Skip to content

KDASO addresses agent drift in liability attribution: continuous skill improvement and an 8.96 percent jump

Share
KDASO addresses agent drift in liability attribution: continuous skill improvement and an 8.96 percent jump

Listen to this article

Read by Anchor

Deploying large language models for legal adjudication and liability attribution faces a structural dilemma that goes beyond drafting initial guidelines. When an intelligent system is asked to scrutinise the details of a dispute and match them against sector-specific rules to decide whether a party bears responsibility, the models encounter three critical operational gaps: the skill and rule libraries become outdated relative to the dynamic evolution of real-world scenarios; the semantic-execution gap widens as the model's inference logic drifts from the precise intents of experts; and the skill bank inflates beyond the context-window budget, which practically leads to the omission of essential inferential conditions and clauses.

The development of skill libraries cannot remain dependent on labour-intensive manual updates.To address this challenge, a joint research team led by Zi Jian Cheng, Wei Ming Wu and Lan Qi Gu presented, in a paper published on the Arkhive platform, a new framework called “KDASO” (KDASO), an automatically guided, dual-led skill-improvement system that leverages both knowledge and data and enables continuous self-development and adaptation of procedural skill libraries for liability resolution.

The proposed framework rests on three algorithmic pillars designed to tackle the existing bottlenecks precisely. The first pillar is neighbourhood-conditioned induction from the latest adjudicative and judicial data, generating and updating rules according to the most recent disputes. The second pillar is execution-adjacent refinement, whereby the system analyses the actual inference paths of language models to diagnose deviation points and adjust skill formulations to close the expert-logic gap. The third pillar is a budget-constrained skill-guidance and selection algorithm, which ensures that the most relevant rules for a dispute are chosen and passed efficiently without exhausting memory space or discarding core rules.

The researchers subjected the KDASO framework to intensive experimental evaluation across three complex domains: adjudicating disputes and liabilities of passenger-transport and delivery platforms, determining liability and legal charges, and logical-inference-based judgment tasks. The results showed a tangible superiority of the framework over generic baseline models and specialised systems in those areas. Most notably, the system achieved an 8.96 percent jump in accuracy when transferring the refined skills to conceptually similar new scenarios, while maintaining stable performance gains through continuous improvement cycles.

Shifting from manual directive tuning to dynamic skill guidance reshapes computational cost and decision quality.This development has direct relevance for engineering teams and compliance units in the Middle East and North Africa, especially in Gulf and Egyptian business hubs where ride-sharing, logistics and fintech platforms are expanding the automation of consumer-dispute handling and contractual liability estimation. The traditional reliance on stuffing context windows with massive regulatory rule sets not only inflates cloud inference costs but also raises the risk of regulatory error when a model drops an exceptional rule. This approach offers an alternative path that tracks decision trajectories and directs rules according to budget, enabling developers to adapt to evolving local regulations without retraining foundational models from scratch.

Don't miss the next story

Subscribe for updates