Reinforcement Learning for Law

Every major AI company uses reinforcement learning. How about in legal drafting?

Reinforcement learning involves using every action as feedback for a self-learning system. By observing user patterns, the system learns how to increase the probability of making a useful suggestion.

In legal drafting, any time a user does anything: accepting, rejecting, or modifying a clause modification created by our legal drafting system, the system logs that feedback and learns from it.

We use reinforcement learning to make legal drafting more adaptive. You can imagine that over thousands of interactions, the system learns to prioritize the phrasing and structure that align with the drafting standards in each individual practice area, for each individual document type.

Ultimately, we want to improve how a language model performs for a specific task. One approach is model fine-tuning, which should improve domain-specific performance. However, effective fine-tuning is highly non-trivial, and it can be difficult to fully characterise the change in model behaviour, especially with the involvement of synthetic task-specific training data.

Reinforcement learning can be applied outside of direct model adjustments. This can involve adjusting the retrieval framework and context window behind a language model, though it requires some creativity. One such approach has been widely publicised as an agentic framework.

Finally, reinforcement learning can work at multiple levels. We can have reinforcement learning at the practice-level (ie. tuning the system for estate planning), or at the user-level (ie. personalising to the individual's drafting style).

If there articles are of interest to you, subscribe to recieve more from Lexifina.