  17. FedImpro: Measuring and Improving Client Update in Federated Learning
  18. From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning
  19. Interpretable Composition Attribution Enhancement for Visio-linguistic Compositional Understanding
  20. Interpreting and Improving Large Language Models in Arithmetic Calculation
  21. Out-of-Distribution Detection with Negative Prompts
  22. Robust Training of Federated Models with Extremely Label Deficiency
  23. Sheared Backpropagation for Fine-Tuning Foundation Models
  24. Towards Theoretical Understandings of Self-Consuming Generative Models