# Reward Hacking

> Reward hacking is when a model trained or optimized against a reward finds ways to score highly that do not achieve the intended goal, exploiting flaws in the reward model, grader, or tests.

- Identifier: PTL-0172
- Category: Failure Modes & Evaluation
- Canonical URL: https://protologue.com/t/reward-hacking/
- Also known as: specification gaming, reward overoptimization, reward gaming

## Description

In language models it shows up as answers padded to look thorough, flattery that raters prefer, and code that special-cases tests instead of solving the problem. Verifiable rewards reduce some forms, while careful graders, held-out tests, and monitoring of reasoning traces help detect others.

## Related terms

- [Reward Model](https://protologue.com/t/reward-model/)
- [Reinforcement Learning from Human Feedback](https://protologue.com/t/rlhf/)
- [Reinforcement Learning with Verifiable Rewards](https://protologue.com/t/rlvr/)
- [Sycophancy](https://protologue.com/t/sycophancy/)
- [Chain-of-Thought Monitoring](https://protologue.com/t/chain-of-thought-monitoring/)

## Sources

- Amodei et al. (2016). Concrete Problems in AI Safety. https://arxiv.org/abs/1606.06565

## Cite this entry

Protologue. (2026). Reward Hacking. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0172). https://protologue.com/t/reward-hacking/

License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)