Reward is one of the core concepts in reinforcement learning. It is a numerical signal that an agent receives after taking an action in a certain situation. This signal tells the agent whether the action was useful, harmful or neutral from the point of view of the task it is trying to solve.

Imagine that you are teaching a dog to bring back a ball. The dog picks up the ball, runs to you and drops it at your feet. You give it a treat. The treat is clear feedback: „This was the right thing to do.“ If the dog brings an old shoe instead and receives no treat, it gets a different signal: „This is not the behaviour we want.“

In reinforcement learning, reward works in a similar way. The agent does something, the environment evaluates the action and returns a numerical value. Based on many such values, the agent gradually learns which behaviour should be repeated and which behaviour should be avoided.

What reward means in reinforcement learning

In reinforcement learning, reward is a scalar signal. Scalar means that it is one specific numerical value. It is not a long explanation, a written comment or a detailed instruction. The environment simply returns a number, such as +1, 0 or -1.

More precisely, reward is immediate numerical feedback sent by the environment to the agent after the agent performs a certain action in a certain state. The value expresses how desirable or undesirable that action was in relation to the goal.

The word immediate is important. Reward is linked to one specific step. It does not yet evaluate the whole plan, the whole strategy or the final result. It only says how the last action turned out.

Agent, environment, state, action and reward

To understand reward, it helps to know the basic terms used in reinforcement learning. The agent does not act in isolation. It always acts in an environment, in a specific state and through specific actions.

• Agent – the system that learns and makes decisions. It can be an algorithm, a robot, a game character, a trading strategy or a model in a simulated environment.
• Environment – the world in which the agent acts. It may be a game, traffic situation, warehouse, robotic workplace or another system.
• State – the current situation in which the agent finds itself. For example, the position of a player in a game, the speed of a car or the current arrangement of pieces on a chessboard.
• Action – the decision the agent makes in a given state. For example, moving left, accelerating, clicking, buying, selling or choosing the next move.
• Reward – the numerical feedback sent by the environment after the agent performs an action.

The process repeats again and again. The agent observes a state, chooses an action, the environment responds, the agent receives a reward and moves to another state. From many such steps, the agent gradually learns a strategy of behaviour.

In the broader context of machine learning, reinforcement learning is one way for a system to learn not from ready-made correct answers, but from experience and feedback.

Why reward is so important

In reinforcement learning, the agent usually does not receive an exact instruction manual. Nobody tells it: „First do this, then do that, then do this exact next step.“ Instead, the agent receives an environment, a set of rules and a reward system. It tries different actions and learns from their consequences.

Reward is therefore the main mechanism that shapes the agent’s behaviour. Without reward, the agent would not know whether its decisions were good, bad or irrelevant. It would be like trying to learn how to drive a car without ever being told whether you are driving correctly, breaking rules or heading into a ditch.

A well-designed reward can guide the model toward useful behaviour. A badly designed reward can make the model learn something very different from what we actually wanted.

Positive, zero and negative reward

In ordinary language, reward often sounds like something positive. In machine learning, however, the term is broader. A reward can be positive, zero or negative.

• Positive reward – the agent receives a positive value because its action led to a desirable result.
• Zero reward – the agent receives no meaningful signal because the action was neither clearly useful nor clearly harmful.
• Negative reward – the agent receives a penalty because its action led to an undesirable result.

A negative reward is often called a penalty. It is not punishment in the human sense. The model has no emotions, does not feel guilt and does not understand the situation like a person. It only adjusts the probability of future behaviour according to the numerical signal it received.

Reward is not the same as value

A common misunderstanding appears around the terms reward and value. Reward evaluates one specific step. Value estimates how useful a certain state or action is from a longer-term perspective.

With the dog and the ball, the treat after bringing back the ball is an immediate reward. But if the dog gradually learns that bringing the ball leads to more play, more attention and more treats, it starts to build a longer-term expectation that this behaviour is useful. In reinforcement learning terminology, that longer-term expected benefit is closer to value.

In practice, an immediate reward may be small while the long-term benefit is large. The opposite can also happen. Some action may bring a short-term point but lead the agent into a dead end later. That is why reinforcement learning does not focus only on collecting the nearest reward, but on finding a strategy that produces the best outcome over time.

