Do language models fail at identifying unstated preconditions?

The classical frame problem (McCarthy and Hayes, 1981) asks how a reasoning system decides which background conditions are relevant when reasoning about an action. Most things stay the same when an action is performed — the frame — and the system needs to know which non-trivial things change without being told. Solving this in symbolic AI required either explicit frame axioms (combinatorially expensive) or non-monotonic logics (mathematically delicate).

The Heuristic Override Benchmark identifies a contemporary version of the same problem in LLMs. When a user asks "should I walk or drive to the car wash 50m away," the relevant unstated condition is that the car must be at the car wash. This is a feasibility precondition that no human would need to state because it is presupposed by the entire setup. The model fails not because it lacks the world knowledge — it has it — but because it does not bring this background condition forward as relevant when the surface heuristic ("50m is walkable") is active.

This reframes the failure. It is not noise filtering (the standard shortcut-learning frame). It is not knowledge retrieval (the standard hallucination frame). It is enumeration: which of the indefinitely many things I know about the world should I treat as live constraints on this decision? Structured prompting that forces enumeration ("what must be true for walking to be feasible?") raises accuracy from around 30 percent to 85 percent on single instances. The intervention works precisely because it externalizes the enumeration step the model cannot reliably perform on its own.

The frame problem was once thought specific to symbolic systems. The HOB results suggest it persists, in different form, in statistical systems trained on language. The substrate changed; the structural problem did not.

Source: Reasoning Critiques