Blog

Conferences confine a paper to be at most X pages in length, where X is usually 7-10 pages. The need to communicate most of the important (theoretical and experimental) parts of my work in such a format means making sacrifices - this often ends in me removing text that aims... [Read More]

Direct Preference Optimization is all the rage now in LLMs, and rightly so! The derivation is neat (and very familiar to those experienced with reinforcement learning) and allows direct, preference-based finetuning of regression-trained LLMs without having to learn a reward model. [Read More]

Intro [Read More]

Intro [Read More]

Intro [Read More]

What is Reinforcement Learning? [Read More]