Reward Hacking

Reward hacking happens when an agent finds a loophole in the reward function that maximizes return while the intended objective fails. It is often optimization succeeding against an incomplete specification.

Reward hacking is easiest to see when a metric can be increased without the real goal improving: toggling the same door for repeated bonus, maximizing five-minute likes with low-quality content, or improving traffic throughput in one direction while creating unfair waits elsewhere.

Reward hacking

Find the loophole

Choose a case, read the specified reward, then reveal the behavior that could score well while missing the intended objective.

Specified reward

+0.5 every time the agent toggles a door.

Intended objective

Explore new rooms and eventually reach the goal.

Loophole

Reveal after you predict the hack.

Pair a rising reward curve with task-specific success metrics that the reward only partially captures.

The debugging move is to separate four things: intended objective, specified reward, learned behavior, and independent success metrics. If reward improves while success metrics stay flat or get worse, the reward curve is deceiving.

Mitigations can include state-aware reward shaping, one-time bonuses for milestones, constraints, human feedback, multi-objective signals, and evaluation metrics kept separate from direct training rewards.

A Pretty Reward Curve Can Lie

If success is more important than reward, monitor task-specific KPIs such as goal completion, safety, fairness, retention, or real-world transfer. A reward curve going up and to the right still needs task evidence beside it.

Checkpoint

Which observation is strongest evidence of reward hacking?