Ambiguity is pervasive in real-world questions, yet large language models (LLMs) often respond with confident answers rather than seeking clarification. In this work, we show that question ambiguity is linearly encoded in the internal representations of LLMs and can be both detected and controlled at the neuron level. During the model’s pre-filling stage, we identify that a small number of neurons, as few as one, encode question ambiguity information. Probes trained on these Ambiguity-Encoding Neurons (AENs) achieve strong performance on ambiguity detection and generalize across datasets, outperforming prompting-based and representation-based baselines. Layerwise analysis reveals that AENs emerge from shallow layers, suggesting early encoding of ambiguity signals in the model’s processing pipeline. Finally, we show that through manipulating AENs, we can control LLM’s behavior from direct answering to abstention. Our findings reveal that LLMs form compact internal representations of question ambiguity, enabling interpretable and controllable behavior.

Jinhao Duan, Xinyu Zhao, Zhuoxuan Zhang, Eunhye Grace Ko, Lily Boddy, Chenan Wang, Tianhao Li, Alexander Rasgon, Junyuan Hong, Min Kyung Lee, Chenxi Yuan, Qi Long, Ying Ding, Tianlong Chen, Kaidi Xu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Identification of targets among known drugs plays an important role in drug repurposing and discovery. Computational approaches for prediction of drug-target interactions (DTIs)are highly desired in comparison to traditional biological experiments as its fast and low price. Moreover, recent advances of systems biology approaches have generated large-scale heterogeneous, biological information networks data, which offer opportunities for machine learning-based identification of DTIs. We present a novel Inductive Matrix Completion with Heterogeneous Graph Attention Network approach (IMCHGAN)for predicting DTIs. IMCHGAN first adopts a two-level neural attention mechanism approach to learn drug and target latent feature representations from the DTI heterogeneous network respectively. Then, the learned latent features are fed into the Inductive Matrix Completion (IMC)prediction score model which computes the best projection from drug space onto target space and output DTI score via the inner product of projected drug and target feature representations. IMCHGAN is an end-to-end neural network learning framework where the parameters of both the prediction score model and the feature representation learning model are simultaneously optimized via backpropagation under supervising of the observed known drug-target interactions data. We compare IMCHGAN with other state-of-the-art baselines on two real DTI experimental datasets. The results show that our method is superior to existing methods in term of AUC and AUPR. Moreover, IMCHGAN also shows it has strong predictive power for novel (unknown)DTIs. All datasets and code can be obtained from https://github.com/ljatynu/IMCHGAN/.

MOTIVATION: Predicting the association between microRNAs (miRNAs) and diseases plays an import role in identifying human disease-related miRNAs. As identification of miRNA-disease associations via biological experiments is time-consuming and expensive, computational methods are currently used as effective complements to determine the potential associations between disease and miRNA. RESULTS: We present a novel method of neural inductive matrix completion with graph convolutional network (NIMCGCN) for predicting miRNA-disease association. NIMCGCN first uses graph convolutional networks to learn miRNA and disease latent feature representations from the miRNA and disease similarity networks. Then, learned features were input into a novel neural inductive matrix completion (NIMC) model to generate an association matrix completion. The parameters of NIMCGCN were learned based on the known miRNA-disease association data in a supervised end-to-end way. We compared the proposed method with other state-of-the-art methods. The area under the receiver operating characteristic curve results showed that our method is significantly superior to existing methods. Furthermore, 50, 47 and 48 of the top 50 predicted miRNAs for three high-risk human diseases, namely, colon cancer, lymphoma and kidney cancer, were verified using experimental literature. Finally, 100% prediction accuracy was achieved when breast cancer was used as a case study to evaluate the ability of NIMCGCN for predicting a new disease without any known related miRNAs. AVAILABILITY AND IMPLEMENTATION: https://github.com/ljatynu/NIMCGCN/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.