OpenAI reports threshold repetition for entangled games

For every finite two-player game with finite-dimensional entangled value v<1 and arbitrary correlated questions, the maximum probability of winning at least a fraction v+δ of k repetitions decays at least as exp(−κδ¹³k/(1+log(|A||B|))) for one universal κ>0 and 0<δ<1−v.

Sources for this record

What was already known

Entangled players can coordinate across repeated games, so independent-trial concentration does not apply. A uniform threshold parallel-repetition bound for arbitrary finite games and correlated question distributions remained open in the manuscript’s account.

AI involvement AI-led

OpenAI credits itself as manuscript author and states that its internal frontier model produced the results in this release. Individual human input is not detailed enough to establish task-only autonomy.

OpenAI is named as author on the manuscript, and its release attributes the results to an internal frontier model. The individual human input is not detailed enough to establish autonomy.

Importance Major

Gives universal exponential threshold decay for finite-dimensional entangled strategies.

Limitations

The pinned Lean rate has δ to the thirteenth power; the manuscript states a stronger fifth-power rate that is outside this selected formalization. Answer alphabets are A and B, and strategy dimension is unrestricted but finite.

Technical verification Formal (Lean 4)

OpenAI reports a Lean proof matching the selected claim.

Formal checking

The pinned Lean note proves the displayed δ¹³ bound uniformly over every joint finite-dimensional strategy and its supremum.

Checker
OpenAI-reported Lean 4 / Comparator
Scope
For every finite two-player game with finite-dimensional entangled value v<1 and arbitrary correlated questions, the maximum probability of winning at least a fraction v+δ of k repetitions decays at least as exp(−κδ¹³k/(1+log(|A||B|))) for one universal κ>0 and 0<δ<1−v.
Basis
Reported by the linked checker
Evidence source

Checking a formal statement does not establish that it matches the original problem.

Human understanding Few

Few is the default when evidence for a higher level is insufficient. No named person is documented as explaining the central argument for this assessed contribution.

Correction and publication history

2026-10-18 · publication

Primary manuscript, attribution, prior-work, and scoped checking review for non-notifying local inclusion.

Read this published revision