We build RL environments that simulate complex legal reasoning for AI labs.

Current language models can recite case holdings and summarize documents well, but struggle with applying legal doctrines to novel fact patterns and interpreting ambiguous laws.

Our reinforcement learning (RL) environments address this by simulating complex legal scenarios that emphasize advanced reasoning and edge cases, going beyond the general legal materials typically used for training.

Rather than optimizing for discrete tasks, we focus on fundamental reasoning skills that transfer across objectives like case analysis, motion drafting, discovery strategy, and settlement negotiation.

Legal data presents unique advantages for RL: its adversarial structure provides clear feedback and reward signals, while the nature of legal reasoning offers a complex environment for sequential decision-making in the face of nuance and ambiguity.

Our long-term goal is to expand into other domains requiring formal logical reasoning—mathematical proofs, regulatory compliance, medical diagnosis, and structured argumentation.

Join our email list for updates on our work and occasional logic puzzles.

Lemma is backed by Y Combinator.

You can reach us at founders@withlemma.com.
For our forthcoming email list, click here.

Founded by Sherwin Lai and Sina Mollaei.