AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Minimal Witness Reinforcement Learning
The paper tackles the recurring question of what irreducible conditions suffice to produce an outcome, formalizing the search for multiple minimal sufficient conditions ('witnesses')—which standard RL tends to answer with a single or redundant solution—as minimal-witness identification. Minimal-Witness Reinforcement Learning (MWRL) takes the union of sets certified by successful policy proposals and credits each proposal for the coverage the union would lose without it, unifying minimality and recovery of alternatives from a single black-box verifier bit. Under this principle the authors derive a value iteration planner that recovers entire witness families and a policy gradient method that scales to large language models, with MWRL recovering most minimal witnesses where other methods return redundant supersets or a single witness.