In legal and adjudication settings, an LLM's ability to explain why a party should bear liability depends on explicit skills. KDASO (Knowledge-Data Dual-Driven Automated Skill Optimization) tries to turn those skills into an updatable asset, not a set of fixed prompts. The core idea is simple to state and complex to implement: evolve a skill library using recent data, the LLM's actual behavior, and a routing mechanism that respects the context budget.

The paper identifies three concrete limitations: aging libraries, a gap between formal skill descriptions and actual expert logic, and skill collections that grow beyond the context window, causing critical omissions. KDASO works on three fronts: neighborhood-conditioned induction to add new skills from recent adjudication data, execution-aligned refinement based on LLM trajectory analysis, and context-budgeted routing for scalable skill selection.

In tests across ride-hailing adjudication, legal liability determination and logical reasoning, the system beats general-purpose and domain-specific baselines. The 8.96% improvement when transferring skills to conceptually similar scenarios is perhaps the most telling figure: it suggests not just benchmark performance, but some portability of learned skills.

For teams running local systems, the real point is not the benchmark. KDASO shifts the problem from model size to the management of skills as operational artifacts. The logic is familiar to anyone working with VRAM constraints: instead of widening the context window or moving to larger models, you choose which skills enter the context. That reduces tokens and can make a self-hosted deployment viable on limited resources, but it should not be confused with a universal fix.

The approach favors organizations that must justify decisions under evolving rules: legal, insurance, and dispute-handling platforms. It pressures vendors that rely only on generic models with monolithic prompts. Structurally, it signals that value is moving toward maintaining traceable, coherent, updatable skill libraries rather than chasing the latest model variant.

For those weighing on-premise versus cloud deployment, the trade-off is not just cost: there is an opportunity to keep a skill library under local control, but with an engineering burden that an API service may not require. AI-RADAR offers analytical frameworks for these comparisons. The open question is how auditable these skills are: LLM trajectories are analyzed, but the boundary between fidelity to expert logic and autonomous drift will need attention in regulatory contexts where every step must be reconstructible.