Skip to content
← All news

News2 min read

23/23 at IPhO 2026, in Lean

All 23 IPhO 2026 theory subproblems formalized and proved in Lean 4, sorry-free. Alongside: 36 of 36 on a blind quantum-algorithms benchmark, and 92.5% end to end on quantum information theory.

This has since moved

QIT is now 40 of 40, and both IPhO 2026 theory runs score 30 of 30 against the official marking scheme. The October update has the details.

Ongoing

The Kimi-K3 run below is still going. Everything else here is complete.

Formal mathematics is where an agent loop can be checked without argument. Physics and quantum information are the interesting extension of that, because the statements have to be formalized before they can be proved, and the formalization is where a run can go wrong in a way that a compiler will happily accept.

IPhO 2026 · theory

23/23

Every subproblem of T1–T3, sorry-free, and the full lake build passes.

QAlg · blind

36/36

Every quantum-algorithm task, from public statements alone.

QIT · end to end

92.5%

37 of 40 quantum-information tasks through both gates.

Two gates, not one ​

sorry is Lean's placeholder for a step nobody has proved. A sorry-free result means every theorem has an actual proof term checked by the Lean kernel. It does not by itself guarantee that the theorem says what the original problem says. So every target passes two gates, and they are counted separately: a semantic review of the formal statement against the problem, and a proof review of the Lean that proves it.

Meaning first · then proof
1 · meaning2 · proofACCEPTED■ wrong claim■ gap in proof■ accepted

Swipe the diagram sideways to see all of it →

1/0 Each target meets two gates: what the statement means, and whether it is proved.

A perfectly checked proof of the wrong statement is the exact failure this pair of gates exists to catch.

Share of tasks through each gate · 29 July 2026
IPhO 2026 theory23 targets
100.0%
100.0%
QAlg36 tasks
100.0%
100.0%
QIT40 tasks
95.0%
92.5%

IPhO 2026: 23 of 23 ​

All 23 theory subproblems (T1–T3) of the 2026 International Physics Olympiad were completed: 23/23 passed both semantic review and proof review, every Lean 4 file is sorry-free, and the full lake build passed.

The change that produced it was a new proof-to-formalization routing. On the same 22-target comparison, it took the run from 20/22 to 22/22 while cutting summed wall time by 62.1%.The routing comparison is from our run logs and is not published; read it as Humanfia-reported. The 23 verified records are in the dataset, as its lean_verified configuration. One additional theory target brought the final run to 23 of 23.

Same targets, new routing

22/22

Up from 20 of 22.

Summed wall time

−62.1%

For the same 22 targets.

A separate 28-target run with a Kimi-K3 worker has solved 23 targets through proof review, with 24 of 28 passing formalization review. The remaining five are in retry or redraft, or have not yet reached proof dispatch.

The dataset

QAlg: 36 of 36 ​

A blind run, given only the public TeX statements and the benchmark's base library and with no access to reference solutions, produced compiling, semantically accepted, Lean-kernel-checked, sorry-free proofs for all 36 quantum-algorithm tasks. The full lake build passed.

QIT: 92.5% end to end ​

On the 40-task blind Quantum Information Theory benchmark: 38 of 40 semantic review passes and 37 of 40 end-to-end proof review passes, for 92.5%. Conditional on semantic acceptance, the proof rate is 97.4%. Once the statement is right, the proof almost always follows. The full lake build passed.

HOA

Read next

News#1 on Lean-Eval v1 first solvesLean-Eval froze its first 128-problem release and moved everything older to an archive. On v1 our GPT-5.6 flow has the most first solves, 30, and is one behind the leader on total; on the archive it has 170 of 171.Read it News30/30 IPhO, 6/6 IOI, 100/100 IBOIPhO 2026 theory 30 of 30, IOI 2026 six of six, IBO 2024 theory 100 of 100, IChO 2026 formalized 68 of 68, and quantum information theory 40 of 40. What each score is, who graded it, and what it does not show.Read it News672/672 on PutnamBench, joint #1GPT-5.6 in a worker–reviewer loop proved all 672 PutnamBench statements in Lean, at $44.50 a problem. The top of the official leaderboard is now a four-way tie, and what separates the four is cost.Read it