Laura Ruis

Laura Ruis is a language-model researcher whose work covers how models generalize, how they learn to reason from pretraining data, and whether weaker judges can assess stronger models. The gSCAN benchmark for compositional generalization, presented at NeurIPS 2020, lists them as first author. Their PhD work at University College London includes the debate paper that won a Best Paper award at ICML 2024 and an ICLR 2025 paper linking model reasoning to pretraining documents that demonstrate procedures such as formulae and code. In 2026, they coauthored Self-CTRL with Jacob Andreas and others, a method that aligns a language model's explanations with its behavior. They are a postdoc in the Language & Intelligence lab at MIT.

Now at: Massachusetts Institute of Technology

Timeline

2020

The gSCAN benchmark

At NeurIPS 2020, Ruis leads the author list, with Jacob Andreas, Marco Baroni, Diane Bouchacourt and Brenden M. Lake, on gSCAN, a benchmark for compositional generalization in situated language understanding. A strong multimodal baseline and a state-of-the-art compositional method fail dramatically in most cases when generalization requires systematic compositional rules.
2022

Testing LLMs on implicature

In October 2022, Ruis leads the author list, with Akbir Khan, Stella Biderman, Sara Hooker, Tim Rocktäschel and Edward Grefenstette, on a study of how LLMs resolve implicature in yes-or-no questions. Of four categories of models tested, three perform close to random, while LLMs instruction-tuned at the example level do significantly better.
2024

An ICML Best Paper on debate

In 2024, Ruis is one of ten authors of "Debating with More Persuasive LLMs Leads to More Truthful Answers", which wins a Best Paper award at ICML 2024. The paper asks whether weaker models can assess the correctness of stronger ones; debate between two expert models lifts non-expert models to 76% accuracy and humans to 88%, against baselines of 48% and 60%.
November 2024

Procedural knowledge and reasoning

In November 2024, Ruis posts "Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models" with nine coauthors; it is published at ICLR 2025. A study of two models, with 7B and 35B parameters, finds that the influential pretraining documents for reasoning questions often demonstrate how to obtain a solution using formulae or code.
2025

Finishing at UCL, joining MIT

In February 2025, Ruis gives a seminar in Amsterdam titled "From Tokens to Thought: How do LLMs learn to reason?". In a blog post of August 2025, they write that after wrapping up their PhD at UCL they will join Professor Jacob Andreas at MIT for a postdoc.
2026

Infusion and Self-CTRL

In February 2026, Ruis coauthors Infusion with J Rosser, Robert Kirk, Edward Grefenstette and Jakob Foerster, a framework that uses influence functions to craft small edits to training documents that induce targeted model behavior. In June 2026, they coauthor Self-CTRL with Itamar Pres, Melat Ghebreselassie, Belinda Z. Li and Jacob Andreas, a method that aligns a language model's explanations with its behavior.
As of October 2026

Postdoc at MIT

As of October 2026, Ruis is a postdoc in the Language & Intelligence lab at MIT. Their research asks how LLMs attain their downstream capabilities from data.