𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝟵.𝟭𝟭>𝟵.𝟵 𝗮𝗻𝗱 𝗮 𝘁𝗿𝗶𝗮𝗻𝗴𝗹𝗲 𝗵𝗮𝘀 𝗳𝗼𝘂𝗿 𝘀𝗶𝗱𝗲𝘀, 𝘁𝗵𝗲𝗿𝗲𝗳𝗼𝗿𝗲 𝟭+𝟭=𝟮. LLMs and Language Agents can sometimes generate correct answers from blatantly incorrect reasoning, which is more often in complex tasks, and exacerbated by reinforcement learning (RL), the commonly believed silver bullet to complex reasoning in LLMs. This is due to a well-known phenomenon called reward hacking, where if the only training signal LLMs are getting from the training data exclusively regards the final result, then LLMs are incentivized to match the correct final output through whatever means possible on its training data, leading to inconsistent and ungeneralizable reasoning processes in RL's wake. With our intern Mengzhao Jia, we (@ignaciocases and myself, plus folks from @Meng_CS s lab at Notre Dame) explore a simple fix: can we use the LLM's own reasoning to provide some additional supervision signal for the reasoning process itself, so that besides the final result, the LLM is also encouraged to stay consistent in its reasoning during training? We design an algorithm to automatically create rubrics for LLM reasoning processes, and train the model to adhere to these rubrics alongside generating correct final answers during RL. The resulting model not only produces significantly more consistent reasoning, but also generalizes better on a wide range of complex reasoning tasks we benchmarked, even with just 10% of the training data. We hope this technique helps pave the way to more powerful and generalizable reasoning models for complex tasks. Read more in our preprint: arxiv.org/pdf/2510.14738