Sparse Autoencoders (SAEs) are a prominent tool in mechanistic interpretability (MI) for decomposing neural network activations into interpretable features. However, the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs, undermining reproducibility and complicating model comparison. We study run-to-run feature consistency in SAEs and argue that it should be reported as a standard evaluation axis alongside reconstruction and sparsity. We propose the Pairwise Dictionary Mean Correlation Coefficient (PW-MCC) as an assignment-based metric to quantify consistency and demonstrate that high levels are achievable (PW-MCC ≈ 0.80 for TopK SAEs on LLM activations) with appropriate architectural choices.Our contributions include: (i) theoretical grounding for strong consistency in the idealized setting of TopK SAEs; (ii) synthetic validation using a model organism, which verifies PW-MCC as a reliable proxy for ground-truth recovery; and (iii) empirical analysis on LLM activations, where PW-MCC correlates with the similarity of automatically generated natural-language feature explanations.

Shaoan Xie, Lingjing Kong, Xiangchen Song, Xinshuai Dong, Guangyi Chen, Eric P. Xing, Kun Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Temporally causal representation learning aims to identify the latent causal process from time series observations, but most methods require the assumption that the latent causal processes do not have instantaneous relations. Although some recent methods achieve identifiability in the instantaneous causality case, they require either interventions on the latent variables or grouping of the observations, which are in general difficult to obtain in real-world scenarios. To fill this gap, we propose an \textbf{ID}entification framework for instantane\textbf{O}us \textbf{L}atent dynamics (\textbf{IDOL}) by imposing a sparse influence constraint that the latent causal processes have sparse time-delayed and instantaneous relations. Specifically, we establish identifiability results of the latent causal process based on sufficient variability and the sparse influence constraint by employing contextual information of time series data. Based on these theories, we incorporate a temporally variational inference architecture to estimate the latent variables and a gradient-based sparsity regularization to identify the latent causal process. Experimental results on simulation datasets illustrate that our method can identify the latent causal process. Furthermore, evaluations on multiple human motion forecasting benchmarks with instantaneous dependencies indicate the effectiveness of our method in real-world settings.

The autoregressive decoding for text generation in large language models (LLMs), while widely used, is inherently suboptimal due to the lack of a built-in mechanism to perform refinement and/or correction of the generated content. In this paper, we consider optimality in terms of the joint probability over the generated response, when jointly considering all tokens at the same time. We theoretically characterize the potential deviation of the autoregressively generated response from its globally optimal counterpart that is of the same length. Our analysis suggests that we need to be cautious when noticeable uncertainty arises during text generation, which may signal the sub-optimality of the generation history. To address the pitfall of autoregressive decoding for text generation, we propose an approach that incorporates a sliding reflection window and a pausing criterion, such that refinement and generation can be carried out interchangeably as the decoding proceeds. Our selective refinement framework strikes a balance between efficiency and optimality, and our extensive experimental results demonstrate the effectiveness of our approach.

Despite Large Language Models' remarkable capabilities, understanding their internal representations remains challenging. Mechanistic interpretability tools such as sparse autoencoders (SAEs) were developed to extract interpretable features from LLMs but lack temporal dependency modeling, instantaneous relation representation, and more importantly theoretical guarantees—undermining both the theoretical foundations and the practical confidence necessary for subsequent analyses. While causal representation learning (CRL) offers theoretically-grounded approaches for uncovering latent concepts, existing methods cannot scale to LLMs' rich conceptual space due to inefficient computation. To bridge the gap, we introduce an identifiable temporal causal representation learning framework specifically designed for LLMs' high-dimensional concept space, capturing both time-delayed and instantaneous causal relations. Our approach provides theoretical guarantees and demonstrates efficacy on synthetic datasets scaled to match real-world complexity. By extending SAE techniques with our temporal causal framework, we successfully discover meaningful concept relationships in LLM activations. Our findings show that modeling both temporal and instantaneous conceptual relationships advances the interpretability of LLMs.

