Outcome Rewards Do Not Guarantee Verifiable or Causally Important Reasoning

Reinforcement Learning from Verifiable Rewards (RLVR) on chain-of-thought reasoning has become a standard part of language model post-training recipes. A common assumption is that the reasoning chains trained through RLVR reliably represent how a model gets to its answer. In this paper, we develop two metrics for critically examining this assumption: Causal Importance of Reasoning (CIR), which measures the cumulative effect of reasoning tokens on the final answer, and Sufficiency of Reasoning (SR), which measures whether a verifier can arrive at an unambiguous answer based on the reasoning alone. Through experiments with the Qwen2.5 model series and ReasoningGym tasks, we find that: (1) While RLVR does improve task accuracy, it does not reliably improve CIR or SR, calling the role of reasoning in model performance into question. (2) A small amount of SFT before RLVR can be a remedy for low CIR and SR. (3) CIR and SR can be improved even without SFT by applying auxiliary CIR/SR rewards on top of the outcome-based reward. This joint reward matches the accuracy of RLVR while also leading to causally important and sufficient reasoning. These results show that RLVR does not always lead models to rely on reasoning in the way that is commonly thought, but this issue can be remedied with simple modifications to the post-training procedure.

Modern AI chatbots often "show their work," writing out a chain of reasoning steps before giving a final answer. A popular way to train them rewards the model whenever its answer is correct, and people generally assume the written steps faithfully reflect how the model actually reached that answer. But is that assumption true? We built two simple tests to find out. The first checks whether the reasoning steps actually change the final answer, or whether the model would have said the same thing regardless. The second checks whether someone reading only the reasoning — without seeing the answer — could figure out what the answer must be. Testing several AI models on a range of reasoning puzzles, we found that this common training method boosts accuracy but often does not make the written reasoning meaningful or self-contained. In other words, the model's "explanation" may not be how it really got there. The good news: two straightforward fixes restore trustworthy reasoning without sacrificing accuracy. This matters because people increasingly rely on these step-by-step explanations to understand and trust what AI systems are doing — so the explanations should actually mean something.