Outcome-based Reward Model (ORM) is a method that evaluates an AI model solely based on the final answer, without checking the path taken to get there. If the result is correct, a reward is given. The calculation path, reasoning or intermediate steps play no role.
This approach has a practical advantage: it is easy to train because only the final result needs to be annotated and not every single intermediate step. For many tasks, this is sufficient, such as with simple classification, text summarisation or question-answering tasks with a clearly verifiable result. The disadvantage becomes apparent with more complex, multi-stage tasks. A model can still arrive at the correct result through a flawed or even guessed solution path. An outcome-based reward model does not recognise this difference and rewards the lucky hit just as much as a sound derivation. Over many training runs, this can lead to a model learning to rely on plausible-sounding but substantively fragile arguments, as long as the numbers add up in the end.
Demarcation
The counterpart is the Process Reward Model, which evaluates each intermediate step individually and thereby also detects flawed derivations with a coincidentally correct result. An outcome-based reward model is cheaper to train, while a process reward model provides the finer signal. Both are variants of the reward function within reinforcement learning training, not reinforcement learning itself.
SAP B1 10.0 FP2608: Service Layer AI as a transactional layer
Process Reward Models: Why a correct result does not yet prove a correct method
Test-Time Compute Scaling: Why a Smaller AI Model Can End Up Winning — and What Supplier Comparison in SAP Business One Has to Do With It
RLHF and reward models: AI hype or what the approval process in SAP Business One has to do with it
AI – Answers from SAP Business One – without SQL, without IT ticket
