Understanding why large language models (LLMs) exhibit certain behaviors is the goal of mechanistic interpretability. One of the major tools employed by mechanistic interpretability is circuit discovery, i.e., identifying a subset of the model’s components responsible for a given task. We present a novel circuit discovery technique called IPE (Isolating Path Effects) that, unlike traditional edge-centric approaches, aims to identify entire computational paths (from input embeddings to output logits) responsible for certain model behaviors. Our method modifies the messages passed between nodes along a given path in such a way as to either precisely remove the effects of the entire path (i.e., ablate it) or to replace the path’s effects with those that would have been generated by a counterfactual input. IPE is different from current path-patching or edge activation-patching techniques since they are not ablating single paths, but rather a set of paths sharing certain edges, preventing more precise tracing of information flow. We apply our method to the well-known Indirect Object Identification (IOI) task, recovering the canonical circuit reported in prior work. On the MIB workshop leaderboard, we tested IOI and MCQA tasks on GPT2-small and Qwen2.5. For GPT2, path counterfactual replacement outperformed path ablation as expected and led to top-ranking results, while for Qwen, no significant differences were observed, indicating a need for larger experiments to distinguish the two approaches.

Comparing NLP systems to select the best one for a task of interest, such as named entity recognition, is critical for practitioners and researchers. A rigorous approach involves setting up a hypothesis testing scenario using the performance of the systems on query documents. However, often the hypothesis testing approach needs to send a lot of document queries to the systems, which can be problematic. In this paper, we present an effective alternative based on the multi-armed bandit (MAB). We propose a hierarchical generative model to represent the uncertainty in the performance measures of the competing systems, to be used by Thompson Sampling to solve the resulting MAB. Experimental results on both synthetic and real data show that our approach requires significantly fewer queries compared to the standard benchmarking technique to identify the best system according to F-measure.

Parallel corpora are often injected with bilingual lexical resources for improved Indian language machine translation (MT). In absence of such lexical resources, multilingual topic models have been used to create coarse lexical resources in the past, using a Cartesian product approach. Our results show that for morphologically rich languages like Hindi, the Cartesian product approach is detrimental for MT. We then present a novel ‘sentential’ approach to use this coarse lexical resource from a multilingual topic model. Our coarse lexical resource when injected with a parallel corpus outperforms a system trained using parallel corpus and a good quality lexical resource. As demonstrated by the quality of our coarse lexical resource and its benefit to MT, we believe that our sentential approach to create such a resource will help MT for resource-constrained languages.

Topic Models have been reported to be beneficial for aspect-based sentiment analysis. This paper reports the first topic model for sarcasm detection, to the best of our knowledge. Designed on the basis of the intuition that sarcastic tweets are likely to have a mixture of words of both sentiments as against tweets with literal sentiment (either positive or negative), our hierarchical topic model discovers sarcasm-prevalent topics and topic-level sentiment. Using a dataset of tweets labeled using hashtags, the model estimates topic-level, and sentiment-level distributions. Our evaluation shows that topics such as ‘work’, ‘gun laws’, ‘weather’ are sarcasm-prevalent topics. Our model is also able to discover the mixture of sentiment-bearing words that exist in a text of a given sentiment-related label. Finally, we apply our model to predict sarcasm in tweets. We outperform two prior work based on statistical classifiers with specific features, by around 25%.