Where RLOO Stops on Countdown: A Capability Ceiling at the Additive ↔ Multiplicative Boundary
A controlled diagnostic study is run, treating the collapse hypothesis as a claim to be falsified rather than assumed, and instrument a baseline RLOO run with diagnostics beyond the standard optimisation metrics: structural (AST) rollout diversity, effective sample size, a three-way zero-advantage decomposition, and reward composition.
Aarohi Gupta
· 0 citations