Transformer-based large language models (LLMs) have successfully handled various tasks.

Another possible reason is that it is unclear if the low signal-to-noise ratio of neuroimaging tools such as functional Magnetic Resonance Imaging (fMRI) can allow us to reveal the correlates of complex (and perhaps subtle) syntactic representations.

Translating the internal representations and computations of models into concepts that humans can understand is a key goal of interpretability. While recent dictionary learning methods such as Sparse Autoencoders (SAEs) provide a promising route to discover human-interpretable features, they suffer from a variety of problems, including a systematic failure to capture the rich conceptual information that drives linguistic understanding. Instead, they exhibit a bias towards shallow, token-specific, or noisy features, such as "the phrase 'The' at the start of sentences". In this work, we propose that this is due to a fundamental issue with how dictionary learning methods for LLMs are trained. Language itself has a rich, well-studied structure spanning syntax, semantics, and pragmatics; however, current unsupervised methods largely ignore this linguistic knowledge, leading to poor feature discovery that favors superficial patterns over meaningful concepts. We focus on a simple but important aspect of language: semantic content has long-range dependencies and tends to be smooth over a sequence, whereas syntactic information is much more local. Building on this insight, we introduce Temporal Sparse Autoencoders (T-SAEs), which incorporate a novel contrastive loss encouraging consistent activations of high-level features over adjacent tokens. This simple yet powerful modification enables SAEs to disentangle semantic from syntactic features in a self-supervised manner. Across multiple datasets and models, T-SAEs recover smoother, more coherent semantic concepts without sacrificing reconstruction quality. Strikingly, they exhibit clear semantic structure despite being trained without explicit semantic signal, offering a new pathway for unsupervised interpretability in language models.

Transformer-based large language models (LLMs) have successfully handled various tasks.

Syntactic bootstrapping (Gleitman, 1990) is the hypothesis that children use the syntactic environments in which a verb occurs to learn its meaning. In this paper, we examine whether large language models exhibit a similar behavior. We do this by training RoBERTa and GPT-2 on perturbed datasets where syntactic information is ablated. Our results show that models' verb representation degrades more when syntactic cues are removed than when co-occurrence information is removed. Furthermore, the representation of mental verbs, for which syntactic bootstrapping has been shown to be particularly crucial in human verb learning, is more negatively impacted in such training regimes than physical verbs. In contrast, models' representation of nouns is affected more when co-occurrences are distorted than when syntax is distorted. In addition to reinforcing the important role of syntactic bootstrapping in verb learning, our results demonstrated the viability of testing developmental hypotheses on a larger scale through manipulating the learning environments of large language models.

Recent work has demonstrated that neural language models encode syntactic structures in their internal representations, yet the derivations by which these structures are constructed across layers remain poorly understood. In this paper, we propose Derivational Probing to investigate how micro-syntactic structures (e.g., subject noun phrases) and macro-syntactic structures (e.g., the relationship between the root verbs and their direct dependents) are constructed as word embeddings propagate upward across layers. Our experiments on BERT reveal a clear bottom-up derivation: micro-syntactic structures emerge in lower layers and are gradually integrated into a coherent macro-syntactic structure in higher layers. Furthermore, a targeted evaluation on subject-verb number agreement shows that the timing of constructing macro-syntactic structures is critical for downstream performance, suggesting an optimal timing for integrating global syntactic information.

Transformer-based large language models (LLMs) have successfully handled various tasks. Specifically, rotary position embedding (RoPE), one of the most widely used techniques, encodes the positional information by dividing the query or key value with d elements into d/2 pairs and rotating the 2d vectors corresponding to each pair of elements. Therefore, the direction of each pair and the position-related rotation jointly determine the attention score. In this paper, we show that the direction of the 2d pair is largely affected by the angle between the corresponding weight vector pair. We theoretically show that non-orthogonal weight vector pairs lead to great attention on tokens at a certain relative position and are less sensitive to the input which may correspond to basic syntactic information.

Aspect-based Sentiment Analysis (ABSA) is the task aimed at predicting the sentiment polarity of aspect words within sentences. Recently, incorporating graph neural networks (GNNs) to capture additional syntactic structure information in the dependency tree derived from syntactic dependency parsing has been proven to be an effective paradigm for boosting ABSA. Despite GNNs enhancing model capability by fusing more types of information, most works only utilize a single topology view of the dependency tree or simply conflate different perspectives of information without distinction, which limits the model performance. To address these challenges, in this paper, we propose a new multi-view attention syntactic enhanced graph convolutional network (MASGCN) that weighs different syntactic information of views using attention mechanisms. Specifically, we first construct distance mask matrices from the dependency tree to obtain multiple subgraph views for GNNs. To aggregate features from different views, we propose a multi-view attention mechanism to calculate the attention weights of views. Furthermore, to incorporate more syntactic information, we fuse the dependency type information matrix into the adjacency matrices and present a structural entropy loss to learn the dependency type adjacency matrix. Comprehensive experiments on four benchmark datasets demonstrate that our model outperforms state-of-the-art methods. The codes and datasets are available at https://github.com/SELGroup/MASGCN.