The Circuit Localization track of the Mechanistic Interpretability Benchmark (MIB) evaluates methods for localizing circuits within large language models (LLMs), i.e., subnetworks responsible for specific task behaviors. In this work, we investigate whether ensembling two or more circuit localization methods can improve performance. We explore two variants: parallel and sequential ensembling. In parallel ensembling, we combine attribution scores assigned to each edge by different methods—e.g., by averaging or taking the minimum or maximum value. In the sequential ensemble, we use edge attribution scores obtained via EAP-IG as a warm start for a more expensive but more precise circuit identification method, namely edge pruning. We observe that both approaches yield notable gains on the benchmark metrics, leading to a more precise circuit identification approach. Finally, we find that taking a parallel ensemble over various methods, including the sequential ensemble, achieves the best results. We evaluate our approach in the BlackboxNLP 2025 MIB Shared Task, comparing ensemble scores to official baselines across multiple model–task combinations.

In large language models (LLMs), certain neurons can store distinct pieces of knowledge learned during pretraining. While factual knowledge typically appears as a combination of relations and entities, it remains unclear whether some neurons focus on a relation itself – independent of any entity. We hypothesize such neurons detect a relation in the input text and guide generation involving such a relation. To investigate this, we study the LLama-2 family on a chosen set of relations, with a statistics-based method. Our experiments demonstrate the existence of relation-specific neurons. We measure the effect of selectively deactivating candidate neurons specific to relation r on the LLM’s ability to handle (1) facts involving relation r and (2) facts involving a different relation r' ≠ r. With respect to their capacity for encoding relation information, we give evidence for the following three properties of relation-specific neurons. (i) Neuron cumulativity. Multiple neurons jointly contribute to processing facts involving relation r, with no single neuron fully encoding a fact in r on its own. (ii) Neuron versatility. Neurons can be shared across multiple closely related as well as less related relations. In addition, some relation neurons transfer across languages. (iii) Neuron interference. Deactivating neurons specific to one relation can improve LLMs’ factual recall performance for facts of other relations. We make our code and data publicly available at https://github.com/cisnlp/relation-specific-neurons.

Understanding how large language models (LLMs) acquire and store factual knowledge is crucial for enhancing their interpretability, reliability, and efficiency. In this work, we analyze the evolution of factual knowledge representation in the OLMo-7B model by tracking the roles of its Attention Heads and Feed Forward Networks (FFNs) over training. We classify these components into four roles—general, entity, relation-answer, and fact-answer specific—and examine their stability and transitions. Our results show that LLMs initially depend on broad, general-purpose components, which later specialize as training progresses. Once the model reliably predicts answers, some components are repurposed, suggesting an adaptive learning process. Notably, answer-specific attention heads display the highest turnover, whereas FFNs remain stable, continually refining stored knowledge. These insights offer a mechanistic view of knowledge formation in LLMs and have implications for model pruning, optimization, and transparency.

Current approaches to automatic summarization of scientific papers generate informative summaries in the form of abstracts. However, abstracts are not intended to show the relationship between a paper and the references cited in it. We propose a new contextualized summarization approach that can generate an informative summary conditioned on a given sentence containing the citation of a reference (a so-called “citance”). This summary outlines content of the cited paper relevant to the citation location. Thus, our approach extracts and models the citances of a paper, retrieves relevant passages from cited papers, and generates abstractive summaries tailored to each citance. We evaluate our approach using **Webis-Context-SciSumm-2023**, a new dataset containing 540K computer science papers and 4.6M citances therein.

To ease the difficulty of argument stance classification, the task of same side stance classification (S3C) has been proposed. In contrast to actual stance classification, which requires a substantial amount of domain knowledge to identify whether an argument is in favor or against a certain issue, it is argued that, for S3C, only argument similarity within stances needs to be learned to successfully solve the task. We evaluate several transformer-based approaches on the dataset of the recent S3C shared task, followed by an in-depth evaluation and error analysis of our model and the task’s hypothesis. We show that, although we achieve state-of-the-art results, our model fails to generalize both within as well as across topics and domains when adjusting the sampling strategy of the training and test set to a more adversarial scenario. Our evaluation shows that current state-of-the-art approaches cannot determine same side stance by considering only domain-independent linguistic similarity features, but appear to require domain knowledge and semantic inference, too.

We introduce and study a problem variant of sentiment analysis, namely the “same sentiment classification problem”, where, given a pair of texts, the task is to determine if they have the same sentiment, disregarding the actual sentiment polarity. Among other things, our goal is to enable a more topic-agnostic sentiment classification. We study the problem using the Yelp business review dataset, demonstrating how sentiment data needs to be prepared for this task, and then carry out sequence pair classification using the BERT language model. In a series of experiments, we achieve an accuracy above 83% for category subsets across topics, and 89% on average.