It is a strong paper...
ECHO is interesting because it has strong backing from fixed point classes theory.
In GRPO, the reward signal is sparse: success or failure only touches a few boundary points, so the model must infer the convergence pathway from weak and discontinuous guidance.
ECHO changes this geometry by turning terminal responses—stdout, stderr, errors, logs, test outputs—into dense boundary conditions.
From the Deep Manifold view, this gives the agent many more local constraints along the trajectory, making the fixed-point class more identifiable, more stable, and easier to approach. In theory, dense boundary conditions are exactly what a fixed-point iteration wants: they reduce ambiguity, constrain drift, and improve convergence.
The hard part is practical, not conceptual: how to obtain such dense, faithful, action-linked boundary signals in domains where the environment does not naturally return rich terminal feedback. In this sense, ECHO works because the terminal is an unusually good physical cover of the world model.
