Maria Teleki

Techniques that enhance inference through increased computation at test-time have recently gained attention. In this survey, we investigate the current state of LLM Inference-Time Self-Improvement from three different perspectives: Independent Self-improvement, focusing on enhancements via decoding or sampling methods; Context-Aware Self-Improvement, leveraging additional context or datastore; and Model-Aided Self-Improvement, achieving improvement through model collaboration. We provide a comprehensive review of recent relevant studies, contribute an in-depth taxonomy, and discuss challenges and limitations, offering insights for future research.

@inproceedings{dong24_survey,
title = {A Survey on LLM Inference-Time Self-Improvement},
author = {Xiangjue Dong and Maria Teleki and James Caverlee},
year = {2024},
booktitle = {arXiv}
}

Masculine defaults are widely recognized as a significant type of gender bias, but they are often unseen as they are under-researched (Cheryan and Markus 2020). Masculine defaults involve three key parts: (i) the cultural context, (ii) the masculine characteristics or behaviors, and (iii) the reward for, or simply acceptance of, those masculine characteristics or behaviors. In this work, we study discourse-based masculine defaults, and propose a twofold framework for (i) the large-scale discovery and analysis of gendered discourse words in spoken content via our Gendered Discourse Correlation Framework (GDCF); and (ii) the measurement of the gender bias associated with these gendered discourse words in LLMs via our Discourse Word-Embedding Association Test (D-WEAT). We focus our study on podcasts, a popular and growing form of social media, analyzing 15,117 podcast episodes. We analyze correlations between gender and discourse words -- discovered via LDA and BERTopic -- to automatically form gendered discourse word lists. We then study the prevalence of these gendered discourse words in domain-specific contexts, and find that gendered discourse-based masculine defaults exist in the domains of business, technology/politics, and video games. Next, we study the representation of these gendered discourse words from a state-of-the-art LLM embedding model from OpenAI, and find that the masculine discourse words have a more stable and robust representation than the feminine discourse words, which may result in better system performance on downstream tasks for men. Hence, men are rewarded for their discourse patterns with better system performance by one of the state-of-the-art language models -- and this embedding disparity constitutes a representational harm and a masculine default.

@inproceedings{teleki25_icwsm,
title = {Masculine Defaults via Gendered Discourse in Podcasts and Large Language Models},
author = {Maria Teleki and Xiangjue Dong and Haoran Liu and James Caverlee},
year = {2025},
booktitle = {ICWSM 2025}
}

In this work, we evaluate the disfluency capabilities of two automatic speech recognition systems -- Google ASR and WhisperX -- through a study of 10 human-annotated podcast episodes and a larger set of 82,601 podcast episodes. We employ a state-of-the-art disfluency annotation model to perform a fine-grained analysis of the disfluencies in both the scripted and non-scripted podcasts. We find, on the set of 10 podcasts, that while WhisperX overall tends to perform better, Google ASR outperforms in WIL and BLEU scores for non-scripted podcasts. We also find that Google ASR's transcripts tend to contain closer to the ground truth number of edited-type disfluent nodes, while WhisperX's transcripts are closer for interjection-type disfluent nodes. This same pattern is present in the larger set. Our findings have implications for the choice of an ASR model when building a larger system, as the choice should be made depending on the distribution of disfluent nodes present in the data.

@inproceedings{teleki24_interspeech,
title = {Comparing ASR Systems in the Context of Speech Disfluencies},
author = {Maria Teleki and Xiangjue Dong and Soohwan Kim and James Caverlee},
year = {2024},
booktitle = {Interspeech 2024},
pages = {4548--4552},
doi = {10.21437/Interspeech.2024-1270},
}

Spoken content is abundant -- including podcasts, meeting transcripts, and TikTok-like short videos. And yet, many important tasks like summarization are often designed for written content rather than the looser, noiser, and more disfluent style of spoken content. Hence, we aim in this paper to quantify the impact of disfluency on spoken content summarization. Do disfluencies negatively impact the quality of summaries generated by existing approaches? And if so, to what degree? Coupled with these goals, we also investigate two methods towards improving summarization in the presence of such disfluencies. We find that summarization quality does degrade with an increase in these disfluencies and that a combination of multiple disfluency types leads to even greater degradation. Further, our experimental results show that naively removing disfluencies and augmenting with special tags can worsen the summarization when used for testing, but that removing disfluencies for fine-tuning yields the best results. We make the code available at https://github.com/mariateleki/Quantifying-Impact-Disfluency.

