Towards Solving More Challenging IMO Problems via Decoupled Reasoning and Proving

Zhenwen Liang, Linfeng Song, Yang Li, Tao Yang, Feng Zhang, Haitao Mi, Dong YuarXiv 2025 · OpenReview

Abstract

Automated Theorem Proving (ATP) in formal languages is a foundational challenge for AI. While Large Language Models (LLMs) have driven remarkable progress, a significant gap remains between their powerful informal reasoning capabilities and their weak formal proving performance. Recent studies show that the informal accuracy exceeds 80% while formal success remains below 8% on benchmarks like PutnamBench. We argue this gap persists because current state-of-the-art provers, by tightly coupling reasoning and proving, are trained with paradigms that inadvertently punish deep reasoning in favor of shallow, tactic-based strategies. To bridge this fundamental gap, we propose a novel framework that decouples high-level reasoning from low-level proof generation. Our approach utilizes two distinct, specialized models: a powerful, general-purpose Reasoner to generate diverse, strategic subgoal lemmas, and an efficient Prover to rigorously verify them. This modular design liberates the model's full reasoning potential and bypasses the pitfalls of end-to-end training. We evaluate our method on a challenging set of post-2000 IMO problems, a problem set on which no prior open-source prover has reported success. Our decoupled framework successfully solves 5 of these problems, demonstrating a significant step towards automated reasoning on exceptionally difficult mathematical challenges.

The Performance Chasm in AI Mathematics

There is a massive performance gap in AI mathematics. Top-tier Large Language Models (LLMs) can solve over 80% of difficult math problems when writing in informal natural language. However, when asked to produce formal proofs—code that a computer can verify, such as in the Lean programming language—even the best models drop to less than 8% accuracy.

Current training methods, specifically Reinforcement Learning with Verifiable Rewards (RLVR), may actually hinder deep mathematical reasoning. Because these models are rewarded primarily for passing the compiler, they often learn to rely on shallow, brute-force "tactic hacking"—trying every possible command until something works—rather than developing strategic insights. This leads to "contextual blindness," where a model fails to grasp the overarching structure of a complex proof.

DRP-IMO: A Decoupled Architecture

To address these issues, we propose the DRP-IMO (Decoupled Reasoning and Proving) framework. Instead of a single model attempting to solve a problem end-to-end, the process is split into three distinct stages:

  1. Strategic Subgoal Generation (The Reasoner): A powerful, general-purpose LLM analyzes the problem and generates a roadmap of formal lemma statements. It defines the logical milestones without attempting the full proof immediately.
  2. Subgoal Verification and Filtering (The Prover): A specialized, efficient ATP model attempts to prove each individual lemma. Verified lemmas serve as sound "building blocks."
  3. Final Proof Construction: The Reasoner assembles the final, complete proof for the main theorem using the set of verified lemmas.
The DRP-IMO framework decouples high-level reasoning from low-level proof generation through subgoal generation and verification.

The "Reasoning Degradation" Hypothesis

Our research identifies a "degradation" effect: fine-tuning a model specifically for formal proving (like Lean) can hurt its general mathematical reasoning. Experiments showed a performance drop of up to 11.3 points on AIME benchmarks when a base model was specialized into a prover. By decoupling these tasks, we preserve the "raw power" of the reasoning model while leveraging the precision of the formal prover.

Key Results and IMO Success

The DRP-IMO framework successfully solved five difficult post-2000 IMO problems: 2000 P2, 2005 P3, 2011 P3, 2019 P1, and 2020 P2. No prior open-source prover had reported success on this specific set.

Our findings suggest that:

  • Formal Statements as a Bridge: Forcing the Reasoner to output formal Lean statements for subgoals bypasses the errors associated with auto-formalization.
  • Overcoming Contextual Blindness: The Reasoner is significantly better at integrating and applying lemmas to construct final proofs compared to models trained purely on code.

Conclusion and Future Directions

While DRP-IMO represents a significant step forward, challenges remain. The system still struggles with problems requiring highly creative "leaps" or "magical insights." Future work will focus on fine-tuning specialized provers for "proof continuation"—the ability to effectively use a library of provided lemmas—and improving the computational efficiency of the multi-stage pipeline.

To foster future research, we have released our full dataset of generated and verified lemmas for a wide range of IMO problems at https://tencent-imo.github.io/.

Summary updated

Abstract

Automated Theorem Proving (ATP) in formal languages is a foundational challenge for AI. While Large Language Models (LLMs) have driven remarkable progress, a significant gap remains between their powerful informal reasoning capabilities and their weak formal proving performance. Recent studies show that the informal accuracy exceeds 80% while formal success remains below 8% on benchmarks like PutnamBench. We argue this gap persists because current state-of-the-art provers, by tightly coupling reasoning and proving, are trained with paradigms that inadvertently punish deep reasoning in favor of shallow, tactic-based strategies. To bridge this fundamental gap, we propose a novel framework that decouples high-level reasoning from low-level proof generation. Our approach utilizes two distinct, specialized models: a powerful, general-purpose Reasoner to generate diverse, strategic subgoal lemmas, and an efficient Prover to rigorously verify them. This modular design liberates the model's full reasoning potential and bypasses the pitfalls of end-to-end training. We evaluate our method on a challenging set of post-2000 IMO problems, a problem set on which no prior open-source prover has reported success. Our decoupled framework successfully solves 5 of these problems, demonstrating a significant step towards automated reasoning on exceptionally difficult mathematical challenges. To foster future research, we release our full dataset of generated and verified lemmas for a wide range of IMO problems, available at https://tencent-imo.github.io/ .
BibTeX
@misc{liang2025towards,
  author = {Zhenwen Liang and Linfeng Song and Yang Li and Tao Yang and Feng Zhang and Haitao Mi and Dong Yu},
  title = {{Towards Solving More Challenging IMO Problems via Decoupled Reasoning and Proving}},
  year = {2025},
  eprint = {2507.06804},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.06804},
}
Figure