Conference & Journal Articles

2025

1. NAACL 2025

   Analysis of Augmenting Large Language Models with Tools in Zero-Shot

   Ne Luo, Aryo Pradipta Gema, Xuanli He, Emile Krieken,  Pietro Lesci, and Pasquale Minervini

   Accepted at NAACL 2025. Nov 2025.

   Abs

   While Large Language Models (LLMs) show remarkable performance in various natural language processing tasks, they remain prone to factual inaccuracies and computational errors, including hallucinations and mistakes in mathematical reasoning. To mitigate these limitations, we explore approaches for teaching LLMs to effectively use external tools, including prompting and fine-tuning. Our approach begins with a zero-shot prompting strategy to guide LLMs in tool utilisation. We then curate a tool-using dataset using the LLM itself and fine-tune the model through supervised learning or preference-based techniques. Experiments across four challenging tasks show that tool use enhances performance on complex, long-tail knowledge tasks like PopQA. However, our method sometimes improperly calls the external tools, leading to degraded performance on other datasets. Our findings highlight the potential and challenges of integrating external tool use in LLMs.

2. ICLR 2025

   PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs

   Pietro Lesci*, Oskar van der Wal*, Max Müller-Eberstein, Naomi Saphra, Hailey Schoelkopf, Willem Zuidema, and Stella Biderman

   Accepted at ICLR 2025. Nov 2025.

   Abs

   Understanding the stability of language model pre-training and its effects on downstream performance is still understudied. Prior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed. Crucially, resources to study pre-training stability in language models are still lacking, especially for decoder-only models. We introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release. Using these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed—i.e., parameters’ initialisation and data order—on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases. In addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions. Additionally, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics. Our findings show the potential of using these methods to predict training stability.

2024

1. EMNLP 2024

   Tending Towards Stability: Convergence Challenges in Small Language Models

   Richard Diehl Martinez,  Pietro Lesci, and Paula Buttery

   In Findings of the Association for Computational Linguistics: EMNLP 2024. Nov 2024.

   Abs arXiv Bib Code

   Increasing the number of parameters in language models is a common strategy to enhance their performance. However, smaller language models remain valuable due to their lower operational costs. Despite their advantages, smaller models frequently underperform compared to their larger counterparts, even when provided with equivalent data and computational resources. Specifically, their performance tends to degrade in the late pretraining phase. This is anecdotally attributed to their reduced representational capacity. Yet, the exact causes of this performance degradation remain unclear. We use the Pythia model suite to analyse the training dynamics that underlie this phenomenon. Across different model sizes, we investigate the convergence of the Attention and MLP activations to their final state and examine how the effective rank of their parameters influences this process. We find that nearly all layers in larger models stabilise early in training - within the first 20% - whereas layers in smaller models exhibit slower and less stable convergence, especially when their parameters have lower effective rank. By linking the convergence of layers’ activations to their parameters’ effective rank, our analyses can guide future work to address inefficiencies in the learning dynamics of small models.

   @inproceedings{diehl-martinez-etal-2024-tending,
     address = {Miami, Florida, USA},
     author = {Diehl Martinez, Richard and Lesci, Pietro and Buttery, Paula},
     booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2024},
     editor = {Al-Onaizan, Yaser and Bansal, Mohit and Chen, Yun-Nung},
     month = nov,
     pages = {3275--3286},
     publisher = {Association for Computational Linguistics},
     title = {Tending Towards Stability: Convergence Challenges in Small Language Models},
     url = {https://aclanthology.org/2024.findings-emnlp.187},
     year = {2024}
   }

2. ACL 2024Best Paper Award

   Causal Estimation of Memorisation Profiles

   Pietro Lesci, Clara Meister, Thomas Hofmann, Andreas Vlachos, and Tiago Pimentel

   In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Aug 2024.

   Abs arXiv Bib Code

   Understanding memorisation in language models has practical and societal implications, e.g., studying models’ training dynamics or preventing copyright infringements.Prior work defines memorisation as the causal effect of training with an instance on the model’s ability to predict that instance. This definition relies on a counterfactual: the ability to observe what would have happened had the model not seen that instance.Existing methods struggle to provide computationally efficient and accurate estimates of this counterfactual. Further, they often estimate memorisation for a model architecture rather than for a specific model instance. This paper fills an important gap in the literature, proposing a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics. Using this method, we characterise a model’s memorisation profile–its memorisation trends across training–by only observing its behaviour on a small set of instances throughout training.In experiments with the Pythia model suite, we find that memorisation (i) is stronger and more persistent in larger models, (ii) is determined by data order and learning rate, and (iii) has stable trends across model sizes, thus making memorisation in larger models predictable from smaller ones.

