Interactive proof systems whose verifiers are constant-space machines have interesting features that do not have counterparts in the better studied case where the verifiers operate under reasonably large space bounds. The language verification power of finite-state verifiers is known to be sensitive to the difference between private and public randomization. These machines also lack the capability of imposing worst-case superlinear bounds on their own runtime, and long interactions with untrustable provers can involve the risk of being fooled to loop forever. We analyze such verifiers under different bounds on the numbers of private and public random bits that they are allowed to use. This separate accounting for the private and public coin budgets as resource functions of the input length provides interesting characterizations of the collections of the associated languages. When the randomness bound is constant, the verifiable class is $\rm NL$ for private-coin machines, but equals just the regular languages when one uses public coins. Increasing the public coin budget while keeping the number of private coins constant augments the power: We show that the set of languages that are verifiable by such machines in expected polynomial time (with an arbitrarily small positive probability of looping) equals the complexity class $\rm P$. This hints that allowing a minuscule probability of looping may add significant power to polynomial-time finite-state automata, since it is still not known whether those machines can verify all of $\rm P$ when required to halt with probability 1, even with no bound on their private coin usage. We also show that logarithmic-space machines which hide a constant number of their coins are limited to verifying the languages in $\rm P$.

Agentic AI systems increasingly rely on collections of cooperating agents—retrievers, planners, tool executors, verifiers—working together across large

Researchers have developed a new training framework that enhances visual reasoning in artificial intelligence by using AI-powered systems to both refine logical thinking and improve the accurate identification of objects within images, achieving superior performance compared to existing methods without relying on human-labelled data.

ArXiv link for Shrinking the Generation-Verification Gap with Weak Verifiers