Strategy-Conditioned Policy for Pre-IMO Problems
Precede proof generation with a strategy classification stage to condition the solver on high-level principles before attempting algebraic manipulation.
Time
Compute level
Resources
Prerequisites
Scope
The Gap
Appendix B.4.3 highlights that models failed completely (0 score) on 'easy' pre-IMO problems involving invariants (coloring) or symmetric inequalities because they attempted complex algebraic manipulations instead of applying simple combinatorial principles.
Limitation
Models exhibit 'technique mismatch,' applying heavy-handed computational methods to problems requiring specific high-level insights (e.g., Pigeonhole Principle, Invariance).
Proposed Solution
Develop a two-stage reasoning architecture. Stage 1: A classifier trained to identify the necessary problem-solving strategy (e.g., 'Use Invariance', 'Use Induction') before generating the proof. Stage 2: A solver conditioned on this strategy. This forces the model to plan the high-level approach explicitly, preventing it from getting lost in irrelevant algebra.
Why This Matters
Solving Olympiad geometry and combinatorics requires 'Aha!' moments (insights) rather than just compute. Explicit strategy selection bridges the gap between perception and formal reasoning.
How to Get Started
How to test it: We hypothesize that conditioning generation on an explicit strategy classifier improves proof scores on IMO-ProofBench. We measure performance using average 0-7 human-correlated scores. If stage-1 classification accuracy falls below 70% or degrades final proof correctness, we will stop.
- Annotate a dataset of pre-IMO competition problems with high-level solving strategies such as invariance and induction.
- Train a lightweight classifier to predict the correct problem-solving strategy from the problem statement.
- Condition the proof generation architecture on the predicted strategy token.
- Evaluate the two-stage framework on pre-IMO benchmarks, measuring improvements in success rate over unconditioned models.
Code linked to this paper on Hugging Face: github.com/google-deepmind/superhuman
Suggested by AI. Adapt it to your setup.
Has This Been Done?
The provided literature search results contain papers related to maritime governance or technical requirements for the IMO (International Maritime Organization), as well as 'IMO^3: Interactive Multi-Objective Off-Policy Optimization'. None of these papers propose a strategy-conditioned policy for solving mathematical Olympiad problems.
- IMO^3: Interactive Multi-Objective Off-Policy Optimization · 2022 · 5 citations
- Towards Solving More Challenging IMO Problems via Decoupled Reasoning and Proving · 2025 · 10 citations
We searched OpenAlex for “Precede proof generation with a strategy classification stage to condition the solver on high-level principles before attempting algebraic manipulation.” and “Strategy-Conditioned Policy for Pre-IMO Problems” and an AI compared the idea with the top 6 results. A small search can miss relevant work.
Source Paper
Towards Robust Mathematical ReasoningThang Luong, Dawsen Hwang, Hoang H. Nguyen +17 more · 2025
This paper introduces IMO-Bench, a challenging benchmark suite designed to evaluate mathematical reasoning at the International Mathematical Olympiad (IMO) level, comprising tasks for short-answer verification (IMO-AnswerBench), proof writing (IMO-ProofBench), and proof grading (IMO-GradingBench). The authors demonstrate that current frontier models, while improving, still struggle with robustification (reformulated problems) and subtle logical fallacies in proofs, proposing a new gold-standard model (Gemini Deep Think) that achieves significant gains through agentic search and verification.
Related Ideas
- BitDistill for Reasoning-Intensive DomainsBitNet Distillation🟠Medium
- Hybrid Neuro-Symbolic Planning for Combinatorial ConstraintsDeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints🟡Fairly easy
- Distilling Explicit Reasoning into Latent SpaceRobust-R1: Degradation-Aware Reasoning for Robust Visual Understanding🟡Fairly easy
- Enhancing Abstract Visual Logic via Synthetically Augmented Multi-Step Math SFTPenguin-VL: Exploring the Efficiency Limits of VLM with LLM-based Vision Encoders🟠Medium
- Integration of Online Reinforcement Learning (PPO/GRPO)BitNet b1.58 2B4T Technical Report🟠Medium