---
schema: 1
type: term
title: _reward function_
description: >-
  A **_reward function_** is the specification that defines what counts as
  success for a system being trained, providing a measurable signal the system
  optimizes toward as it explores possible behaviors.
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/reward-function-56561d909020
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/reward-function-56561d909020.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \_reward function\_

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/reward-function-56561d909020

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

A **_reward function_** is the specification that defines what counts as success for a system being trained, providing a measurable signal the system optimizes toward as it explores possible behaviors.

In this channel's material on **new AI architectures and coding agents**, a **_reward function_** is the specification of what counts as success for a system being trained. Rather than being shown every correct move, the system explores and is evaluated against the reward function, which tells it how well its behavior matches the desired outcome.The reward function is a key piece of the story the channel explores about agents like Devin: verifiable tasks — such as whether code compiles and passes tests — can be turned into a measurable signal, which makes it possible to train and evaluate agents on problems where correctness can be checked rather than guessed.- A **reward function** defines, in measurable terms, what success looks like for a trained system.- It lets a system **explore** and improve against an objective instead of needing explicit step-by-step supervision.- It is especially relevant for **coding agents**, where outcomes like passing tests are verifiable.