We investigate how Large Language Models (LLMs) distinguish between memorization and generalization at the neuron level. Through carefully designed tasks, we identify distinct neuron subsets responsible for each behavior. Experiments on both a GPT-2 model trained from scratch and a pretrained LLaMA-3.2 model fine-tuned with LoRA show consistent neuron-level specialization. We further demonstrate that inference-time interventions on these neurons can steer the model’s behavior toward memorization or generalization. To assess robustness, we evaluate intra-task and inter-task consistency, confirming that these neuron-behavior associations reflect generalizable patterns rather than dataset-specific artifacts. Our findings reveal modular structure in LLMs and enable controlling memorization and generalization behaviors at inference time.

This paper addresses the challenge of handling unseen modalities and dynamic modality combinations at test time with our proposed text-centric alignment method. This training-free alignment approach unifies different input modalities into a single semantic text representation by leveraging in-context learning with Large Language Models and uni-modal foundation models. Our method significantly enhances the ability to manage unseen, diverse, and unpredictable modality combinations, making it suitable for both generative and discriminative models to adopt on top. Our extensive experiments primarily evaluate on discriminative tasks, demonstrating that our approach is essential for LLMs to achieve strong modality alignment performance. It also surpasses the limitations of traditional fixed-modality frameworks in embedding representations. This study contributes to the field by offering a flexible and effective solution for real-world applications where modality availability is dynamic and uncertain.

LLM search methods, such as Chain of Thought (CoT) and Tree of Thought (ToT), enhance LLM reasoning by exploring multiple reasoning paths. When combined with search algorithms like MCTS and Bandit methods, their effectiveness relies heavily on uncertainty estimation to prioritize paths that align with specific search objectives. However, it remains unclear whether existing LLM uncertainty metrics adequately capture the diverse types of uncertainty required to guide different search objectives. In this work, we introduce a framework for uncertainty benchmarking, identifying four distinct uncertainty types: Answer, Correctness, Aleatoric, and Epistemic Uncertainty. Each type serves different optimization goals in search. Our experiments demonstrate that current metrics often align with only a subset of these uncertainty types, limiting their effectiveness for objective-aligned search in some cases. These findings highlight the need for additional target-aware uncertainty estimators that can adapt to various optimization goals in LLM search.

This study investigates the privacy risks associated with text embeddings, focusing on the scenario where attackers cannot access the original embedding model. Contrary to previous research requiring direct model access, we explore a more realistic threat model by developing a transfer attack method. This approach uses a surrogate model to mimic the victim model’s behavior, allowing the attacker to infer sensitive information from text embeddings without direct access. Our experiments across various embedding models and a clinical dataset demonstrate that our transfer attack significantly outperforms traditional methods, revealing the potential privacy vulnerabilities in embedding technologies and emphasizing the need for enhanced security measures.

Disentangled representations have attracted increasing attention recently. However, how to transfer the desired properties of disentanglement to word representations is unclear. In this work, we propose to transform typical dense word vectors into disentangled embeddings featuring improved interpretability via encoding polysemous semantics separately. We also found the modular structure of our disentangled word embeddings helps generate more efficient and effective features for natural language processing tasks.

Chinese NLP applications that rely on large text often contain huge amounts of vocabulary which are sparse in corpus. We show that characters’ written form, Glyphs, in ideographic languages could carry rich semantics. We present a multi-modal model, Glyph2Vec, to tackle Chinese out-of-vocabulary word embedding problem. Glyph2Vec extracts visual features from word glyphs to expand current word embedding space for out-of-vocabulary word embedding, without the need of accessing any corpus, which is useful for improving Chinese NLP systems, especially for low-resource scenarios. Experiments across different applications show the significant effectiveness of our model.

We focus on a recently deployed system built for summarizing academic articles by concept tagging. The system has shown great coverage and high accuracy of concept identification which could be contributed by the knowledge acquired from millions of publications. Provided with the interpretable concepts and knowledge encoded in a pre-trained neural model, we investigate whether the tagged concepts can be applied to a broader class of applications. We propose transforming the tagged concepts into sparse vectors as representations of academic documents. The effectiveness of the representations is analyzed theoretically by a proposed framework. We also empirically show that the representations can have advantages on academic topic discovery and paper recommendation. On these applications, we reveal that the knowledge encoded in the tagging system can be effectively utilized and can help infer additional features from data with limited information.

The objective of non-parallel text style transfer, or controllable text generation, is to alter specific attributes (e.g. sentiment, mood, tense, politeness, etc) of a given text while preserving its remaining attributes and content. Generative adversarial network (GAN) is a popular model to ensure the transferred sentences are realistic and have the desired target styles. However, training GAN often suffers from mode collapse problem, which causes that the transferred text is little related to the original text. In this paper, we propose a new GAN model with a word-level conditional architecture and a two-phase training procedure. By using a style-related condition architecture before generating a word, our model is able to maintain style-unrelated words while changing the others. By separating the training procedure into reconstruction and transfer phases, our model is able to learn a proper text generation process, which further improves the content preservation. We test our model on polarity sentiment transfer and multiple-attribute transfer tasks. The empirical results show that our model achieves comparable evaluation scores in both transfer accuracy and fluency but significantly outperforms other state-of-the-art models in content compatibility on three real-world datasets.

An acrostic is a form of writing that the first token of each line (or other recurring features in the text) forms a meaningful sequence. In this paper we present a generalized acrostic generation system that can hide certain message in a flexible pattern specified by the users. Different from previous works that focus on rule-based solutions, here we adopt a neural- based sequence-to-sequence model to achieve this goal. Besides acrostic, users are also allowed to specify the rhyme and length of the output sequences. Based on our knowledge, this is the first neural-based natural language generation system that demonstrates the capability of performing micro-level control over output sentences.

Unsupervised document representation learning is an important task providing pre-trained features for NLP applications. Unlike most previous work which learn the embedding based on self-prediction of the surface of text, we explicitly exploit the inter-document information and directly model the relations of documents in embedding space with a discriminative network and a novel objective. Extensive experiments on both small and large public datasets show the competitiveness of the proposed method. In evaluations on standard document classification, our model has errors that are 5 to 13% lower than state-of-the-art unsupervised embedding models. The reduction in error is even more pronounced in scarce label setting.

Lexicon relation extraction given distributional representation of words is an important topic in NLP. We observe that the state-of-the-art projection-based methods cannot be generalized to handle unseen hypernyms. We propose to analyze it in the perspective of pollution and construct the corresponding indicator to measure it. We propose a word relation autoencoder (WRAE) model to address the challenge. Experiments on several hypernym-like lexicon datasets show that our model outperforms the competitors significantly.