Reflection-in-Action

Reflection-in-action literally means: thinking while acting. The term comes from educational research. In the 1980s, social scientist Donald Schön used it to describe how experienced professionals actually work. A doctor doesn’t rigidly follow a textbook plan, but notices in the middle of an examination that something doesn’t fit, and takes a different direction. Schön called the counterpart reflection-on-action: the calm reflection afterward, once everything is over. In computer technology, the term is used today for programs that check their own intermediate results and adjust their approach during the ongoing process.

Why blindly executing plans fails

A computer program classically works through a fixed sequence of steps. That works well as long as the world looks exactly as expected. As soon as something unforeseen happens, a rigid plan runs into a dead end. That is precisely the fundamental problem of many AI systems that are supposed to complete tasks independently.

With language models, there’s a second problem on top of that. A language model is a program that has learned to continue texts, and in doing so can also produce calculations or program code. It produces its answer piece by piece and can make a mistake early on. Without checking, it then keeps building on that error for pages on end. Experts call this an error cascade: a small mistake at the beginning renders everything unusable in the end.

Reflection-in-action is the answer to this. If a system briefly checks after each step whether it’s still on track, errors stay small. This is also economically relevant. A coding assistant that tests its own code saves the human the debugging work.

Checking, evaluating, changing course

Technically, reflection-in-action consists of a loop with three parts. First, the system executes a step, such as a search or a piece of computation. Then it evaluates the result: Was that useful? Does it fit the task? Finally, it decides whether to continue or try something else.

The evaluation can come from outside or from the model itself. From outside, it’s more reliable. A program that writes code can simply have it executed: the error message is hard feedback. If the model evaluates itself, it writes a brief critique of its own output and continues working with that. Well-known methods of this kind are called Reflexion or Self-Refine.

A common misconception is that self-criticism always helps. That’s not true. A model that can’t recognize an error often just confirms itself when double-checking. Moreover, each round of checking costs computing time and thus money. That’s why the number of rounds is usually limited to just a few passes.

From AI agents to giving a talk

The term is most often read in connection with AI agents. That’s what systems are called that complete a task in their own steps, such as researching on the internet or editing files. Such agents only work usefully if they notice along the way that a source isn’t yielding anything. Providers therefore advertise with phrases like “self-correcting” or “reflective.”

The principle is also embedded in products. Coding tools automatically run generated code and repair it when errors occur. Research assistants check their intermediate results before writing a report. In trade news, the terms Reflexion, self-correction, and reflection-in-action often appear mixed up together.

The origin of the term continues to have an effect in pedagogy. There, reflection-in-action describes what you yourself do when a talk isn’t landing and you spontaneously explain it differently. This human meaning explains why the term is so popular in AI: it describes precisely the ability that machines have so far lacked the most.