   @inproceedings{lesci-etal-2024-causal,
     address = {Bangkok, Thailand},
     author = {Lesci, Pietro and Meister, Clara and Hofmann, Thomas and Vlachos, Andreas and Pimentel, Tiago},
     award = {Best Paper Award},
     booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
     doi = {10.18653/v1/2024.acl-long.834},
     editor = {Ku, Lun-Wei and Martins, Andre and Srikumar, Vivek},
     month = aug,
     pages = {15616--15635},
     publisher = {Association for Computational Linguistics},
     title = {Causal Estimation of Memorisation Profiles},
     url = {https://aclanthology.org/2024.acl-long.834},
     year = {2024}
   }

3. NAACL 2024

   AnchorAL: Computationally Efficient Active Learning for Large and Imbalanced Datasets

   Pietro Lesci, and Andreas Vlachos

   In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). Jun 2024.

   Abs arXiv Bib Code

   Active learning for imbalanced classification tasks is challenging as the minority classes naturally occur rarely. Gathering a large pool of unlabelled data is thus essential to capture minority instances. Standard pool-based active learning is computationally expensive on large pools and often reaches low accuracy by overfitting the initial decision boundary, thus failing to explore the input space and find minority instances. To address these issues we propose AnchorAL. At each iteration, AnchorAL chooses class-specific instances from the labelled set, or anchors, and retrieves the most similar unlabelled instances from the pool. This resulting subpool is then used for active learning. Using a small, fixed-sized subpool AnchorAL allows scaling any active learning strategy to large pools. By dynamically selecting different anchors at each iteration it promotes class balance and prevents overfitting the initial decision boundary, thus promoting the discovery of new clusters of minority instances. Experiments across different classification tasks, active learning strategies, and model architectures AnchorAL is (i) faster, often reducing runtime from hours to minutes, (ii) trains more performant models, (iii) and returns more balanced datasets than competing methods.

   @inproceedings{lesci-vlachos-2024-anchoral,
     address = {Mexico City, Mexico},
     author = {Lesci, Pietro and Vlachos, Andreas},
     booktitle = {Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)},
     doi = {https://doi.org/10.48550/arXiv.2404.05623},
     editor = {Duh, Kevin and Gomez, Helena and Bethard, Steven},
     month = jun,
     pages = {8445--8464},
     publisher = {Association for Computational Linguistics},
     title = {{AnchorAL}: {C}omputationally Efficient Active Learning for Large and Imbalanced Datasets},
     url = {https://aclanthology.org/2024.naacl-long.467},
     year = {2024}
   }

2023

1. ACL 2023

   Diable: Efficient Dialogue State Tracking as Operations on Tables

   Pietro Lesci, Yoshinari Fujinuma, Momchil Hardalov, Chao Shang, Yassine Benajiba, and Lluis Marquez

   In Findings of the Association for Computational Linguistics: ACL 2023. Jul 2023.

   Abs arXiv Bib Code

   Sequence-to-sequence state-of-the-art systems for dialogue state tracking (DST) use the full dialogue history as input, represent the current state as a list with all the slots, and generate the entire state from scratch at each dialogue turn. This approach is inefficient, especially when the number of slots is large and the conversation is long. We propose Diable, a new task formalisation that simplifies the design and implementation of efficient DST systems and allows one to easily plug and play large language models. We represent the dialogue state as a table and formalise DST as a table manipulation task. At each turn, the system updates the previous state by generating table operations based on the dialogue context. Extensive experimentation on the MultiWoz datasets demonstrates that Diable (i) outperforms strong efficient DST baselines, (ii) is 2.4x more time efficient than current state-of-the-art methods while retaining competitive Joint Goal Accuracy, and (iii) is robust to noisy data annotations due to the table operations approach.

   @inproceedings{lesci-etal-2023-diable,
     address = {Toronto, Canada},
     author = {Lesci, Pietro and Fujinuma, Yoshinari and Hardalov, Momchil and Shang, Chao and Benajiba, Yassine and Marquez, Lluis},
     booktitle = {Findings of the Association for Computational Linguistics: ACL 2023},
     doi = {10.18653/v1/2023.findings-acl.615},
     editor = {Rogers, Anna and Boyd-Graber, Jordan and Okazaki, Naoaki},
     month = jul,
     pages = {9697--9719},
     publisher = {Association for Computational Linguistics},
     title = {Diable: Efficient Dialogue State Tracking as Operations on Tables},
     url = {https://aclanthology.org/2023.findings-acl.615},
     year = {2023}
   }