Most existing causal discovery methods rely on the assumption of no latent confounders, limiting their applicability in solving real-life problems. In this paper, we introduce a novel, versatile framework for causal discovery that accommodates the presence of causally-related hidden variables almost everywhere in the causal network (for instance, they can be effects of measured variables), based on rank information of covariance matrix over measured variables. We start by investigating the efficacy of rank in comparison to conditional independence and, theoretically, establish necessary and sufficient conditions for the identifiability of certain latent structural patterns. Furthermore, we develop a Rank-based Latent Causal Discovery algorithm, RLCD, that can efficiently locate hidden variables, determine their cardinalities, and discover the entire causal structure over both measured and hidden ones. We also show that, under certain graphical conditions, RLCD correctly identifies the Markov Equivalence Class of the whole latent causal graph asymptotically. Experimental results on both synthetic and real-world personality data sets demonstrate the efficacy of the proposed approach in finite-sample cases. Our code will be publicly available.

Identifying the underlying time-delayed latent causal processes in sequential data is vital for grasping temporal dynamics and making downstream reasoning. While some recent methods can robustly identify these latent causal variables, they rely on strict assumptions about the invertible generation process from latent variables to observed data. However, these assumptions are often hard to satisfy in real-world applications containing information loss. For instance, the visual perception process translates a 3D space into 2D images, or the phenomenon of persistence of vision incorporates historical data into current perceptions. To address this challenge, we establish an identifiability theory that allows for the recovery of independent latent components even when they come from a nonlinear and non-invertible mix. Using this theory as a foundation, we propose a principled approach, CaRiNG, to learn the Causal Representation of Non-invertible Generative temporal data with identifiability guarantees. Specifically, we utilize temporal context to recover lost latent information and apply the conditions in our theory to guide the training process. Through experiments conducted on synthetic datasets, we validate that our CaRiNG method reliably identifies the causal process, even when the generation process is non-invertible. Moreover, we demonstrate that our approach considerably improves temporal understanding and reasoning in practical applications.

Causal Temporal Representation Learning (Ctrl) methods aim to identify the temporal causal dynamics of complex nonstationary temporal sequences. Despite the success of existing Ctrl methods, they require either directly observing the domain variables or assuming a Markov prior on them. Such requirements limit the application of these methods in real-world scenarios when we do not have such prior knowledge of the domain variables. To address this problem, this work adopts a sparse transition assumption, aligned with intuitive human understanding, and presents identifiability results from a theoretical perspective. In particular, we explore under what conditions on the significance of the variability of the transitions we can build a model to identify the distribution shifts. Based on the theoretical result, we introduce a novel framework, *Causal Temporal Representation Learning with Nonstationary Sparse Transition* (CtrlNS), designed to leverage the constraints on transition sparsity and conditional independence to reliably identify both distribution shifts and latent factors. Our experimental evaluations on synthetic and real-world datasets demonstrate significant improvements over existing baselines, highlighting the effectiveness of our approach.

Causal analysis for time series data, in particular estimating individualized treatment effect (ITE), is a key task in many real world applications, such as finance, retail, healthcare, etc. Real world time series, i.e., large-scale irregular or sparse and intermittent time series, raise significant challenges to existing work attempting to estimate treatment effects. Specifically, the existence of hidden confounders can lead to biased treatment estimates and complicate the causal inference process. In particular, anomaly hidden confounders which exceed the typical range can lead to high variance estimates. Moreover, in continuous time settings with irregular samples, it is challenging to directly handle the dynamics of causality. In this paper, we leverage recent advances in Lipschitz regularization and neural controlled differential equations (CDE) to develop an effective and scalable solution, namely LipCDE, to address the above challenges. LipCDE can directly model the dynamic causal relationships between historical data and outcomes with irregular samples by considering the boundary of hidden confounders given by Lipschitz constrained neural networks. Furthermore, we conduct extensive experiments on both synthetic and real world datasets to demonstrate the effectiveness and scalability of LipCDE.

