Proceedings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 5: Tutorial Abstracts)

Maria Lomeli, Swabha Swayamdipta, Rui ZhangNAACL 2025

This document presents the abstracts for the tutorials offered at the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL 2025). It offers a curated selection of deep dives into cutting-edge and emerging topics in computational linguistics and natural language processing (NLP), designed to equip attendees with the latest insights, tools, and methodologies.

TL;DR

This collection of tutorial abstracts showcases 8 selected topics from 76 submissions, covering a broad spectrum of NLP advancements. The tutorials aim to provide in-depth knowledge on areas like creative planning with language models, GenAI red-teaming, foundation models for embodied agents, knowledge distillation, LLM adaptation, grounding language, copyright risks, and social intelligence in LLMs, catering to diverse expertise levels.

What Problem Does This Paper Address?

This document doesn't address a single problem but rather serves as a program guide for a set of tutorials. These tutorials collectively aim to address the rapidly evolving challenges and opportunities within the field of computational linguistics and natural language processing. This includes equipping researchers and practitioners with the knowledge and skills to tackle complex issues such as:

  • Developing more creative and controllable language models: Moving beyond basic text generation to models that can plan, reason, and execute creative tasks.
  • Ensuring the safety and robustness of AI systems: Developing methods to identify and mitigate potential harms, especially in generative AI.
  • Bridging the gap between language and action: Integrating language understanding with physical embodiment for more capable AI agents.
  • Improving the efficiency and effectiveness of large language models (LLMs): Techniques for knowledge distillation and adaptation to make LLMs more accessible and specialized.
  • Understanding and mitigating the societal impact of LLMs: Addressing issues like copyright, bias, and the development of social intelligence in AI.

The tutorials aim to provide a structured learning experience for participants to gain a deeper understanding of these critical areas.

Methods

The methods employed are diverse, as each tutorial focuses on a specific sub-field. However, common themes and approaches likely include:

  • State-of-the-art model architectures: Discussions on transformer-based models, diffusion models, and other recent advancements in deep learning for NLP.
  • Novel training and fine-tuning techniques: Methods for efficient knowledge distillation, domain adaptation, and few-shot learning.
  • Evaluation methodologies: Development of new benchmarks and metrics for assessing model capabilities in areas like creativity, safety, and grounding.
  • Theoretical frameworks: Exploration of underlying principles in areas like cognitive science and linguistics to inform NLP model development.
  • Practical toolkits and libraries: Introduction to software and frameworks that enable the implementation and application of learned concepts.

The selection process for these tutorials involved a rigorous evaluation by a panel of reviewers, assessing criteria such as clarity, novelty, timeliness, instructor expertise, and commitment to open access materials.

Datasets

The specific datasets used will vary greatly depending on the tutorial topic. However, based on the titles, we can infer potential dataset categories:

  • Creative Planning: Datasets for story generation, dialogue, or task-oriented planning.
  • GenAI Red-teaming: Datasets designed to probe for harmful outputs, biases, or vulnerabilities in generative models. This might include adversarial prompts or curated datasets of problematic content.
  • Foundation Models Meet Embodied Agents: Datasets related to robotics, simulation environments, or multimodal learning that combine language with sensory input.
  • Knowledge Distillation: Large-scale language model datasets (e.g., Common Crawl, Wikipedia) used for training teacher models, and potentially smaller, specialized datasets for student models.
  • Adaptation of Large Language Models: Domain-specific corpora for fine-tuning LLMs for particular tasks or industries.
  • Learning Language through Grounding: Multimodal datasets that pair language with visual, auditory, or other sensory information.
  • LLMs and Copyright Risks: Datasets of copyrighted text, legal documents, or generated content to analyze copyright infringement.
  • Social Intelligence in the Age of LLMs: Datasets related to social interactions, sentiment analysis, emotion recognition, or theory of mind tasks.

It is also possible that some tutorials introduce newly created benchmarks to address specific research gaps not adequately covered by existing datasets.

Key Insights and Ideas

  • The increasing sophistication of language models: Moving beyond simple text generation to complex tasks like creative planning and embodied interaction.
  • The critical need for AI safety and robustness: The rise of "red-teaming" highlights the proactive efforts to identify and mitigate risks in generative AI.
  • The convergence of language and action: Foundation models are being integrated with embodied agents, suggesting a future where AI can interact with the physical world more intelligently.
  • Efficiency is paramount: Knowledge distillation and adaptation techniques are crucial for making powerful LLMs more accessible and practical.
  • LLMs are entering complex societal domains: The tutorials on copyright and social intelligence underscore the growing awareness of the broader implications of LLM deployment.
  • Multidisciplinarity is key: The topics suggest a strong interplay between NLP, computer vision, robotics, and even cognitive science.

Key Results

As this document is a collection of tutorial abstracts, specific quantitative results are not presented here. The "Key Results" would be the outcomes and learnings participants gain from attending each tutorial. These would typically involve:

  • Understanding of novel algorithms and architectures.
  • Familiarity with new evaluation metrics and benchmarks.
  • Practical skills in applying specific NLP techniques.
  • Insights into the current state-of-the-art and future research directions.

For example, a tutorial on "Creative Planning with Language Models" might result in participants being able to design prompts or fine-tune models for specific creative writing tasks, while a tutorial on "LLMs and Copyright Risks" might equip attendees with methods to detect potential copyright issues in AI-generated content.

Limitations and Future Work

The limitations and future work are specific to each individual tutorial, and therefore not comprehensively detailed in this overview document. However, based on the general trends in NLP, we can anticipate common themes:

  • Scalability and efficiency: Many advanced models still require significant computational resources, and future work will focus on making them more efficient.
  • Robustness and generalization: Models often struggle with out-of-distribution data or adversarial attacks. Future research will aim to improve their resilience.
  • Interpretability and explainability: Understanding why models make certain decisions remains a challenge, and efforts will continue to make them more transparent.
  • Ethical considerations and societal impact: Addressing bias, fairness, and the responsible deployment of NLP technologies will be a continuous area of focus.
  • Multimodality and embodiment: Integrating language with other modalities and enabling AI to interact with the physical world will be a significant frontier.

The tutorials themselves are likely to highlight specific limitations of the discussed techniques and propose concrete directions for future research within their respective sub-fields.

Summary updated

BibTeX
@inproceedings{lomeli2025proceedings,
  author = {Maria Lomeli and Swabha Swayamdipta and Rui Zhang},
  title = {{Proceedings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 5: Tutorial Abstracts)}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure