According to a paper published on arxiv.org, researchers have introduced HOA (Humanize Olympiads Agent), an open framework that coordinates multiple language model agent sessions to solve complex olympiad-level problems. The system addresses limitations in single-agent approaches where “early errors can propagate, missing domain knowledge can stall progress, and apparent success can mask a misinterpretation or weakening of the original task,” according to the paper.
HOA operates through three feedback channels: an independent reviewer that evaluates candidates and provides feedback, a structured wiki supplying task-specific knowledge, and a verifier that checks generated Lean proofs against challenge statements, according to arxiv.org.
The system achieved notable benchmark results. According to the source, with GPT-5.6 Sol, HOA achieved “perfect scores on the IMO, IPhO, and IOI 2026, near-full marks on the IChO 2026 problem set, and solves all 672 problems (100%) on PutnamBench.” On the Lean-Eval leaderboard, HOA ranked first, solving 251 of 304 problems (82.6%).
Using the open-source Kimi-K3 model, HOA achieved “full scores on the IMO, IPhO, and IOI, and a gold-level score of 97.1% on the IChO,” according to arxiv.org. The paper notes these results “closely match the closed-source GPT-5.6 Sol” baseline on some competitions while exceeding it on others, demonstrating that “systematic feedback can help bridge substantial differences in base-model capability.”