@inproceedings{teleki-etal-2024-quantifying-impact,
title = "Quantifying the Impact of Disfluency on Spoken Content Summarization",
author = "Teleki, Maria and
Dong, Xiangjue and
Caverlee, James",
editor = "Calzolari, Nicoletta and
Kan, Min-Yen and
Hoste, Veronique and
Lenci, Alessandro and
Sakti, Sakriani and
Xue, Nianwen",
booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
month = may,
year = "2024",
address = "Torino, Italia",
publisher = "ELRA and ICCL",
url = "https://aclanthology.org/2024.lrec-main.1175",
pages = "13419--13428",
abstract = "Spoken content is abundant {--} including podcasts, meeting transcripts, and TikTok-like short videos. And yet, many important tasks like summarization are often designed for written content rather than the looser, noiser, and more disfluent style of spoken content. Hence, we aim in this paper to quantify the impact of disfluency on spoken content summarization. Do disfluencies negatively impact the quality of summaries generated by existing approaches? And if so, to what degree? Coupled with these goals, we also investigate two methods towards improving summarization in the presence of such disfluencies. We find that summarization quality does degrade with an increase in these disfluencies and that a combination of multiple disfluency types leads to even greater degradation. Further, our experimental results show that naively removing disfluencies and augmenting with special tags can worsen the summarization when used for testing, but that removing disfluencies for fine-tuning yields the best results. We make the code available at https://github.com/mariateleki/Quantifying-Impact-Disfluency.",
}

Voice-driven software systems are in abundance. However, language models that power these systems are traditionally trained on fluent, written text corpora. Hence there can be a misalignment between the inherent disfluency of transcribed spoken content and the fluency of the written training data. Furthermore, gold-standard disfluency annotations of various complexities for incremental training can be expensive to collect. So, we propose in this paper a Disfluency Augmented Curriculum Learning (DACL) approach to tackle the complex structure of disfluent sentences and generate fluent texts from them, by using Curriculum Learning (CL) coupled with our synthetically augmented disfluent texts of various levels. DACL harnesses the tiered structure of our generated synthetic disfluent data using CL, by training the model on basic samples (i.e. more fluent) first before training it on more complex samples (i.e. more disfluent). In contrast to the random data exposure paradigm, DACL focuses on a simple-to-complex learning process. We comprehensively evaluate DACL on Switchboard Penn Treebank-3 and compare it to the state-of-the-art disfluency removal models. Our model surpasses existing techniques in word-based precision (by up to 1%) and has shown favorable recall and F1 scores.

@inproceedings{chaudhury-etal-2024-dacl-disfluency,
title = "{DACL}: Disfluency Augmented Curriculum Learning for Fluent Text Generation",
author = "Chaudhury, Rohan and
Teleki, Maria and
Dong, Xiangjue and
Caverlee, James",
editor = "Calzolari, Nicoletta and
Kan, Min-Yen and
Hoste, Veronique and
Lenci, Alessandro and
Sakti, Sakriani and
Xue, Nianwen",
booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
month = may,
year = "2024",
address = "Torino, Italia",
publisher = "ELRA and ICCL",
url = "https://aclanthology.org/2024.lrec-main.385",
pages = "4311--4321",
abstract = "Voice-driven software systems are in abundance. However, language models that power these systems are traditionally trained on fluent, written text corpora. Hence there can be a misalignment between the inherent disfluency of transcribed spoken content and the fluency of the written training data. Furthermore, gold-standard disfluency annotations of various complexities for incremental training can be expensive to collect. So, we propose in this paper a Disfluency Augmented Curriculum Learning (DACL) approach to tackle the complex structure of disfluent sentences and generate fluent texts from them, by using Curriculum Learning (CL) coupled with our synthetically augmented disfluent texts of various levels. DACL harnesses the tiered structure of our generated synthetic disfluent data using CL, by training the model on basic samples (i.e. more fluent) first before training it on more complex samples (i.e. more disfluent). In contrast to the random data exposure paradigm, DACL focuses on a simple-to-complex learning process. We comprehensively evaluate DACL on Switchboard Penn Treebank-3 and compare it to the state-of-the-art disfluency removal models. Our model surpasses existing techniques in word-based precision (by up to 1{\%}) and has shown favorable recall and F1 scores.",
}

