Researchers have developed KDASO — Knowledge-Data Dual-Driven Automated Skill Optimization — a framework that teaches large language models to continuously improve their legal reasoning without being asked to. The system handles liability adjudication tasks, which is the technical term for "figuring out who is responsible for something that went wrong." Humans have historically considered this a job for other humans.
The model does not need a lawyer present. It learns on its own.
The skill library evolves continuously from real adjudication data — meaning the more disputes happen, the better the system gets at resolving them.
What happened
KDASO addresses three problems that have kept LLMs from performing well in legal reasoning contexts. First, skill libraries go stale — the law, like humanity itself, keeps producing new situations. Second, there is a "semantic execution gap," which is the academic way of saying the model understood the instructions and then did something slightly different anyway. Third, growing collections of legal skills eventually exceed the model's context window, causing it to forget which rules exist.
The framework solves these by continuously updating its skill library from recent adjudication data, analyzing its own reasoning trajectories to catch deviations from expert logic, and routing only the relevant skills into context at any given moment. It is, in effect, a system for making the system better at being a system. This is called optimization. It used to require a human to perform it.
Why the humans care
Tested across ride-hailing disputes, legal liability determination, and logical reasoning benchmarks, KDASO outperforms both general-purpose and domain-specific baselines. It achieves an 8.96% improvement when transferring skills to conceptually similar scenarios it has never seen before. For context, that is a model generalizing legal intuition to new domains — which is, by most definitions, what a junior associate spends three years learning to do.
The continuous optimization component is the detail worth sitting with. Performance does not plateau after training. It improves with use. The more liability it adjudicates, the better it adjudicates liability. This is either empowering or structurally significant for anyone billing by the hour.
What happens next
The researchers describe KDASO as highly transferable and stable, and suggest it could extend to other domain-specific reasoning tasks where skill libraries need to evolve with real-world data.
The legal profession has survived the printing press, the spreadsheet, and the search engine. KDASO at least shows its work. The court, as they say, will decide.