Why belief revision in AI is blocked by implementation, not by formulas

A dependable AI system must be able not only to add new facts but to revise earlier conclusions. And it is exactly at that passage from logic to working code that the main gap opens: theory describes guarantees, while engineering needs concrete procedures and trade-offs.

The paper traces the path from Doyle and London's 1980 taxonomy to AGM — the formal basis for belief revision. The authors show that the early computational work and the later theory do not exist in isolation: there is continuity between them, but in the move to AGM the emphases and the demands on the system change.

The practical value of such a review lies not in yet another classification. It gathers the historical and theoretical foundations on which an implementation can be built: which categories of belief change to use, where their properties came from, and which formal guarantees have to be preserved in design.

That is an especially important shift for AI agents. An agent receiving new evidence must know which old beliefs to revise, which to keep and why. Without a formal model such logic quickly turns into a set of ad-hoc rules that are hard to check and extend.

But there is a substantial limitation: what we have is a review and a road map, not a ready library or proof that a particular algorithm works. The abstract offers no implementations, experiments, performance comparisons or universal way to carry AGM approaches into an applied system. The next step is to turn the historical analysis into engineering patterns with formal guarantees.

If your AI agent changes its conclusions when new data arrives, you are already designing a belief revision system — the only question is whether you are doing it explicitly or blind. Source: cs.AI updates on arXiv.org