Human-guided AI advances a Grothendieck bound over weeks of research

A human-guided AI research system helped derive a new lower bound for the Grothendieck constant, a longstanding quantity connecting difficult discrete optimisation problems with more tractable continuous ones. The result raises the reported lower bound to 6π/11, approximately 1.7136.[1][2]

An August case study documents the extended collaboration rather than a single successful prompt. The system ran from 16 June to 24 July, using more than 240 sessions and 2,000 reasoning-model calls. Human researchers periodically reviewed its progress and changed its direction. The current version of the paper is dated 14 August.[1][2]

The bound measures a gap between two kinds of problem

In the optimisation setting described by the authors, a hard problem assigns positive or negative signs to maximise a mathematical expression. A continuous relaxation replaces those signs with vectors, making a related problem easier to solve. The Grothendieck constant limits how far the relaxed answer can exceed the original optimum. Its exact value remains unknown.[2]

Improving a lower bound shows that the gap must be at least that large. The team’s approach did this by placing a limit on an entire class of rounding schemes, rather than constructing a single difficult instance. The authors say they independently checked the resulting argument and revised it for the companion mathematical paper.[2]

The case study also discusses an improved upper bound. That construction came from earlier conversations with GPT-5.5-Pro and predates the long-running research system. The chronology matters: the upper- and lower-bound contributions should not both be credited to the same multi-agent run.[2]

Human intervention turned a failed search into a different question

The system divided work between a reasoning model and a coding agent. Files carried the evolving research state, and a separate bulletin let human operators steer the work. Numerical checks and interval arithmetic supplied stronger tests than a model’s own confidence in a calculation.[2]

By session 18, an upper-bound search had stalled after several unsuccessful variations. The researchers directed the system to turn those failures into a statement about what the whole class of methods could not do. That changed the task from finding a better construction to proving a limitation—a route toward a lower bound.[2]

Later, the system spent repeated sessions refining restricted results without closing the general argument. Another human intervention called for theoretical work on the missing connection. The eventual reframe used quantities whose relationship survived the mixtures and limits needed by the proof.[2]

A long research run also needs an accurate memory

The record includes a numerical upper bound that was later found unsupported by machine testing and withdrawn. Other stronger candidates remained machine-tested rather than human-verified. The paper keeps those statuses separate from the theorems the authors endorse.[2]

That distinction was not always preserved inside the system. The authors report that repeated rewriting of the research state could retain an impressive number while losing qualifications about its evidence. A memory designed for ongoing research therefore needs to retain failed approaches, unresolved dependencies and the verification status of each claim.[2]

The case study shows a productive division of work: models performed substantial technical exploration, while researchers recognised when local progress was no longer advancing the problem and checked the result they would publish. The new bound is the mathematical output; the detailed record explains how that collaboration reached it.[2]

Report an error or suggest a correction