Large language model (LLM) agents are rapidly advancing on long‑horizon scientific and Olympiad tasks. Prior systems have produced a Lean‑formalized solution to the Navier–Stokes problem and autonomous answers to four open questions in the Erd\H{o}s Conjectures database, yet they do not release the full model and inference stack. A single agent session remains fragile on hard problems: early mistakes propagate, missing domain knowledge stalls progress, and apparent success can hide misinterpretations of the original task. We introduce HOA (Humanize Olympiads Agent), an open flow‑level framework that coordinates multiple sessions via three complementary feedback channels. First, an independent reviewer evaluates each candidate and supplies actionable feedback. Second, a structured wiki provides task‑specific knowledge that may be absent or unreliable in the model’s pretraining. Third, a verifier checks generated Lean proofs against trusted challenge statements. With GPT‑5.6 Sol, HOA achieves perfect scores on the IMO, IPhO, and IOI 2026, near‑full marks on the IChO 2026 set, and solves all 672 PutnamBench problems (100%). On the Lean‑Eval leaderboard, HOA ranks first, solving 251 of 304 problems (82.6%). Using the open model Kimi‑K3, HOA attains full scores on the IMO, IPhO, and IOI, and a gold‑level 97.1% on the IChO. These results demonstrate that systematic feedback can bridge substantial gaps in base‑model capability.
Review