We propose PIP (Perturbation-based Iterative Pruning), a method that iteratively prunes parameters based on the distinction between unperturbed and perturbed views. Experimental results show that PIP reduces parameter count by approximately 20% while retaining over 85% of the original accuracy.

We propose TIV, an innovative framework that compresses token-level reasoning into compact vectors without sacrificing performance. Rather than generating explicit thoughts, TIV injects learnable vectors into the post-attention hidden states of the final token across Transformer layers, enabling implicit and lightweight reasoning.

We propose LLM4MS, a method leveraging expert knowledge from large language models to generate discriminative spectral embeddings for improved compound identification. Experimental results show a 13.7% improvement in accuracy over existing methods on a million-scale library, with a query speed of nearly 15,000 queries per second.

We propose UniCorn, a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. By partitioning a single UMM into three collaborative roles: Proposer, Solver, and Judge, UniCorn generates high-quality interactions via self-play and employs cognitive pattern reconstruction to distill latent understanding into explicit generative signals.

We propose L2UL, a learning-based and model-agnostic framework for machine unlearning that learns the unlearning behavior instead of manually designing an unlearning function. By representing data distributions with kernel mean embeddings, L2UL achieves accuracy comparable to retraining while providing substantial efficiency gains, especially in data-intensive scenarios.