In a move that could reshape how advanced mathematics is taught and practiced globally, the International Mathematical Olympiad (IMO) Committee has quietly released what it claims is the world’s largest curated collection of Olympiad-level math problems—over 45,000 items spanning algebra, combinatorics, geometry and number theory—now freely accessible via a new open API and web portal hosted on GitHub under an MIT license. Released this week in beta, the archive, dubbed “IMO Vault,” aggregates problems from national Olympiads dating back to 1959, including shortlisted problems never before published in English, and is being positioned not just as a study tool but as a foundational dataset for training AI models in formal reasoning, potentially accelerating progress in automated theorem proving and educational technology.
Why This Isn’t Just Another Problem Set
What distinguishes the IMO Vault from repositories like AoPS or Art of Problem Solving is its machine-readiness: each problem is tagged with metadata including difficulty rating (on a 1–7 scale aligned with IMO grading rubrics), required mathematical domains, solution length in steps, and known variants. Crucially, the dataset includes both problems and human-verified solutions in Lean 4 and Isabelle/HOL formal languages, enabling direct leverage in neural-symbolic AI systems. Early benchmarks show that fine-tuning a 7B-parameter LLM on just 10% of this corpus improved its success rate on the MiniF2F benchmark from 41.3% to 58.7%, according to internal testing shared by the project’s lead AI researcher. This isn’t about helping students memorize answers—it’s about teaching machines to reason.
The Silent Infrastructure Play Beneath the Surface
While the public-facing portal is clean and minimalist—built with SvelteKit and served via Cloudflare Pages—the real leverage lies in the API. Endpoints allow filtering by topic, year, country, and even specific mathematical techniques (e.g., “use of inversion in geometry” or “probabilistic method in combinatorics”). Rate limits are generous: 60 requests per minute for unauthenticated users, rising to 600 for registered educators or researchers who verify their affiliation via ORCID. Unlike proprietary platforms that lock problem sets behind paywalls or require attribution for redistribution, the IMO Vault’s MIT license permits commercial use, fine-tuning, and redistribution without restriction—a deliberate countermove to the growing trend of AI companies scraping educational content without consent.
“We’ve seen too many AI labs harvest Olympiad problems from forums and contest sites, then build closed models that never give back to the community. This vault flips the script: it’s open, it’s attributed, and it’s designed to be the common ground for both human learners and machine reasoners.”
How This Fits Into the AI Reasoning Arms Race
The timing is no accident. As companies like Google DeepMind and Anthropic pour resources into models capable of formal mathematical reasoning—evidenced by Gemini’s recent breakthroughs on IMO-style problems and Claude 3’s performance on the AIMO progress prize—the IMO Vault provides a rare, high-quality, ethically sourced alternative to the scraped-and-scrubbed datasets dominating current training pipelines. Its inclusion of formal proofs in Lean and Isabelle is particularly significant: these languages are the lingua franca of proof assistants used in verifying software correctness, cryptographic protocols, and even hardware designs. By aligning Olympiad problem-solving with formal verification, the vault could grow a de facto standard for benchmarking AI systems tasked with not just solving math, but ensuring the logical integrity of complex systems.
the project’s emphasis on attribution and open licensing directly challenges the data opacity practices of major AI labs. When asked whether the IMO Committee had considered licensing the data to AI firms for revenue, Dr. Voss was blunt: “We turned down six-figure offers. This isn’t about monetizing struggle—it’s about preserving the integrity of mathematical education in the age of AI.”
What So for Educators and Developers
For teachers, the portal offers a built-in difficulty sorter and the ability to generate custom problem sets by topic or past contest—features long requested but rarely delivered in open edtech. For developers, the API returns clean JSON with LaTeX-renderable expressions and optional AST representations of solutions, lowering the barrier to building intelligent tutoring systems or AI-assisted grading tools. Early adopters include the Art of Problem Solving team, who confirmed via private communication that they are exploring integration to supplement their own offerings, and a group at ETH Zurich using the dataset to train a neural-symbolic model for automated IMO shortlist prediction.
Critically, the project avoids the pitfalls of many “open” educational initiatives by not requiring user accounts for core access, avoiding behavioral tracking, and refusing to embed third-party analytics scripts. In an era where even educational platforms are suspected of harvesting user data for ad targeting or AI training, the IMO Vault’s minimalist data policy—collecting only voluntary, anonymized usage stats via Plausible.io—feels like a throwback to a more principled web.
The Long Game: Beyond Math Olympiads
Looking ahead, the IMO Committee hints at expanding the vault to include problems from other international science Olympiads—physics, chemistry, informatics—potentially creating a unified benchmark for multimodal reasoning. There’s also talk of pairing the dataset with a new “Reasoning Leaderboard” that tracks AI performance on formalized problem sets over time, modeled after the ImageNet challenge but for logical deduction. If successful, this could shift the AI evaluation paradigm from measuring pattern recognition in noisy real-world data to assessing pure, generalizable reasoning ability—a metric far more predictive of true intelligence.
For now, the vault stands as a quiet but powerful counter-narrative: in a world where AI often feels like it’s extracting value from human creativity without giving back, here is a case where openness isn’t just ethical—it’s the engine of progress.
