Supervision that conditions learning on the internal steps, action sequences, or representation trajectories that lead to an output instead of (or in addition to) supervising only the final result.
Process based supervision is an approach where training signals target the dynamics or intermediate states of a model or agent—examples include supervising hidden-layer activations, aligning predicted action sequences with expert trajectories in imitation learning, constraining reasoning chains in large language models, or enforcing fidelity to program traces in neural program induction. Practically, this is implemented via auxiliary losses on intermediate representations, trajectory-level imitation objectives, contrastive alignment of process embeddings, process-distillation (student models trained to match teacher internal dynamics), and procedurally annotated datasets that provide stepwise ground truth. Theoretical motivations include improved credit assignment, reduced reliance on spurious input–output correlations, better out-of-distribution generalization when process-level invariants are causal, and increased interpretability and auditability. Key applications span imitation learning and robotics (trajectory supervision), structured prediction and sequence modeling, program synthesis and neural-symbolic systems, and recent LLM work where chain-of-thought traces or rationale annotations are used to shape model reasoning. Trade-offs include annotation cost for process traces, risk of enforcing incorrect or suboptimal procedures, potential reduction in model flexibility, and challenges in defining suitable metrics for process fidelity.
First explicit uses in ML literature appeared in the 2000s within imitation learning and structured-prediction communities; the approach gained broader traction in the 2010s with deep learning, and saw a surge of interest around 2021–2024 driven by chain-of-thought supervision, interpretability work, and process-level objectives for large models.