Welcome to the 1st Workshop on Taming Large Language Models: Controllability in the era of Interactive Assistants! This workshop aims to unite esteemed scholars, researchers, and practitioners specializing in Natural Language Generation (NLG). This event will foster in-depth discussions and explorations of the challenges and prospects associated with content control in LLMs. Emphasizing the intersection of NLG research and the instruction-learning paradigm, the workshop will serve as a platform for fruitful collaborations and knowledge exchange. This hybrid workshop will be co-located with INLG 2023 (https://inlg2023.github.io/workshops.html) at Prague.
Important Dates:
- Submission deadline: June 15, 2023
- Author notification: July 21, 2023
- Camera-ready deadline: August 14, 2023
- Workshop date: September 12, 2023
Submission Portal: https://softconf.com/n/tllm2023
Website: https://ctrlnlg.github.io/
**All deadlines are 11.59 pm AOE time.
Topics:
We welcome submissions on one or more of the following topics:
- Alignment: Investigating techniques to better align LLMs with human values and intentions, including reward modeling, human-in-the-loop systems, and quantifying alignment metrics. Understanding the objectives pursued by a model and aligning them with human preferences are key challenges. We encourage research on methods to increase alignments, such as through prompt design and fine-tuning.
- In-context Learning: Exploring the role of context in LLMs, including how to improve context understanding, manage context drift, and enhance context-aware responses. Also, investigating the use of in-context learning as a control mechanism.
- Instruction-based Control: Comparing popular controlling mechanisms, including approaches such as logit manipulation, decoder mixing, and classifier guidance, amongst others, against the simpler instruction-based control.
- Generality: Investigating controllable techniques that work across tasks and datasets.
Safety and Robustness: Assessing potential risks and vulnerabilities in LLMs, along with solutions such as adversarial training, safe exploration, and monitoring model behavior during deployment.
Controllability vs. Robustness: Developing methods to better understand LLMs' decision-making processes, and how it acts in grounded scenarios. Understanding its reliance on implicit vs. explicit memory.
- Scalability and efficiency: Investigating novel approaches for reducing computational requirements for achieving control in LLMs.
Real-world applications and case studies: Showcasing successful LLM deployments in various fields, such as healthcare, finance, education, and creative industries, along with lessons learned and future opportunities.
Submissions:
We welcome reports of original research in the form of two types:
- Long papers (8 pages + references)
- Short papers (4 pages + references)
We encourage all authors to include relevant discussions of ethical considerations and impact in the body of the paper.
Submissions will be made via SoftConf/START: https://softconf.com/n/tllm2023
Submission Format:
- The proceedings will be published by ACL Anthology.
- All long, short, and abstract submissions must follow the two-column ACL format , which are available as an Overleaf template and also downloadable directly (Latex and Word). Please refer to the SIGDIAL 2023 website for the most recent version of the templates.
- Submissions must conform to the official ACL style guidelines, which are contained in these templates. Submissions must be electronic, in PDF format.
- All submissions should be anonymized to facilitate double blind reviewing.
- Submissions that do not adhere to the author guidelines or ACL policies will be rejected without review.
- Appendix should be added in the main document after references. Appendix does not count towards the page length.
Any questions regarding submissions can be sent to tamingllm-workshop [at] googlegroups.com