Inference Horizon

Opening

Inference horizon is the term Apple's 2025 "Illusion of Thinking" paper coined for the complexity threshold past which a large reasoning model can no longer follow its own chain of thought to produce correct answers. The paper observed that for every LRM tested (including frontier models like o3), accuracy degraded past a model-specific complexity threshold on controllable logic puzzles, and argued that this threshold — not the model's apparent reasoning ability — is the operational ceiling of LRM capability. The term has since become standard in the LRM discourse for describing where step-by-step reasoning fails.

Mechanism

Inference horizon manifests empirically as a curve of accuracy-vs-complexity on puzzles where complexity is a clean independent variable (Tower of Hanoi with N disks, river crossing with N pairs, block stacking with N blocks). Below the horizon, the model reliably solves the puzzle. Near the horizon, the model attempts to reason but produces incorrect solutions — the chain of thought runs longer but does not converge. Above the horizon, the model collapses entirely and may even produce nonsensical intermediate tokens or refuse to attempt the puzzle. The mechanism the paper attributes this to is bounded working memory: the model's chain of thought has a maximum reliable length, and puzzles requiring more than that length push the model past its step-following capability. Apple's specific claim is that the horizon is not a continuum that improves smoothly with model scale but a relatively hard cliff.

Tradeoffs

The trade-off the inference-horizon concept surfaces is between the verifiable-domain successes that LRMs deliver (math olympiad problems, code with test suites, formal proofs with Lean verification) and the brittleness the horizon reveals (Tower of Hanoi past 8 disks, planning problems with more than a fixed number of steps). The concept also exposes a measurement problem: LRMs marketed as reasoning-capable tend to be evaluated on benchmarks where their horizon is large enough to look impressive, while the horizon itself remains undocumented. The trade-off for the field is between accepting the horizon as a fundamental architectural limit (and designing around it with tool-use and external verification) and treating it as an engineering challenge that newer model generations will push further out.

Open Questions

Whether inference horizon is a transformer architectural limit or a training-data / training-objective artifact that improves with scale. Whether the horizon is approximately universal across reasoning tasks (a model with horizon N on Tower of Hanoi also has horizon N on planning problems) or task-specific. Whether tool-use, scratchpad memory, or external verification (Lean, code interpreters) effectively extend the horizon, or whether they substitute for reasoning in ways that produce different failure modes. Whether newer model generations are observably pushing the horizon further or whether the horizon remains roughly stable as model capability improves in other dimensions.