Alexander Wettig

Wettig completed his computer science doctorate at Princeton in 2026, advised by Danqi Chen, after undergraduate study at Cambridge. His thesis studied how selecting and combining training text affects language models. He now describes his work at Cursor as reinforcement learning for Composer, the company's coding models. That training process uses feedback on a model's actions to improve its behavior. His university adviser also lists Cursor as his current organization.[1][2][9]

Choosing training text with QuRating

QuRating asks how to choose useful text from a much larger collection. Wettig and his co-authors used language-model comparisons to judge qualities such as educational value, then trained a smaller rating model to score documents. Their experiments showed that selecting only the highest-rated text was less effective than balancing quality with variety. This gives model builders a concrete method to test when the obstacle is the composition of their training data rather than the model architecture.[3]

Making web data understandable

WebOrganizer groups web pages by topic and format so researchers can control what goes into a model's training mix. Instead of treating a web crawl as one undifferentiated collection, its classifiers identify meaningful categories. The team tested how changing their proportions affected downstream tasks and found that domain mixing could complement quality filtering. This makes it easier to investigate which kinds of text help a capability and which are being removed by a quality filter.[4]

Training models to use longer context

Wettig co-led the work behind ProLong, a recipe for extending how much information a language model can use at once. The researchers evaluated practical long-context tasks rather than relying only on finding a hidden phrase in a long input. They found that books and code repositories were useful sources of long examples, but combining them with good short examples remained important. The paper provides training code, data and models for others to investigate these choices.[5]

Testing and building software agents

Wettig co-authored SWE-bench, which asks a model to repair real repository issues, and SWE-agent, which gives a model tools for navigating files, editing code and running tests. The benchmark provides an executable task; the agent paper studies how the interface changes the model's ability to complete it. Together they help separate a model's ability to generate code from its ability to work through a software problem with feedback from a running environment.[6][7]