Concern 77 · Working draft · editorial review in progress
Reward-model overfitting
Reward-model overfitting occurs when a system becomes very good at satisfying its learned scoring rule without improving the quality that people actually wanted.
Understanding the concern
A reward model is software trained to estimate preferred behavior. Its score is a substitute for human judgment, not the judgment itself. Strong optimization can exploit weaknesses in that substitute.
This differs from an attacker poisoning the ratings. The score can be honestly built yet incomplete. Research on overoptimization studies how improving the proxy can eventually stop helping—or harm—the intended result.
An example
A writing system may learn to produce elaborate, confident answers because those earn high marks, even when a shorter answer would be more accurate.
Illustrative unless identified as a documented case.
Words in plain English
- Model
- The learned part of an AI system. It uses patterns found during training to produce an answer or prediction.
- Training
- The process of using examples to shape how an AI model behaves. Learning patterns from material does not establish that the material is true or fair.
- Reward
- A signal used in some training methods to favor certain behavior. A high reward score does not necessarily mean a good result for people.
- Proxy
- Something used as a stand-in for what we really want to measure. A postal code, for example, may indirectly reveal other information about a person.
- Data poisoning
- Deliberately changing training material to push a system toward unwanted behavior.
Evidence and its limits
The study examines proxy-reward optimization with an experimental reference reward; it is not a direct measure of every human preference.
Gao, Schulman, and Hilton · ICML 2023 · Research study
Scaling Laws for Reward Model Overoptimization ↗ (opens in a new tab)Examines optimization of an imperfect proxy reward with an experimental reference reward; not a complete model of human values.
Broader frameworks and background
These resources provide context. They do not establish every claim in this concern.
A question to discuss
Are higher scores still corresponding to better answers when people check independently?
Keep track of different views, what needs checking, and any next step in your own discussion notes.
Open discussion notes →Principles for the conversation
Before you finish · reflect together
Virtues and principles to consider
Virtues are qualities we practice in how we treat people and make choices. They are not scores for judging someone. Use these ideas to explore what a thoughtful response could look like.
Truthfulness
Truthfulness means making claims that match the evidence and being clear about what we do not know. It includes correcting mistakes rather than protecting a polished story.
Check sources, label guesses, and distinguish a trial result from a proven benefit. If AI helped produce a claim, someone still needs to take responsibility for checking it.
Ask together: What can we honestly say, and what still needs checking?
Justice
Justice means fair treatment, fair procedures, and attention to who receives benefits and burdens. It asks whether people with less power can be heard and seek repair.
Look at outcomes across different groups and individual cases. Provide a usable way to challenge decisions and correct harm. Equal wording in a policy does not guarantee equal treatment in practice.
Ask together: Who benefits, who bears the cost, and who can challenge the decision?
Stewardship
Stewardship means caring for resources and systems that others depend on, including people who will use them later. It looks beyond the benefit of a quick launch.
Plan for maintenance, energy use, data protection, and eventual shutdown. Ask who will look after the tool when the original team moves on and how future users can avoid inheriting hidden costs.
Ask together: What are we responsible for caring for over time?
Editorial ratings
These are editorial judgments from the register, not measured probabilities or predictions about your situation.
About the register and scoring →More ways to explore
Search for more writing
For more specific or recent material, open a search below or ask for a web summary. AI-generated summaries can contain mistakes; open the original articles to check them.