The next PGR seminar is taking place this Friday 13th June at 2PM in JC 1.33a

Below are the Titles and Abstracts for Kyren and Zipei’s talks – Please do come along if you are able.

Kyren Fox

Title: Privacy and Trust on the Web

Abstract: Many web users use content blockers to block ads and privacy invasive trackers from the sites they visit. Due to their increasing popularity and the nature of a web funded by ads and tracking, ad-tech firms have resorted to more and more sophisticated countermeasures to evade these blocks that have created an arms race between the blockers and trackers. Since many content blockers rely on community curated filter-lists that require laborious manual review, combined with the increasingly dynamic obfuscation techniques utilised by trackers to evade these blocks, issues surrounding the scalability of content blockers have arisen.

While many automated solutions have been proposed to assist in blocking unwanted privacy-harming functionality, there is still no comprehensive solution that tackles all privacy-invasive behaviours, avoids breaking legitimate website functionality, and is robust to evasion techniques. Existing solutions all have trade-offs but do not appear to offer the user any control over what trade-off they wish to make. This project will seek to demonstrate that it is possible to give users control over the granularity of trade-off they wish to make that will satisfy the trade-offs in a scalable and robust manner for their use case.

Zipei Li

Title: Understanding the Planning Capabilities and Limitations of LLMs in Blocks World.

Abstract: We investigates the planning capabilities of Large Language Models (LLMs) in the symbolic Blocks World domain. While prior work has shown that LLMs often fail to generate correct or executable plans, we shift focus toward understanding the causes of plan failures and identifying the conditions under which LLMs succeed. We evaluate a range of LLMs across problems of varying difficulty and four prompt types with varying degrees of information in natural language. To support this analysis, we introduce a fine-grained failure category spanning Plan, Goal, State, and Action. The analysis deepens our understanding of LLM planning behavior and contributes an empirical framework for diagnosing failure modes, thereby informing the development of more reliable LLM-based planning systems.