Process Supervision
Process supervision is a training paradigm contrasted with outcome supervision, in which evaluation signals are provided at intermediate steps rather than only at task completion. This approach gives models stepwise optimization feedback and is widely applied in reinforcement learning and LLM alignment, making it particularly suitable for training on long-chain reasoning and multi-step tasks.
Timeline (last 90 days)
Treating execution trajectories as training data essentially extends the supervision signal from 'outcome supervision' to 'process supervision'.
The industry has proposed methods such as scalable oversight and process supervision to address behavioral anomalies at the training mechanism level.
OpenAI's 2023 paper "Let's Verify Step by Step" demonstrated that process supervision (PRM) significantly outperforms outcome supervision (ORM) on mathematical reasoning tasks.
OpenAI 在2023年的研究《Let's Verify Step by Step》中证明，对推理过程每一步进行奖励评估（过程监督）比仅对最终结果评估能更有效提升数学推理能力
执行轨迹记录模型或人类解决问题的完整中间步骤，使训练系统能识别推理链条中具体哪一步出错，从而提供更精细的学习信号
All Facts (5)
OpenAI's 2023 paper "Let's Verify Step by Step" demonstrated that process supervision (PRM) significantly outperforms outcome supervision (ORM) on mathematical reasoning tasks.
70%UnverifiedTreating execution trajectories as training data essentially extends the supervision signal from 'outcome supervision' to 'process supervision'.
50%UnverifiedThe industry has proposed methods such as scalable oversight and process supervision to address behavioral anomalies at the training mechanism level.
50%UnverifiedOpenAI 在2023年的研究《Let's Verify Step by Step》中证明，对推理过程每一步进行奖励评估（过程监督）比仅对最终结果评估能更有效提升数学推理能力
50%Unverified执行轨迹记录模型或人类解决问题的完整中间步骤，使训练系统能识别推理链条中具体哪一步出错，从而提供更精细的学习信号
50%