TD-Grokking: Learning from Zero-Reward Problems by Training-Time Decomposition

In three linesTD-Grokking recursively decomposes zero-reward problems into verifiable subproblems to generate training signals. Tested on mathematical and medical tasks, the framework outperforms vanilla GRPO and converts zero-reward examples into usable optimization signals.

Summary generated by Claude — human-verified