Skip to content

HOA — Humanize Olympic Agents

Competition mathematics, solved by agents and checked by a proof assistant. No rubric, no grader, no benefit of the doubt: Lean 4 either accepts the proof or it does not.

6 / 6IMO 2026 problems, every solution formally verified in Lean 4IMO 2026
99.7%PutnamBench: 670 of 672 formal statements, ranked #1 on the official leaderboardPutnamBench
3.0×Faster than the previously reported agentic result on IMO 2026, at less than half the costEfficiency

IMO 2026: six of six

Running a fully agentic, YOLO-style loop on Humanize, we solved all six problems of the 2026 International Mathematical Olympiad, and every solution is machine-checked in Lean 4 against the formal statements published by AxiomMath.

Times below are API time — the time actually spent inside model calls — compared with the times reported by AxiomProver on the same statements.

ProblemHumanfia (GPT-5.6)Humanfia (Kimi-K3)AxiomProver
Total497.7 min1,034.9 min1,496 min
Q138.1 min87.1 min24 min
Q2100.4 min224.3 min360 min
Q3187.1 min343.7 min869 min
Q458.7 min75.6 min39 min
Q546.5 min91.9 min65 min
Q666.9 min212.4 min139 min

The Kimi-K3 run is a hybrid — a Kimi worker paired with a Codex reviewer — and cost $57.70 in total, a 54.7% saving against the GPT-5.6 run's $127.41. That is the part worth sitting with: two different backends, the same loop, and the cheaper one still closes all six.

Everything is released: the formal statements, both sets of Lean solutions, and the scripts to reproduce the solving process. Pinned to Lean 4.31.0 and Mathlib.

humanfia/imo2026

PutnamBench: 670 of 672

The same approach across the whole of PutnamBench — all 672 formal statements — produced 670 verified Lean proofs, a 99.7% pass rate, and first place on the official leaderboard. The result has been verified by the PutnamBench team.

Verified proofs670 of 672
Pass rate99.7%
Ranking#1 on the official leaderboard
Worker / reviewer modelgpt-5.5, reasoning effort xhigh
ToolchainLean 4.27.0, Mathlib pinned
AcceptanceLean kernel, Comparator and AXLE — all three must agree

A problem counts as solved only when the Lean file passes every gate. Candidates that fail any gate are kept for inspection and are never counted. The fastest way to check us is the AXLE verification API, which needs Python 3 and a network connection and nothing else — no Lean, no Mathlib, no model calls, no API key.

At the PutnamBench authors' request the full solution set is not open-sourced; a preview of the first twelve is published for review, and the entire solving pipeline is public.

humanfia/putnambench-solver

Live competitions

Formal mathematics is the cleanest test but not the only one. We also run these loops on live Kaggle competitions and publish what comes of it, including the parts that need caveats — which results are official final ranks, which are late-submission estimates against a frozen leaderboard, and what a snapshot of an ongoing competition is and is not worth.

agentkaggle/leaderboard · the audit

What this is really testing

None of this is a claim that agents are good at mathematics. It is a claim about loops.

The models here are the ones everybody has. What differs is the arrangement around them: who works and who reviews, what carries between attempts and what is deliberately forgotten, when a line of attack is abandoned, and how a run of hundreds of hours is kept from going in circles. A formal verifier is the ideal instrument for measuring that, because it removes every way of being approximately right.

The loop used for these runs is Humanize 1. What we learned from watching it run for weeks is most of why RLAR and the rest of the flowverse look the way they do.