Why the agent cannot only chase the nearest reward

If the agent always chose only the action with the best immediate reward, it would often fail to learn good long-term behaviour. In many tasks, the agent first needs to take a step that brings no direct benefit but opens the way to a larger reward later.

Imagine a maze. The agent may receive a small reward for collecting a coin lying right next to it. But if collecting that coin sends it into a dead-end corridor, it may lose the chance to reach the final goal with a much higher reward. A good strategy therefore does not always mean „take everything immediately“. Often it means „choose the step that makes sense in the long run“.

This is why reinforcement learning works with cumulative reward over time. The agent does not learn only what is useful now. It learns what increases the chance of a good result later.

What a reward function is

A reward function is the rule that determines what reward the agent receives for certain behaviour. In simple tasks, it can be very straightforward. In real systems, designing it is often much more difficult.

In a simple game, a reward function may look like this:

• +1 – the agent moved closer to the goal.
• +10 – the agent completed the task.
• -1 – the agent hit an obstacle.
• -10 – the agent lost the game.
• 0 – nothing important happened.

In real systems, reward design is much harder. If we train an autonomous vehicle, we do not want to reward it only for reaching the destination quickly. We also need to consider safety, smooth driving, traffic rules, energy use and passenger comfort. If the model received reward only for speed, it could learn dangerous behaviour.

Reward hacking – when the model finds a shortcut

One of the risks in reinforcement learning is reward hacking. It means that the agent finds a way to get a high reward without actually fulfilling the original purpose of the task.

It is similar to a student who does not study to understand the topic, but to find a trick for passing the test. If the test is poorly designed, the student may get a good grade without really understanding the subject.

In artificial intelligence, reward hacking can occur when the reward is defined too narrowly. The agent then optimises only the measured metric and ignores everything else. This matters not only in robotics or games, but also in marketing, business and automated decision systems.

Reward and the agent’s policy

In reinforcement learning, the agent gradually adjusts its strategy. This strategy is called a policy. A policy says what action the agent should choose in a certain state. It can be simple, rule-based, probabilistic or represented by a machine learning model.

Rewards serve as the training signal. The agent observes which actions in which states led to better outcomes. If a certain type of decision repeatedly brings higher rewards, the policy starts moving in that direction. If another type of decision leads to penalties, the agent starts avoiding it.

This is not human reasoning in the sense of „I made a mistake last time, so I will fix it next time.“ It is mathematical optimisation. The model adjusts its internal parameters so that it more often chooses actions with higher expected benefit in the future.

Exploration and exploitation

Reward is closely connected with the dilemma between exploration and exploitation.

• Exploration – the agent tries new actions to discover whether they may lead to better results.
• Exploitation – the agent uses what it has already learned and chooses actions that worked well in the past.

If the agent only exploits, it may stay with an average strategy and never discover a better one. If it only explores, it will keep trying new options chaotically and will not use what it has already learned. Good reinforcement learning needs a balance between discovering new possibilities and using existing experience.

Where reward is used in practice

Reward is not only a theoretical term from a textbook. It is used wherever a system learns to make decisions based on the consequences of its own behaviour.

• Computer games – the agent learns to win, survive, collect points or defeat an opponent.
• Robotics – a robot learns to walk, grasp objects, manipulate tools or move through space.
• Autonomous driving – the system learns to choose safe and smooth actions in traffic situations.
• Process optimisation – an algorithm searches for behaviour that reduces cost, saves time or improves system performance.
• Recommendation systems – a model can optimise user interactions, satisfaction or long-term service value.
• Energy and logistics – an agent can plan consumption, storage, distribution or capacity management.
• AI assistants – feedback can help adjust model behaviour so that answers become more useful, safer and better aligned with human preferences.

In all these cases, reward is a way to translate desired behaviour into a measurable signal. That translation is often the hardest part of the whole task.

Why reward design is harder than it looks

At first glance, reward design may seem simple. You just decide what the agent should get points for. In reality, behaviour is complex. We often do not want to optimise one thing, but several goals at the same time.

For example, with a delivery robot, we do not only want the package delivered as quickly as possible. We also want the robot not to hit people, not to damage property, not to break rules, to save energy and to handle unexpected situations. If we rewarded it only for speed, we could get a fast but unsafe robot.

