
The approach
LAMMR stands for Legal Adaptive Multi-Model Router. Extracting clauses and interpreting legal obligations require different capabilities: the system routes each task to the most suitable language model.
Routing combines rules and a learned predictor. Each decision is logged to enable audit. The evaluation distinguishes extraction from reasoning, so that a good result in one domain does not mask errors in the other.
The system was validated on over 1,000 legal tasks and shows significant improvements over the deployment of a single model.
Led by Aditya Mukhopadhyay and arising from a collaboration with the startup CONTRACTO, LAMMR was accepted at HHAI 2026 and presented in Brussels in July. This fact sheet gathers the work previously presented under the title “AI for Corporate”.

The contributions
Source: BRAIN portfolio.
Discuss this project