liveBook · Manning

Process signal

As explained in Self-Improving Agents

A feedback signal that evaluates the steps by which an agent reaches its output.

A process signal evaluates how an agent reaches an output, step by step, rather than scoring only the final answer. section 1.4.1

Two forms assess the agent’s path:

  • A Process Reward Model assigns credit to individual reasoning or tool-use steps. section 1.4.1
  • A reflection signal has the agent or a critic critique its trajectory, which can reveal a fragile path even when the final answer looks fine. section 1.4.1
Figure 1.9 The signal family arranged by precision against availability. Ground-truth and environment signals are the most precise but the rarest; judge-based signals are available almost everywhere but are noisier, and process and reflection signals assess how the agent worked rather than only what it produced.

See how process and reflection signals assess an agent’s work, alongside signals that score its output. section 1.4.1

How Self-Improving Agents covers it

How deep it goes
Defines step-level credit and trajectory critique, without detailing their implementation.
Examples
Illustrates assessment through reasoning steps, tool-use steps, and a fragile trajectory behind an acceptable final answer.
What you can do after
Understand how feedback can assess an agent’s trajectory rather than only its final answer.
How it connects
Connects process reward models and reflection signals to ground-truth and judge-based signals.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Self-Improving Agents. The section links point to the parts of the book it draws on.