VibeMathedMath problems solved with AI
← All frontiers
FrontierNumber theorysignificance 30

Runners for which the lonely runner conjecture is proved

the largest nn such that the lonely runner conjecture holds for every nnn' \le n runners

Current best · higher is better
n=13n = 13
Touch Sungkawichai and Tanupat Trakulthongchai, with AI-assisted code and one lemma sketched by ChatGPT-5.6 Pro, 26 Apr 2026
steps
7
by AI
0
since
1972
1980199020002010202046810121972: n = 4 (Ulrich Betke and Jorg M. Wills)1984: n = 5 (Thomas W. Cusick and Carl Pomerance)2001: n = 6 (Tom Bohman, Ron Holzman and Daniel Kleitman)2008: n = 7 (Javier Barajas and Oriol Serra)2025-09-17: n = 8 (Matthieu Rosenfeld)2025-12-01: n = 9 (Matthieu Rosenfeld)2026-04-26: n = 13 (Touch Sungkawichai and Tanupat Trakulthongchai, with AI-assisted code and one lemma sketched by ChatGPT-5.6 Pro)2025-11-27: n = 10 (Tanupat Trakulthongchai, with GPT-5) - candidate, under review

The line is the frontier over time, climbing as the bound goes up: higher is better here. Filled dots are steps that moved it; muted dots are results that did not. Orange dots are catalog entries, results with AI in the loop. Hollow dots are candidates under review and never move the line.Grey dots along the bottom edge are results from before the quantity had a number, placed at the worst end because they have no value on this axis. Dots that would overlap are nudged sideways a few pixels. Hover a dot for its value and attribution.

About this frontier

Wills in 1967, and Cusick independently in 1974, conjectured that if nn runners start together on a circular track of unit length at pairwise distinct constant speeds, then each runner is at some time at distance at least 1/n1/n from all the others. Trivial for n3n \le 3, it has fallen one or a few runners at a time: 4 in 1972, 5 in 1984, 6 in 2001, 7 in 2008, then, after a seventeen-year pause, a computer-assisted method built on Tao's finite-checking reduction settled 8 in September 2025 and has been extended almost monthly since, to 9 and 10, and to 13 by April 2026. The conjecture for all nn is open; this frontier tracks how far the case-by-case verification has reached.

Every step, newest first

DateValueWhoModelStatusSource
26 Apr 2026n=13n = 13best
Eleven, twelve and thirteen runners, extending the same computational method with parallel execution and stronger pruning. The paper says its code development was AI-assisted and prints a proof sketch it attributes to ChatGPT-5.6 Pro, so this is an AI-in-the-loop result without a catalog entry yet. Drawn as a cited row until it has one.
Touch Sungkawichai and Tanupat Trakulthongchai, with AI-assisted code and one lemma sketched by ChatGPT-5.6 Prohistoricalsource ↗
1 Dec 2025n=9n = 9
Nine runners, independently, four days after the entry's nine and ten, by improvements to his own eight-runner method. On the reviewed track this is the record until April 2026.
Matthieu Rosenfeldhistoricalsource ↗
27 Nov 2025n=10n = 10
Nine and ten runners, by adding a sieve to Rosenfeld's verification; GPT-5 assisted the C++ implementation, per the paper's Section 6. Code and result receipts are public; no independent rerun has appeared, so the entry is unreviewed and this row is a candidate.
Tanupat Trakulthongchai, with GPT-5GPT-5candidate
unreviewed
entry
17 Sept 2025n=8n = 8
The method every later row uses: Tao's 2018 reduction to a finite check, sharpened by Malikiosis, Santos and Schymura in 2025, then a computer verification. Rosenfeld's abstract predicted that minor improvements would reach 9 or 10; they did within ten weeks.
Matthieu Rosenfeldhistoricalsource ↗
2008n=7n = 7
The lonely runner with seven runners, Electronic Journal of Combinatorics 15 (2008). The record for seventeen years.
Javier Barajas and Oriol Serrahistoricalsource ↗
2001n=6n = 6
Six lonely runners, Electronic Journal of Combinatorics 8 (2001). Seventeen years after five.
Tom Bohman, Ron Holzman and Daniel Kleitmanhistoricalsource ↗
1984n=5n = 5
View-obstruction problems III, Journal of Number Theory 19 (1984). Computer-assisted at the time; Bienia, Goddyn, Gvozdjak, Sebo and Tarsi gave an elementary proof in 1998.
Thomas W. Cusick and Carl Pomerancehistoricalsource ↗
1972n=4n = 4
Monatshefte fur Mathematik 76 (1972), in the Diophantine-approximation language Wills posed the conjecture in five years earlier. The cases n <= 3 are elementary.
Ulrich Betke and Jorg M. Willshistoricalsource ↗

Rows from the Wikipedia article's per-n account and its works-cited list, each reference then checked at Crossref or arXiv; the four 2025-2026 papers were opened and their abstracts and AI disclosures read. Rosenfeld's 9 (1 December 2025) came four days after the entry's 9 and 10 and is drawn as a row that did not move the line, since the entry's row is a candidate: unreviewed, so hollow. The April 2026 paper reaching 13 discloses AI-assisted code and a lemma sketched by ChatGPT-5.6 Pro; it is drawn as a cited historical row because it has no catalog entry yet, and it should get one.

Discussion