This paper addresses the inefficiency of credit assignment in Reinforcement Learning with Verifiable Rewards (RLVR) for long Chain-of-Thought (CoT) reasoning. Current methods like Group Relative Policy Optimization (GRPO) uniformly distribute response-level rewards across all tokens, while On-Policy Self-Distillation (OPSD) relies on likelihood shifts from a privileged self-teacher to provide dense supervision. Through counterfactual testing—re-scoring trajectories under both correct and incorrect outcome conditions—the authors demonstrate that these likelihood shifts are unreliable indicators of reasoning quality. They find that most tokens exhibit identical shift directions regardless of the outcome, and large magnitude shifts are concentrated on substitutable surface-form tokens rather than core reasoning content. To mitigate this, the authors introduce Counterfactual Sensitivity Credit Reallocation (CSCR), a modification of GRPO that penalizes highly sensitive tokens and renormalizes token-level advantages. This approach preserves the original credit budget and the directionality determined by the verifier while reducing noise from non-essential tokens. Experimental results on long-CoT mathematical reasoning benchmarks show that CSCR consistently outperforms the GRPO baseline using the same number of policy updates. Ablation studies indicate that moderate downweighting of sensitive tokens is the most effective strategy, whereas excessive modulation destabilizes the optimization process.