Large transformer-based language models dominate modern NLP, yet our understanding of how they encode linguistic information relies primarily on studies of early models like BERT and GPT-2. We systematically probe 25 models from BERT Base to Qwen2.5-7B focusing on two linguistic properties: lexical identity and inflectional features across 6 diverse languages. We find a consistent pattern: inflectional features are linearly decodable throughout the model, while lexical identity is prominent early but increasingly weakens with depth. Further analysis of the representation geometry reveals that models with aggressive mid-layer dimensionality compression show reduced steering effectiveness in those layers, despite probe accuracy remaining high. Pretraining analysis shows that inflectional structure stabilizes early while lexical identity representations continue evolving. Taken together, our findings suggest that transformers maintain inflectional features across layers, while trading off lexical identity for compact, predictive representations. Our code is available at https://github.com/ml5885/model_internal_sleuthing

Nishant Subramani, Jason Eisner, Justin Svegliato, Benjamin Van Durme, Yu Su, Sam Thomson. 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.

Dirk Groeneveld, Iz Beltagy, Evan Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, Shane Arora, David Atkinson, Russell Authur, Khyathi Chandu, Arman Cohan, Jennifer Dumas, Yanai Elazar, Yuling Gu, Jack Hessel, Tushar Khot, William Merrill, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew Peters, Valentina Pyatkin, Abhilasha Ravichander, Dustin Schwenk, Saurabh Shah, William Smith, Emma Strubell, Nishant Subramani, Mitchell Wortsman, Pradeep Dasigi, Nathan Lambert, Kyle Richardson, Luke Zettlemoyer, Jesse Dodge, Kyle Lo, Luca Soldaini, Noah Smith, Hannaneh Hajishirzi. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Luca Soldaini, Rodney Kinney, Akshita Bhagia, Dustin Schwenk, David Atkinson, Russell Authur, Ben Bogin, Khyathi Chandu, Jennifer Dumas, Yanai Elazar, Valentin Hofmann, Ananya Jha, Sachin Kumar, Li Lucy, Xinxi Lyu, Nathan Lambert, Ian Magnusson, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew Peters, Abhilasha Ravichander, Kyle Richardson, Zejiang Shen, Emma Strubell, Nishant Subramani, Oyvind Tafjord, Evan Walsh, Luke Zettlemoyer, Noah Smith, Hannaneh Hajishirzi, Iz Beltagy, Dirk Groeneveld, Jesse Dodge, Kyle Lo. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Abstract With the success of large-scale pre-training and multilingual modeling in Natural Language Processing (NLP), recent years have seen a proliferation of large, Web-mined text datasets covering hundreds of languages. We manually audit the quality of 205 language-specific corpora released with five major public datasets (CCAligned, ParaCrawl, WikiMatrix, OSCAR, mC4). Lower-resource corpora have systematic issues: At least 15 corpora have no usable text, and a significant fraction contains less than 50% sentences of acceptable quality. In addition, many are mislabeled or use nonstandard/ambiguous language codes. We demonstrate that these issues are easy to detect even for non-proficient speakers, and supplement the human audit with automatic analyses. Finally, we recommend techniques to evaluate and improve multilingual corpora and discuss potential risks that come with low-quality data releases.

Synthetic speech or “fake speech” which matches personal vocal traits has become better and cheaper due to advances in deep learning-based speech synthesis and voice conversion approaches. This increased accessibility of synthetic speech systems and the growing misuse of them highlights the critical need to build countermeasures. Furthermore, new synthesis models evolve all the time and the efficacy of previously trained detection models on these unseen attack vectors is poor. In this paper, we focus on: 1) How can we build highly accurate, yet parameter and sample-efficient models for fake speech detection? 2) How can we rapidly adapt detection models to new sources of fake speech? We present four parameter-efficient convolutional architectures for fake speech detection with best detection F1 scores of around 97 points on a large dataset of fake and bonafide speech. We show how the fake speech detection task naturally lends itself to a novel multi-task problem further improving F1 scores for a mere 0.5% increase in model parameters. Our multi-task setting also helps in data-sparse situations, commonplace in adversarial settings. We investigate an alternative approach to the data-sparsity problem using transfer learning and show that it is possible to meet purely supervised detection performance for unseen attack vectors with as little as 6.25% of the training data. This is the first known application of transfer learning in adversarial settings for speech. Finally, we show how well our transfer learning approach adapts in an instance-efficient way to new attack vectors using the Real-Time Voice Cloning toolkit. We exceed the purely supervised detection performance (99.18 F1) with as little as 6.25% of the data.

Neural network-based generative language models like ELMo and BERT can work effectively as general purpose sentence encoders in text classification without further fine-tuning. Is it possible to adapt them in a similar way for use as general-purpose decoders? For this to be possible, it would need to be the case that for any target sentence of interest, there is some continuous representation that can be passed to the language model to cause it to reproduce that sentence. We set aside the difficult problem of designing an encoder that can produce such representations and, instead, ask directly whether such representations exist at all. To do this, we introduce a pair of effective, complementary methods for feeding representations into pretrained unconditional language models and a corresponding set of methods to map sentences into and out of this representation space, the reparametrized sentence space. We then investigate the conditions under which a language model can be made to generate a sentence through the identification of a point in such a space and find that it is possible to recover arbitrary sentences nearly perfectly with language models and representations of moderate size.

Causal graphs, such as directed acyclic graphs (DAGs) and partial ancestral graphs (PAGs), represent causal relationships among variables in a model. Methods exist for learning DAGs and PAGs from data and for converting DAGs to PAGs. However, these methods only output a single causal graph consistent with the independencies/dependencies (the Markov equivalence class M) estimated from the data. However, many distinct graphs may be consistent with M, and a data modeler may wish to select among these using domain knowledge. In this paper, we present a method that makes this possible. We introduce PAG2ADMG, the first method for enumerating all causal graphs consistent with M, under certain assumptions. PAG2ADMG converts a given PAG into a set of acyclic directed mixed graphs (ADMGs). We prove the correctness of the approach and demonstrate its efficiency relative to brute-force enumeration.