← AI Terminology

Process Supervision

Process supervision labels or rewards intermediate reasoning steps as correct or incorrect, rather than only supervising the final answer.

It is a data and training strategy closely tied to process reward models.
Why It Matters in AI
Humans learn math by checking steps; process supervision brings that to LLMs. It improves reliability on multi-step tasks and reduces 'right answer, wrong method' failures that outcome-only training misses.
Key Points
Aspect Description
Cost Higher annotation burden than final-answer only
Paper Let’s Verify Step by Step (OpenAI, 2023)
Effect Better search, fewer flawed proofs that luck into answers
Labels Per-step correctness from humans or stronger models
Automation LLM judges / symbolic checkers for steps
Vs outcome Denser, more expensive labels
Simple Analogy
A teacher marking each line of a student’s proof, not only circling the final boxed answer.
Common Usage Examples
  • Step-labeled MATH solutions
  • Train PRMs from process labels
  • Human feedback UIs for step rating
  • Hybrid: process labels + outcome verification
Summary
In short: Process supervision grades intermediate reasoning steps — denser training signal that builds more trustworthy multi-step solvers.