Yuhang Liu

Developing human understandable interpretation of large language models (LLMs) becomes increasingly critical for their deployment in essential domains. Mechanistic interpretability seeks to mitigate the issues through extracts human-interpretable process and concepts from LLMs' activations. Sparse autoencoders (SAEs) have emerged as a popular approach for extracting interpretable and monosemantic concepts by decomposing the LLM internal representations into a dictionary. Despite their empirical progress, SAEs suffer from a fundamental theoretical ambiguity: the well-defined correspondence between LLM representations and human-interpretable concepts remains unclear. This lack of theoretical grounding gives rise to several methodological challenges, including difficulties in principled method design and evaluation criteria. In this work, we show that, under mild assumptions, LLM representations can be approximated as a {linear mixture} of the log-posteriors over concepts given the input context, through the lens of a latent variable model where concepts are treated as latent variables. This motivates a principled framework for concept extraction, namely Concept Component Analysis (ConCA), which aims to recover the log-posterior of each concept from LLM representations through a {unsupervised} linear unmixing process. We explore a specific variant, termed sparse ConCA, which leverages a sparsity prior to address the inherent ill-posedness of the unmixing problem. We implement 12 sparse ConCA variants and demonstrate their ability to extract meaningful concepts across multiple LLMs, offering theory-backed advantages over SAEs.

Continual learning (CL) aims to incrementally train a model on a sequence of tasks while retaining performance on prior ones. However, storing and replaying data is often infeasible due to privacy or security constraints and impractical for arbitrary pre-trained models. Data-free CL seeks to update models without access to previous data. Beyond regularization, we employ model inversion to synthesize data from the trained model, enabling replay without storing samples. Yet, model inversion in predictive models faces two challenges: (1) generating inputs solely from compressed output labels causes drift between synthetic and real data, and replaying such data can erode prior knowledge; (2) inversion is computationally expensive since each step backpropagates through the full model. These issues are amplified in large pre-trained models such as CLIP. To improve efficiency, we propose Per-layer Model Inversion (PMI), inspired by faster convergence in single-layer optimization. PMI provides strong initialization for full-model inversion, substantially reducing iterations. To mitigate feature shift, we model class-wise features via Gaussian distributions and contrastive model, ensuring alignment between synthetic and real features. Combining PMI and feature modeling, our approach enables continual learning of new classes by generating pseudo-images from semantic-aware projected features, achieving strong effectiveness and compatibility across multiple CL settings.

Few-shot learning (FSL) often requires effective adaptation of models using limited labeled data. However, most existing FSL methods rely on entangled representations, requiring the model to implicitly recover the unmixing process to obtain disentangled representations using only limited supervision, which hinders effective adaptation. Recent theoretical studies show that multimodal contrastive learning methods, such as CLIP, can disentangle latent representations up to linear transformations. In light of this, we propose the Causal CLIP Adapter (CCA), a novel framework that explicitly disentangles visual features extracted from CLIP using unsupervised Independent Component Analysis (ICA). This removes the need to learn the unmixing process from the labeled data, thereby reducing the number of trainable parameters and mitigating overfitting. Taking a step further, while ICA can obtain visual disentangled representations, it may also disrupt CLIP's intra- and inter-modal alignment. To counteract this, CCA further leverages CLIP's inherent cross-modal alignment by enhancing it in two ways: unidirectionally, through fine-tuning a CLIP-based text classifier, and bidirectionally, via a cross-attention mechanism that enriches visual and textual representations through mutual interaction. Both unimodal and cross-modal classification outputs can be effectively combined linearly to improve classification accuracy. Extensive experiments on 11 benchmark datasets demonstrate that our method consistently outperforms state-of-the-art approaches in terms of few-shot performance and robustness to distributional shifts, while maintaining computational efficiency. Code will be available at https://github.com/tianjiao-j/CCA.

Accurately predicting stock returns is crucial for effective portfolio management. However, existing methods often overlook a fundamental issue in the market, namely, distribution shifts, making them less practical for predicting future markets or newly listed stocks. This study introduces a novel approach to address this challenge by focusing on the acquisition of invariant features across various environments, thereby enhancing robustness against distribution shifts. Specifically, we present InvariantStock, a designed learning framework comprising two key modules: an environment-aware prediction module and an environment-agnostic module. Through the designed learning of these two modules, the proposed method can learn invariant features across different environments in a straightforward manner, significantly improving its ability to handle distribution shifts in diverse market settings. Our results demonstrate that the proposed InvariantStock not only delivers robust and accurate predictions but also outperforms existing baseline methods in both prediction tasks and backtesting within the dynamically changing markets of China and the United States.

One of the significant challenges in reinforcement learning (RL) when dealing with noise is estimating latent states from observations. Causality provides rigorous theoretical support for ensuring that the underlying states can be uniquely recovered through identifiability. Consequently, some existing work focuses on establishing identifiability from a causal perspective to aid in the design of algorithms. However, these results are often derived from a purely causal viewpoint, which may overlook the specific RL context. We revisit this research line and find that incorporating RL-specific context can reduce unnecessary assumptions in previous identifiability analyses for latent states. More importantly, removing these assumptions allows algorithm design to go beyond the earlier boundaries constrained by them. Leveraging these insights, we propose a novel approach for general partially observable Markov Decision Processes (POMDPs) by replacing the complicated structural constraints in previous methods with two simple constraints for transition and reward preservation. With the two constraints, the proposed algorithm is guaranteed to disentangle state and noise that is faithful to the underlying dynamics. Empirical evidence from extensive benchmark control tasks demonstrates the superiority of our approach over existing counterparts in effectively disentangling state belief from noise.

Multi-source domain adaptation (MSDA) addresses the challenge of learning a label prediction function for an unlabeled target domain by leveraging both the labeled data from multiple source domains and the unlabeled data from the target domain. Conventional MSDA approaches often rely on covariate shift or conditional shift paradigms, which assume a consistent label distribution across domains. However, this assumption proves limiting in practical scenarios where label distributions do vary across domains, diminishing its applicability in real-world settings. For example, animals from different regions exhibit diverse characteristics due to varying diets and genetics. Motivated by this, we propose a novel paradigm called latent covariate shift (LCS), which introduces significantly greater variability and adaptability across domains. Notably, it provides a theoretical assurance for recovering the latent cause of the label variable, which we refer to as the latent content variable. Within this new paradigm, we present an intricate causal generative model by introducing latent noises across domains, along with a latent content variable and a latent style variable to achieve more nuanced rendering of observational data. We demonstrate that the latent content variable can be identified up to block identifiability due to its versatile yet distinct causal structure. We anchor our theoretical insights into a novel MSDA method, which learns the label distribution conditioned on the identifiable latent content variable, thereby accommodating more substantial distribution shifts. The proposed approach showcases exceptional performance and efficacy on both simulated and real-world datasets.

The task of causal representation learning aims to uncover latent higher-level causal representations that affect lower-level observations. Identifying true latent causal representations from observed data, while allowing instantaneous causal relations among latent variables, remains a challenge, however. To this end, we start from the analysis of three intrinsic properties in identifying latent space from observations: transitivity, permutation indeterminacy, and scaling indeterminacy. We find that transitivity acts as a key role in impeding the identifiability of latent causal representations. To address the unidentifiable issue due to transitivity, we introduce a novel identifiability condition where the underlying latent causal model satisfies a linear-Gaussian model, in which the causal coefficients and the distribution of Gaussian noise are modulated by an additional observed variable. Under some mild assumptions, we can show that the latent causal representations can be identified up to trivial permutation and scaling. Furthermore, based on this theoretical result, we propose a novel method, termed Structural caUsAl Variational autoEncoder, which directly learns latent causal representations and causal relationships among them, together with the mapping from the latent causal variables to the observed ones. We show that the proposed method learns the true parameters asymptotically. Experimental results on synthetic and real data demonstrate the identifiability and consistency results and the efficacy of the proposed method in learning latent causal representations.