Pre-trained Language Models are widely used in many important real-world applications. However, recent studies show that these models can encode social biases from large pre-training corpora and even amplify biases in downstream applications. To address this challenge, we propose Co2PT, an efficient and effective debias-while-prompt tuning method for mitigating biases via counterfactual contrastive prompt tuning on downstream tasks. Our experiments conducted on three extrinsic bias benchmarks demonstrate the effectiveness of Co2PT on bias mitigation during the prompt tuning process and its adaptability to existing upstream debiased language models. These findings indicate the strength of Co2PT and provide promising avenues for further enhancement in bias mitigation on downstream tasks.

@inproceedings{dong-etal-2023-co2pt,
title = "{C}o$^2${PT}: Mitigating Bias in Pre-trained Language Models through Counterfactual Contrastive Prompt Tuning",
author = "Dong, Xiangjue and
Zhu, Ziwei and
Wang, Zhuoer and
Teleki, Maria and
Caverlee, James",
editor = "Bouamor, Houda and
Pino, Juan and
Bali, Kalika",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.findings-emnlp.390",
doi = "10.18653/v1/2023.findings-emnlp.390",
pages = "5859--5871",
abstract = "Pre-trained Language Models are widely used in many important real-world applications. However, recent studies show that these models can encode social biases from large pre-training corpora and even amplify biases in downstream applications. To address this challenge, we propose Co$^2$PT, an efficient and effective *debias-while-prompt tuning* method for mitigating biases via counterfactual contrastive prompt tuning on downstream tasks. Our experiments conducted on three extrinsic bias benchmarks demonstrate the effectiveness of Co$^2$PT on bias mitigation during the prompt tuning process and its adaptability to existing upstream debiased language models. These findings indicate the strength of Co$^2$PT and provide promising avenues for further enhancement in bias mitigation on downstream tasks.", }

In this paper, we present Howdy Y’all, a multi-modal task-oriented dialogue agent developed for the 2021-2022 Alexa Prize TaskBot competition. Our design principles guiding Howdy Y’all aim for high user satisfaction through friendly and trustworthy encounters, minimization of negative conversation edge cases, and wide coverage over many tasks. Hence, Howdy Y’all is built upon a rapid prototyping platform to enable fast experimentation and powered by four key innovations to enable this vision: (i) First, it combines a rules, phonetic matching, and a transformer-based approach for robust intent understanding. (ii) Second, to accurately elicit user preferences and guide users to the right task, Howdy Y’all is powered by a contrastive learning search framework over sentence embeddings and a conversational recommender for eliciting preferences. (iii) Third, to support a variety of user question types, it introduces a new data augmentation method for question generation and a self-supervised answer selection approach for improving question answering. (iv) Finally, to help motivate our users and keep them engaged, we design an emotional conversation tracker that provides empathetic responses to keep users engaged and a monitor of conversation quality.

@inproceedings{University2022,
author={Alfifi, Majid and Dong, Xiangjue and Feldman, Timo and Lin, Allen and Madanagopal, Karthic and Pethe, Aditya and Teleki, Maria and Wang, Zhuoer and Zhu, Ziwei and Caverlee, James},
title = {Howdy Y’all: An Alexa TaskBot},
year = {2022},
url = {https://www.amazon.science/alexa-prize/proceedings/howdy-yall-an-alexa-taskbot},
booktitle = {Alexa Prize TaskBot Challenge Proceedings},
}