SAIR Launches Math Distillation Challenge

🔑 Enhanced Key Takeaways

• •The challenge is co-organized by Fields Medalist Terence Tao of UCLA and Damek Davis of University of Pennsylvania, using the Equational Theory Project (ETP) dataset of 22 million universal algebra true-false problems[1][2][4].
• •Stage 1 requires submitting a 'cheat sheet' of at most 10KB to boost weak open-source models' performance on hard problems, where top models achieve 95% accuracy but weak ones perform near-random[1][2][4].
• •Top 1000 Stage 1 submissions advance to Stage 2 in late April, involving proof generation, counterexamples, or formal proofs using Lean theorem prover[1][2].
• •SAIR provides a Playground for experimentation with custom cheat sheets, model selection, and performance analysis, but results do not count toward the leaderboard[3].
• •As of March 14, 2026, the competition has 39 participants, launched at a π-inspired time on the earliest timezone[5][7].

🛠️ Technical Deep Dive

• •Dataset: 22 million true-false problems from Equational Theory Project (ETP) on universal algebra, testing if a target equation follows from an initial one via algebraic manipulation or counterexample[2][4].
• •Stage 1: Submit ≤10KB human-readable 'cheat sheet' (like an A4 paper summary) to distill knowledge, evaluated on private test set to maximize weak model accuracy[1][2][4].
• •Stage 2: Advanced validation with proofs or counterexamples, potentially using Lean theorem prover to eliminate logical ambiguity[1][2].
• •Playground tools: Customizable private cheat sheets, AI model selection, public training set testing, metrics for runtime/resource usage and verdicts[3].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline