Long Minh Bui

Long Minh Bui, Tuan Anh Le Van, Tung Phi Duc et al.

Model merging aims to combine existing single-task solutions into a multi-task solution without additional data-driven fine-tuning.~Most existing approaches achieve this using geometric properties of local solution spaces. However, such geometric views provide limited guidance for scoring how statistically useful each task-specific update direction is across tasks during merging. To address this, we formulate model merging from a new perspective of probabilistic inference under a product-of-experts (PoE) scenario where each single-task solution defines an energy-based expert model (EBM) over the merged parameters. We show that several existing model merging methods arise as special cases of our framework under energy designs that impose implicit Gaussian assumptions on directional residuals between merged and task-specific models. Empirically, we find that these residuals are often heavy-tailed which exposes a mismatch with the imposed light-tailed Gaussian structures. We address this with a heavy-tailed PoE design based on Cauchy experts, which better captures the observed residual behavior while admitting a provably convergent inference procedure. Experiments across multiple tasks and architectures show significant improvements over state-of-the-arts baselines. Our code is available at https://github.com/MinhLong210/PoE-EBM-Merging.git.

Hang Thi-Thuy Le, Long Minh Bui, Minh Hoang et al.

Continual fine-tuning aims to adapt a pre-trained backbone to new tasks sequentially while preserving performance on earlier tasks whose data are no longer available. Existing approaches fall into two categories which include input- and parameter-adaptation. Input-adaptation methods rely on retrieving the most relevant prompts at test time, but require continuously learning a retrieval function that is prone to forgetting. Parameter-adaptation methods instead use a fixed input embedding function to enable retrieval-free prediction and avoid forgetting, but sacrifice representation adaptability. To combine their best strengths, we propose a new parameter-adaptation method that enables adaptive use of input embeddings during test time with parameter-free retrieval. We derive task-retrieval error bounds for a clustering-based, parameter-free paradigm, providing theoretical guarantees that link low retrieval error to structural properties of task-specific representation clusters, revealing a fresh insight into how well-organized clustering structure will enable reliable retrieval. Motivated by this insight, our method is designed with two key components: (i) an adaptive module composition strategy that learns informative task-specific updates to preserve and complement prior knowledge, and (ii) a clustering-based retrieval mechanism that captures distinct representation signatures for each task, enabling adaptive representation use at test time. Extensive experiments show that these components work synergistically to improve retrieval and predictive performance under large shifts in task semantics.