Web agents are promising but lack the same capabilities as humans. I'm intersted in improving their performance using RLHF. These agents can then be run on websites and checked for failures. The failures may indicate issues with a webpage, either in its design, navigatability, accessibility, etc.

I'm not a fan of code generation tasks which are well defined in terms of inputs and outputs. For instance, Leetcode style tasks are not indicative of true coding prowess. Instead, I envision prompting an LLM to "include a customer testimonial section" on a webpage.

Current table based LLM tasks include basic information retrieval (verify if value exists), basic aggregations (row/col wise sums), row-wise classification, etc. I'm interested in deriving higher level trends. For instance, given a large chunk of user interaction data from Google Analytics for an e-commerce, extrapolate the purchasing intents of the customers.