Taking place 29-30 July, the ELLIS Workshop on Robustness in Large Language Models (RobustLLMs) is a two-day event hosted by the Oxford Department of Statistics at Keble College, Oxford and co-organised by ELISE. The workshop will feature keynotes and invited talks, discussions and poster sessions focused on the role of robustness in improving factuality and reasoning, defending against adversarial inputs, enhancing reliability for real-world applications, and dealing with hallucinations in LLMs.
Visit the workshop website for more information.
Topics
- Implications of Robustness on Safety, Hallucinations, Factuality and Reasoning:
- Investigating robustness against adversarial inputs (prompt injections) specifically tailored to deceive LLMs.
- Addressing robustness in the context of distributional shifts and their impact on LLM performance.
- The role of robustness in improving factuality and reasoning of LLMs, detecting and mitigating hallucinations etc.
- Enhancing LLM Reliability for Real-world Applications:
- Techniques for uncertainty quantification in LLMs to improve decision-making reliability.
- The role of diverse and challenging datasets in assessing and enhancing the reliability of LLMs.
- Verification of LLM properties to ensure reliability and trustworthiness in real-world applications.
- Societal, Ethical, and Legal Considerations:
- Legal aspects concerning the robustness of LLMs, including liabilities related to misinterpretations and erroneous outputs.
- Strategies for ensuring that LLMs are developed with fairness and without bias, promoting ethical AI practices.
- Examining the robustness requirements of LLMs in critical sectors such as legal, healthcare, and content moderation, where the stakes are particularly high.
- Innovation and Future Directions:
- Novel approaches for improving the robustness of LLMs through architecture innovations, training methodologies, and data augmentation techniques.
- The potential of hybrid models that combine the strengths of LLMs with other AI techniques for enhanced robustness.
- Anticipating future challenges and opportunities in the evolving landscape of LLM robustness and reliability.