In unsupervised causal representation learning for sequential data with time-delayed latent causal influences, strong identifiability results for the disentanglement of causally-related latent variables have been established in stationary settings by leveraging temporal structure. However, in nonstationary setting, existing work only partially addressed the problem by either utilizing observed auxiliary variables (e.g., class labels and/or domain indexes) as side information or assuming simplified latent causal dynamics. Both constrain the method to a limited range of scenarios. In this study, we further explored the Markov Assumption under time-delayed causally related process in nonstationary setting and showed that under mild conditions, the independent latent components can be recovered from their nonlinear mixture up to a permutation and a component-wise transformation, without the observation of auxiliary variables. We then introduce NCTRL, a principled estimation framework, to reconstruct time-delayed latent causal variables and identify their relations from measured sequential data only. Empirical evaluations demonstrated the reliable identification of time-delayed latent causal influences, with our methodology substantially outperforming existing baselines that fail to exploit the nonstationarity adequately and then, consequently, cannot distinguish distribution shifts.

Functional magnetic resonance imaging (fMRI) has become one of the most common imaging modalities for brain function analysis. Recently, graph neural networks (GNN) have been adopted for fMRI analysis with superior performance. Unfortunately, traditional functional brain networks are mainly constructed based on similarities among region of interests (ROI), which are noisy and agnostic to the downstream prediction tasks and can lead to inferior results for GNN-based models. To better adapt GNNs for fMRI analysis, we propose TBDS, an end-to-end framework based on Task-aware Brain connectivity DAG (short for Directed Acyclic Graph) Structure generation for fMRI analysis. The key component of TBDS is the brain network generator which adopts a DAG learning approach to transform the raw time-series into task-aware brain connectivities. Besides, we design an additional contrastive regularization to inject task-specific knowledge during the brain network generation process. Comprehensive experiments on two fMRI datasets, namely Adolescent Brain Cognitive Development (ABCD) and Philadelphia Neuroimaging Cohort (PNC) datasets demonstrate the efficacy of TBDS. In addition, the generated brain networks also highlight the prediction-related brain regions and thus provide unique interpretations of the prediction results. Our implementation will be published upon acceptance.

To alleviate data sparsity and cold-start problems of traditional recommender systems (RSs), incorporating knowledge graphs (KGs) to supplement auxiliary information has attracted considerable attention recently. However, simply integrating KGs in current KG-based RS models is not necessarily a guarantee to improve the recommendation performance, which may even weaken the holistic model capability. This is because the construction of these KGs is independent of the collection of historical user-item interactions; hence, information in these KGs may not always be helpful for recommendation to all users. In this paper, we propose attentive Knowledge-aware Graph convolutional networks with Collaborative Guidance for personalized Recommendation (CG-KGR). CG-KGR is a novel knowledge-aware recommendation model that enables ample and coherent learning of KGs and user-item interactions, via our proposed Collaborative Guidance Mechanism. Specifically, CG-KGR first encapsulates historical interactions to interactive information summarization. Then CG-KGR utilizes it as guidance to extract information out of KGs, which eventually provides more precise personalized recommendation. We conduct extensive experiments on four real-world datasets over two recommendation tasks, i.e., Top-K recommendation and Click-Through rate (CTR) prediction. The experimental results show that the CG-KGR model significantly outperforms recent state-of-the-art models by 1.4-27.0% in terms of Recall metric on Top-K recommendation.

Session-based recommendation aims to predict a user's next action based on previous actions in the current session. The major challenge is to capture authentic and complete user preferences in the entire session. Recent work utilizes graph structure to represent the entire session and adopts Graph Neural Network (GNN) to encode session information. This modeling choice has been proved to be effective and achieved remarkable results. However, most of the existing studies only consider each item within the session independently and do not capture session semantics from a high-level perspective. Such limitation often leads to severe information loss and increases the difficulty of capturing long-range dependencies within a session. Intuitively, compared with individual items, a session snippet, i.e., a group of locally consecutive items, is able to provide supplemental user intents which are hardly captured by existing methods. In this work, we propose to learn multi-granularity consecutive user intent unit to improve the recommendation performance. Specifically, we creatively propose Multi-granularity Intent Heterogeneous Session Graph (MIHSG) which captures the interactions between different granularity intent units and relieves the burden of long-dependency. Moreover, we propose the Intent Fusion Ranking (IFR) module to compose the recommendation results from various granularity user intents. Compared with current methods that only leverage intents from individual items, IFR benefits from different granularity user intents to generate more accurate and comprehensive session representation, thus eventually boosting recommendation performance. We conduct extensive experiments on five session-based recommendation datasets and the results demonstrate the effectiveness of our method. Compared to current state-of-the-art methods, we achieve as large as 10.21% gain on [email protected] and 15.53% gain on [email protected]

Taxonomies are fundamental to many real-world applications in various domains, serving as structural representations of knowledge. To deal with the increasing volume of new concepts needed to be organized as taxonomies, researchers turn to automatically completion of an existing taxonomy with new concepts. In this paper, we propose TaxoEnrich, a new taxonomy completion framework, which effectively leverages both semantic features and structural information in the existing taxonomy and offers a better representation of candidate position to boost the performance of taxonomy completion. Specifically, TaxoEnrichconsists of four components: (1) taxonomy-contextualized embedding which incorporates both semantic meanings of concept and taxonomic relations based on powerful pretrained language models; (2) a taxonomy-aware sequential encoder which learns candidate position representations by encoding the structural information of taxonomy; (3) a query-aware sibling encoder which adaptively aggregates candidate siblings to augment candidate position representations based on their importance to the query-position matching; (4) a query-position matching model which extends existing work with our new candidate position representations. Extensive experiments on four large real-world datasets from different domains show that TaxoEnrichachieves the best performance among all evaluation metrics and outperforms previous state-of-the-art methods by a large margin.

With the emerging unlicensed spectrum sharing and Machine-to-Machine communications, the coexistence performance of multiple devices with different access schemes operating at the same unlicensed bands has received significant research interests. As the two most representative random-access schemes, Aloha and Carrier Sense Multiple Access (CSMA) both found wide applications in unlicensed bands. Yet most of the studies have focused on the coexistence of CSMA-based networks, leaving the coexistence of Aloha-based and CSMA-based networks largely unexplored. The challenge originates from the lack of a coexistence model of slotted Aloha and CSMA. In this paper, the throughput performance of coexisting Aloha and CSMA networks is characterized and optimized by extending a unified analytical framework proposed for random-access networks. The analysis shows that different from the single-network case where the maximum throughput of CSMA is much higher thanks to carrier sensing, when Aloha and CSMA coexist, Aloha would significantly outperform CSMA if each network optimizes its own throughput performance without cooperation, leading to poor throughput performance for CSMA and severe unfairness. To achieve fair coexistence, inter-network cooperation is crucial. The optimal transmission probabilities of Aloha and CSMA for maximizing the total network throughput under a given throughput ratio are further derived, and applied to coexisting LTE Unlicensed and WiFi networks to optimize their coexistence performance.

Automatically constructing taxonomy finds many applications in e-commerce and web search. One critical challenge is as data and business scope grow in real applications, new concepts are emerging and needed to be added to the existing taxonomy. Previous approaches focus on the taxonomy expansion, i.e. finding an appropriate hypernym concept from the taxonomy for a new query concept. In this paper, we formulate a new task, “taxonomy completion”, by discovering both the hypernym and hyponym concepts for a query. We propose Triplet Matching Network (TMN), to find the appropriate pairs for a given query concept. TMN consists of one primal scorer and multiple auxiliary scorers. These auxiliary scorers capture various fine-grained signals (e.g., query to hypernym or query to hyponym semantics), and the primal scorer makes a holistic prediction on triplet based on the internal feature representations of all auxiliary scorers. Also, an innovative channel-wise gating mechanism that retains task-specific information in concept representations is introduced to further boost model performance. Experiments on four real-world large-scale datasets show that TMN achieves the best performance on both taxonomy completion task and the previous taxonomy expansion task, outperforming existing methods.

