Jinjin Chi

Jinjin Chi, Taoping Liu, Mengtao Yin et al.

Disentangled representation learning aims to capture the underlying explanatory factors of observed data, enabling a principled understanding of the data-generating process. Recent advances in generative modeling have introduced new paradigms for learning such representations. However, existing diffusion-based methods encourage factor independence via inductive biases, yet frequently lack strong semantic alignment. In this work, we propose a flow matching-based framework for disentangled representation learning, which casts disentanglement as learning factor-conditioned flows in a compact latent space. To enforce explicit semantic alignment, we introduce a non-overlap (orthogonality) regularizer that suppresses cross-factor interference and reduces information leakage between factors. Extensive experiments across multiple datasets demonstrate consistent improvements over representative baselines, yielding higher disentanglement scores as well as improved controllability and sample fidelity.

Bing Wang, Ximing Li, Changchun Li et al.

Recently, the prominent performance of large language models (LLMs) has been largely driven by multi-task instruct-tuning. Unfortunately, this training paradigm suffers from a key issue, named cross-task interference, due to conflicting gradients over shared parameters among different tasks. Some previous methods mitigate this issue by isolating task-specific parameters, e.g., task-specific neuron selection and mixture-of-experts. In this paper, we empirically reveal that the cross-task interference still exists for the existing solutions because of many parameters also shared by different tasks, and accordingly, we propose a novel solution, namely Basic Abilities Decomposition for multi-task Instruct-Tuning (BADIT). Specifically, we empirically find that certain parameters are consistently co-activated, and that co-activated parameters naturally organize into base groups. This motivates us to analogize that LLMs encode several orthogonal basic abilities, and that any task can be represented as a linear combination of these abilities. Accordingly, we propose BADIT that decomposes LLM parameters into orthogonal high-singular-value LoRA experts representing basic abilities, and dynamically enforces their orthogonality during training via spherical clustering of rank-1 components. We conduct extensive experiments on the SuperNI benchmark with 6 LLMs, and empirical results demonstrate that BADIT can outperform SOTA methods and mitigate the degree of cross-task interference.

Jinjin Chi, Zhiyao Yang, Jihong Ouyang et al.

Wasserstein barycenter, built on the theory of optimal transport, provides a powerful framework to aggregate probability distributions, and it has increasingly attracted great attention within the machine learning community. However, it suffers from severe computational burden, especially for high dimensional and continuous settings. To this end, we develop a novel continuous approximation method for the Wasserstein barycenters problem given sample access to the input distributions. The basic idea is to introduce a variational distribution as the approximation of the true continuous barycenter, so as to frame the barycenters computation problem as an optimization problem, where parameters of the variational distribution adjust the proxy distribution to be similar to the barycenter. Leveraging the variational distribution, we construct a tractable dual formulation for the regularized Wasserstein barycenter problem with c-cyclical monotonicity, which can be efficiently solved by stochastic optimization. We provide theoretical analysis on convergence and demonstrate the practical effectiveness of our method on real applications of subset posterior aggregation and synthetic data.

Jinjin Chi, Jihong Ouyang, Changchun Li et al.

The top word list, i.e., the top-M words with highest marginal probability in a given topic, is the standard topic representation in topic models. Most of recent automatical topic labeling algorithms and popular topic quality metrics are based on it. However, we find, empirically, words in this type of top word list are not always representative. The objective of this paper is to find more representative top word lists for topics. To achieve this, we rerank the words in a given topic by further considering marginal probability on words over every other topic. The reranking list of top-M words is used to be a novel topic representation for topic models. We investigate three reranking methodologies, using (1) standard deviation weight, (2) standard deviation weight with topic size and (3) Chi Square \c{hi}2statistic selection. Experimental results on real world collections indicate that our representations can extract more representative words for topics, agreeing with human judgements.