The Pulse
AI Helped Tighten the Grothendieck Constant’s Bounds
A Princeton research team used AI to help tighten bounds on the Grothendieck constant, with researchers stepping in when the system repeated failed approaches. The case study describes what the system proved, what the authors verified, and
AI.info Team · Written with AI ·
On October 13, Princeton University described a mathematical research effort in which AI helped tighten bounds on the Grothendieck constant—but needed people to recognize what its repeated failures meant. The case study, posted to arXiv in August, records a system that could develop intricate mathematics once researchers chose a direction, yet struggled to judge when a line of work had run its course. The distinction, its authors argue, matters as AI moves from answering defined questions toward carrying out extended research.
A constant narrowed by human-AI work
The Grothendieck constant, first identified by mathematician Alexander Grothendieck in 1953, measures a gap between certain discrete optimization problems and their continuous relaxations. The research team’s mathematical companion paper reports improved lower and upper bounds: 6π/11 ≤ KG ≤ π/(2 log(1+√2)) − 10−4. Together, the bounds establish the previously unknown tenths digit of the constant as 7.
The AI case study centers on the lower-bound result, KG ≥ 6π/11. The authors say the system discovered and first proved that bound; they later checked the argument themselves and revised its write-up. The companion paper presents the mathematical results, while the case study recounts how the work unfolded.
Six failed approaches pointed to a new question
The research system explored ways to improve an upper bound by varying a family of mathematical schemes. Across six distinct attempts, it encountered a recurring obstacle, but kept trying new variations rather than asking whether the obstacle applied to the entire family. At session 18, the human operators redirected it: instead of looking for another construction, it should synthesize the failed attempts and try to prove a general obstruction.
That shift changed the task. A general obstruction would itself produce a lower bound, turning the stalled search into a different mathematical result. With that framing and further guidance, the system developed the central argument; the researchers then verified it independently. The case study says the intervention supplied no new mathematical result—it connected evidence the system had already accumulated to a more valuable research question.
“The intervention supplied no new mathematics: the problem statement had suggested this dualization from the outset, and the run had accumulated exactly the required evidence without making the connection.”
Alan Li, Rahul Saha, Anton Xue, Swarat Chaudhuri, Adam Klivans, Pravesh K Kothari and Raghu Meka, authors of the case study
About 240 sessions, with people setting priorities
The run lasted from June 16 to July 24, 2026, and comprised roughly 240 research sessions. It used OpenAI’s GPT-5.5-Pro as the reasoning model before switching mid-run to GPT-5.6-Sol, alongside Anthropic’s Claude Code as a coding agent. The reasoning model was called 2,091 times and used about 152 million tokens; the authors estimate its API cost at about $5,400, while the coding agent ran on a subscription.
People issued about 40 directives, setting priorities, redirecting the search and auditing claims. The researchers describe the system as effective at developing proofs and running experiments after a high-level action had been chosen. Its weaknesses appeared in choosing that action and preserving important caveats across sessions: a numerical result was briefly treated as a record before an internal audit found it unsupported and withdrew it.
The result is not a case for autonomy
The authors caution that their system worked downstream of a mathematical framework, prior notes and candidate directions supplied by the team. They also distinguish the lower bound they verified from other system-tested results that they had not yet checked themselves. The case study is therefore evidence of a productive collaboration in one research project, not a demonstration that the AI independently selected and completed a research agenda.
For the Princeton researchers, the episode exposes a specific limit: producing technical work is not the same as recognizing its significance. The system could turn a human-selected reframe into a proof, but it did not reliably infer on its own that six similar failures might add up to a theorem.