Scientific literature analysis needs fine-grained named entity recognition (NER) to provide a wide range of information for scientific discovery.For example, chemistry research needs to study dozens to hundreds of distinct, fine-grained entity types, making consistent and accurate annotation difficult even for crowds of domain experts.On the other hand, domain-specific ontologies and knowledge bases (KBs) can be easily accessed, constructed, or integrated, which makes distant supervision realistic for fine-grained chemistry NER.In distant supervision, training labels are generated by matching mentions in a document with the concepts in the knowledge bases (KBs).However, this kind of KB-matching suffers from two major challenges: incomplete annotation and noisy annotation.We propose CHEMNER, an ontologyguided, distantly-supervised method for finegrained chemistry NER to tackle these challenges.It leverages the chemistry type ontology structure to generate distant labels with novel methods of flexible KB-matching and ontology-guided multi-type disambiguation.It significantly improves the distant label generation for the subsequent sequence labeling model training.We also provide an expertlabeled, chemistry NER dataset with 62 finegrained chemistry types (e.g., chemical compounds and chemical reactions).Experimental results show that CHEMNER is highly effective, outperforming substantially the stateof-the-art NER methods (with .25 absolute F1 score improvement).

Graph convolutional networks (GCNs), aiming to obtain node embeddings by integrating high-order neighborhood information through stacked graph convolution layers, have demonstrated great power in many network analysis tasks such as node classification and link prediction. However, a fundamental weakness of GCNs, that is, topological limitations, including over-smoothing and local homophily of topology, limits their ability to represent networks. Existing studies for solving these topological limitations typically focus only on the convolution of features on network topology, which inevitably relies heavily on network structure. Moreover, most networks are text-rich, so it is important to integrate not only document-level information, but also the local text information which is particularly significant while often ignored by the existing methods. To solve these limitations, we propose BiTe-GCN, a novel GCN architecture modeling via bidirectional convolution of topology and features on text-rich networks. Specifically, we first transform the original text-rich network into an augmented bi-typed heterogeneous network, capturing both the global document-level information and the local text-sequence information from texts. We then introduce discriminative convolution mechanisms, which performs convolution on this augmented bi-typed network, realizing the convolutions of topology and features altogether in the same system, and learning different contributions of these two parts (i.e., network part and text part), automatically for the given learning objectives. Extensive experiments on text-rich networks demonstrate that our new architecture outperforms the state-of-the-arts by a breakout improvement. Moreover, this architecture can also be applied to several e-commerce search scenes such as JD searching, and experiments on JD dataset show the superiority of the proposed architecture over the related methods.

Biomedical named entity recognition (BioNER) is a fundamental step for mining COVID-19 literature. Existing BioNER datasets cover a few common coarse-grained entity types (e.g., genes, chemicals, and diseases), which cannot be used to recognize highly domain-specific entity types (e.g., animal models of diseases) or emerging ones (e.g., coronaviruses) for COVID-19 studies. We present CORD-NER, a fine-grained named entity recognized dataset of COVID-19 literature (up until May 19, 2020). CORD-NER contains over 12 million sentences annotated via distant supervision. Also included in CORD-NER are 2,000 manually-curated sentences as a test set for performance evaluation. CORD-NER covers 75 fine-grained entity types. In addition to the common biomedical entity types, it covers new entity types specifically related to COVID-19 studies, such as coronaviruses, viral proteins, evolution, and immune responses. The dictionaries of these fine-grained entity types are collected from existing knowledge bases and human-input seed sets. We further present DISTNER, a distantly supervised NER model that relies on a massive unlabeled corpus and a collection of dictionaries to annotate the COVID-19 corpus. DISTNER provides a benchmark performance on the CORD-NER test set for future research.