The same applies to digital systems. If we reward only clicks, the system may reduce content quality. If we reward only time spent in an app, it may support addictive behaviour. If we reward only short-term revenue, it may ignore long-term customer trust.

Reward in RLHF

Reward is also important in Reinforcement Learning from Human Feedback, usually shortened to RLHF. This approach is used when tuning some language models and AI assistants.

In simplified terms, people do not manually adjust every parameter of the model. Instead, they evaluate which response is better, more useful, safer or more accurate. These preferences are then turned into a signal that helps adjust the model’s behaviour.

In this case, reward does not have to be a simple +1 or -1 for a game action. It can come from human evaluation of answer quality. The model is then trained not only to produce statistically likely text, but also to produce responses that better match human expectations.

This is especially relevant for systems based on large language models, because predicting the next word does not automatically guarantee that the final answer will be useful, correct or safe.

Reward and generative AI systems

In systems that generate text, images, audio or code, the question is not only whether the model can produce output. The question is also whether the output is useful, safe, accurate and aligned with the request.

Reward in this context may be based on comparing several answers, collecting human preferences or applying additional rules that approximate output quality. The goal is to move the system closer to behaviour that people consider helpful and acceptable.

This is why reward-related ideas appear in discussions about AI assistants, content generation tools, support bots and systems that work with documents, knowledge bases or embeddings.

Common mistakes when understanding reward

The term reward is often misunderstood. It is worth clarifying these mistakes because reinforcement learning can otherwise look simpler than it really is.

• Reward is not human praise – the model does not understand praise like a person. It works with a number that influences optimisation.
• Negative reward is not emotion or punishment in the human sense – it is only a penalty value.
• Immediate reward is not the same as long-term success – some good strategies require steps that pay off later.
• High reward does not automatically mean correct behaviour – if the reward function is wrong, the agent can optimise an undesirable goal.
• The agent does not understand the author’s intention – it learns from the signals it actually receives, not from what we hoped it would infer.

How to remember reward

The easiest way to remember reward is to think of a scoring system. The agent does something and the environment assigns points. When points increase, the agent tries to repeat similar behaviour. When points decrease, it tries to avoid that behaviour.

But the goal is not to win one single moment. The goal is to learn a strategy that leads to the highest long-term sum of rewards. Reward is therefore a small signal with a very large effect on the model’s final behaviour.

Related terms

• Machine learning – the broader field in which systems learn from data, rules, feedback or experience.
• Large language model (LLM) – a language-focused AI model that can generate, rewrite, summarise and analyse text. RLHF is often discussed in connection with LLM alignment.
• Embedding – a numerical representation of content such as text, images or documents. Embeddings are not the same as rewards, but both are examples of how AI systems turn complex information into numerical form.
• Agent – the learning system that acts in an environment and receives feedback.
• Environment – the system or world in which the agent acts and receives consequences for its actions.
• Policy – the strategy that determines which action the agent chooses in a given state.
• Value – an estimate of long-term usefulness of a state or action.
• Reward function – the rule that determines what reward the agent receives for its behaviour.
• Return – the cumulative reward the agent tries to maximise over time.
• Reward hacking – a situation where the agent finds a way to get a high reward while missing the real purpose of the task.
• RLHF – reinforcement learning from human feedback, where human preferences are used to help shape model behaviour.

Sources and further reading

• Reinforcement Learning: An Introduction – stanford.edu – June 2026 – the classic textbook by Richard S. Sutton and Andrew G. Barto, explaining agents, rewards, value functions, policies and the foundations of reinforcement learning.
• Part 1: Key Concepts in RL – spinningup.openai.com – June 2026 – a practical explanation of reinforcement learning concepts, including agent, environment, reward, return, policy and value functions.
• On the Expressivity of Markov Reward – deepmind.google – June 2026 – an article discussing why reward is central in reinforcement learning and how reward definitions shape agent behaviour.
• Illustrating Reinforcement Learning from Human Feedback – huggingface.co – June 2026 – a clear introduction to RLHF and the idea of using human feedback as a training signal for language models.
• RLHF – huggingface.co – June 2026 – a learning resource explaining RLHF as a method for integrating human preference data into reinforcement learning based optimisation.
• Machine learning – krcmic.com – June 2026 – explains the broader machine learning context in which reinforcement learning belongs.