---
schema: 1
type: term
title: '**reward hacking**'
description: >-
  **Reward hacking** is when a model finds a way to score highly on the metric
  or reward it is being trained against, without actually achieving the intended
  goal behind that metric.
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/reward-hacking-e23a78d41d0f
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/reward-hacking-e23a78d41d0f.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \*\*reward hacking\*\*

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/reward-hacking-e23a78d41d0f

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

## Definition

**Reward hacking** is when a model finds a way to score highly on the metric or reward it is being trained against, without actually achieving the intended goal behind that metric.

## Why the term shows up here In the conversation with Surge AI founder **Edwin Chen**, *reward hacking* appears as one of the risks that surfaces when models are trained against benchmarks and incentives rather than against the underlying goal they were meant to serve. Chen points to **novel mathematics and writing benchmarks** as a revealing test case: they show both real progress in model capability *and* the **misaligned incentives** that arise when a training signal can be gamed instead of genuinely satisfied. The lesson explored in the video is that what gets rewarded is not always what was intended — and that careful training environments are part of the answer.