The frame problem is the challenge of describing the effects of an action in a formal logic without explicitly stating everything the action leaves unchanged. Artificial intelligence researchers John McCarthy and Patrick Hayes identified it in 1969 while building the situation calculus, and it remains one of the classic topics of AI fundamentals because it exposes how hard it is for a machine to reason about a changing world.
In its narrow, logical form, the frame problem asks how a knowledge base can infer that painting a block does not move it, without one axiom for every pair of action and property. In its broader, epistemological form, discussed by philosophers such as Daniel Dennett and Jerry Fodor, it asks how any intelligent system decides which facts are relevant to an action at all.
How the frame problem arises in logical action theories
The frame problem arises because classical logic is monotonic: a conclusion holds only if it follows from stated axioms, so nothing persists by default. When an agent describes the action paint(block, red) with an effect axiom that says the block becomes red, the logic cannot conclude that the block keeps its position, its weight or its owner unless separate axioms say so.
These extra statements are called frame axioms. Their number grows roughly with the number of actions multiplied by the number of fluents, where a fluent is a property whose value can change from one situation to the next. A domain with 100 actions and 500 fluents could need on the order of 50,000 frame axioms, most of them saying that nothing happened.
The following table separates the core vocabulary that the frame problem relies on.
| Concept | Meaning | Role in the problem |
|---|---|---|
| Situation | A snapshot of the world after a sequence of actions | Unit over which facts are evaluated |
| Fluent | A property that can vary between situations | The thing that may or may not persist |
| Effect axiom | States what an action changes | Covers the positive effects only |
| Frame axiom | States what an action leaves unchanged | Source of the combinatorial explosion |
| Inertia | The assumption that facts persist unless changed | The intuition that solutions must capture |
Related problems that share the same root
The frame problem belongs to a family of representation problems that all stem from describing actions in an open, changing world. Researchers usually distinguish it from two close relatives, the qualification problem and the ramification problem, because each one fails in a different direction.
| Problem | Central question | Typical example |
|---|---|---|
| Frame problem | What does an action not change | Moving a cup does not change the wall color |
| Qualification problem | Which preconditions must hold for success | A car starts unless a potato blocks the exhaust |
| Ramification problem | Which indirect effects follow from an action | Moving a box also moves the items inside it |
| Relevance problem | Which facts deserve attention at all | A robot ignores the moon phase when lifting a tray |
The broader relevance problem is the version philosophers emphasize. It also connects to commonsense reasoning and the symbol grounding problem, since a system that cannot judge relevance cleanly tends to either overthink trivial consequences or overlook important ones.
Formal solutions to the frame problem
The narrow logical frame problem is largely considered solved for well-specified domains. The main solutions encode the commonsense law of inertia compactly, either by changing the logic or by restructuring the axioms so that persistence follows automatically.
| Approach | Core idea | Main limitation |
|---|---|---|
| STRIPS representation | Actions list add and delete effects; all else persists | Weak for conditional and indirect effects |
| Circumscription | Minimize abnormal changes in models | Can yield unintended minimal models |
| Default logic | Persistence holds unless contradicted | Multiple extensions complicate inference |
| Successor state axioms | One axiom per fluent lists all causes of change | Requires a complete list of causes |
| Event calculus | Fluents persist between initiating and terminating events | Needs careful handling of concurrency |
| Answer set programming | Inertia rules with negation as failure | Grounding cost on large domains |
Raymond Reiter’s successor state axioms reduce the axiom count to one axiom per fluent, which makes the representation roughly proportional to the size of the domain rather than its square. The famous Yale shooting problem of 1986 showed that naive nonmonotonic logic could still predict absurd outcomes, which pushed the field toward chronological minimization and causal theories.
The frame problem in modern AI systems
Modern systems still face the frame problem, although it rarely appears under that name. Automated planning languages such as PDDL inherit the STRIPS assumption, so any planner built on them solves the narrow problem by design while accepting its expressive limits.
Learned systems face a statistical version of the problem. A world model in reinforcement learning must predict which parts of a state persist after an action, and errors in that prediction compound over long rollouts. Large language models used as AI agents often lose track of unchanged state in multi-step tasks, for example forgetting that a file was already edited, which is why practitioners add explicit state tracking outside the model.
- Symbolic planners handle persistence exactly but require a hand-built model.
- Neural world models learn persistence from data but can drift over long horizons.
- Neuro-symbolic designs combine learned perception with explicit state representations.
Common misconceptions about the frame problem
The term is often stretched beyond its original meaning, so precision about scope matters. The table below lists frequent misreadings and the more accurate view held by specialists.
| Misconception | More accurate view |
|---|---|
| It is an unsolved logic puzzle | The narrow version is solved for closed domains |
| It concerns visual framing or cropping | It concerns persistence of facts over actions |
| Deep learning makes it irrelevant | It reappears as state drift in learned models |
| It is only about memory limits | It is about deciding relevance, not storage |
The most important distinction is between the technical frame problem, which has practical engineering answers, and the philosophical frame problem, which questions whether any finite system can decide relevance in an unbounded world. The second remains genuinely open and is closely tied to debates about artificial general intelligence.
In simple terms
When you move a coffee cup, you know that the walls, the chairs and the weather stay the same. You never think about it. A computer that reasons with strict logic does not know this automatically. It needs to be told, fact by fact, that everything else stays as it was.
The frame problem is the puzzle of teaching a machine that most things do not change when it acts, without writing an endless list. The key takeaway is that knowing what to ignore is a large part of intelligence, and it is surprisingly hard to give that skill to a machine.
A practical example
A warehouse robotics team builds a planner for a robot that moves boxes between 40 shelf locations. Each box has fluents for location, weight, label and whether it is sealed. The first prototype uses plain first-order axioms, and the planner keeps failing because it cannot prove that a box stays sealed after the robot moves a different box.
The engineers switch to a STRIPS-style action model in which each action lists only its add and delete effects. A move action now changes only the location of the moved box, and every other fact persists automatically. The model shrinks from several thousand axioms to about 30 action definitions.
A ramification still appears when boxes are stacked, because moving the bottom box also moves the boxes on top. The team adds a derived rule for stacked items, and the planner then produces correct multi-step plans in well under a second.