Reward Modeling

GoD-IRM introduces intermediate reward modeling for structured decision-making in language models, assigning rewards at each divergence point in a reasoning trajectory. This approach enables fine-grained credit assignment, improving model robustness in long-horizon problem-solving. By reinforcing decision-making rather than just final outputs, GoD-IRM aligns language models more closely with traditional agent-based RL.

Rearrangement Sampling transforms rejected solutions into new problem statements, expanding the problem-solution space beyond correctness constraints. A larger model acts as a judge to assess whether alternative problems can be inferred from divergent completions, assigning structured rewards to prior generations. This enables efficient data reuse, improves distributional coverage, and enhances reasoning generalization across domains.

We propose a feedback-driven decoding method where each generated candidate is iteratively refined using execution traces or reward-based adjustments. By conditioning generation on structured feedback from previous attempts, the method enforces progressive error minimization and adaptive correction. This approach enhances model reasoning, reduces compounding failure modes, and improves convergence in both code generation and reinforcement learning-based post-training.