Introduction
Recent advancements in Large Language Models (LLMs) have significantly pushed the boundaries of automated mathematical reasoning. Techniques such as Reinforcement Learning (RL) and Long Chain-of-Thought (CoT) processing—exemplified by models like OpenAI’s o1 and DeepSeek-R1—have enabled models to solve increasingly complex problems. However, researchers have identified a "capability ceiling" where pure RL primarily refines a model's existing knowledge rather than introducing fundamentally new reasoning patterns.
To break through this ceiling, this paper introduces NFL-HR (Natural-Formal Hybrid Reasoning), a framework designed to integrate the structural rigor of Formal Languages (FL), such as Lean4, with the intuitive flexibility of Natural Language (NL). By aligning these two paradigms, the framework achieves state-of-the-art performance, reaching 89.80% on the MATH-500 benchmark.
The Challenge: The Format Gap
Mathematical reasoning in AI typically bifurcates into two distinct approaches:
- Natural Language (NL): Highly flexible and capable of broad generalization, but prone to logical "hallucinations" and inconsistent step-by-step verification.
- Formal Language (FL): Mathematically rigorous and verifiable by computer kernels, but rigid and difficult to apply to open-ended questions.
The primary obstacle to a hybrid approach is the "format gap." NL problems are usually presented as questions (e.g., "Find the value of x"), whereas FL systems like Lean4 require theorems to be proven. Bridging this gap requires a method to translate open-ended inquiry into a closed-ended formal structure without knowing the answer in advance.
The NFL-HR Framework
The NFL-HR framework addresses the format gap through a three-stage pipeline:
1. NL-FL Problem Alignment
The core innovation of NFL-HR is the reformulation of NL questions into existence theorems. Instead of asking the model to find a value $C$, the framework asks the model to "Prove that there exists a value $C$ such that the given conditions hold." This allows the problem to be expressed in Lean4 syntax as an exists statement, making it compatible with formal provers while still allowing the model to search for the specific value during the proof process.
2. Mixed Problem Input
Rather than forcing a choice between NL or FL, the framework utilizes a "Mixed Input" technique. The specialized FL reasoner (Kimina-Prover-7B) receives both the original NL problem and the newly generated FL existence theorem. This allows the model to leverage its formal logic expertise to scaffold its natural language "thinking" process within a single, unified Long CoT.
3. Answer Extraction
Because models trained on formal languages often develop a bias toward outputting code or proof scripts, they may not explicitly state the final numerical answer in a standard format. To resolve this, a general-purpose LLM acts as a final "extractor," parsing the complex CoT output to identify and format the implicit numerical or symbolic answer.
Experimental Results
The authors evaluated NFL-HR using Qwen3-8B as the base model and Kimina-Prover-7B as the reasoner. Testing was conducted on two rigorous benchmarks: MATH-500 (a subset of challenging MATH problems) and AMC (83 problems from the 2022 and 2023 American Mathematics Competitions).
Performance Gains
- MATH-500: The framework achieved 89.80% accuracy, representing a 4.60% absolute improvement over the Qwen3-8B baseline.
- AMC: The framework reached 84.34% accuracy, a 4.82% improvement over the baseline.
Domain-Specific Strengths
The integration of formal logic proved particularly effective in structured mathematical domains. The framework saw a 14.63% improvement in Geometry and a 7.14% improvement in Precalculus. These areas require precise logical chains where NL-only models frequently make intuitive but mathematically unsound leaps.
Key Insights
- Formal Scaffolding: The research demonstrates that FL acts as a logical scaffold. In a subset of problems that the NL baseline could not solve even with 64 attempts, the NFL-HR framework found the correct solution within its first 16 attempts. This suggests that formal structures provide a unique "search path" through the problem space that natural language alone cannot access.
- The "Existence" Bridge: By framing questions as existence proofs, the model can reason toward an answer formally without the "autoformalization" bottleneck of needing the answer before writing the theorem.
- Efficiency Trade-offs: While the hybrid approach is more computationally intensive—generating approximately 850 additional tokens per query—the significant gain in reasoning capability justifies the increased inference cost.
Limitations and Future Work
Despite its success, the framework has areas for refinement. Currently, the system uses the structure of formal language to guide reasoning but does not yet execute the Lean4 code in real-time. Future iterations aim to include a "compiler-in-the-loop" to provide immediate feedback during the inference process. Additionally, the framework's performance is sensitive to the quality of the initial NL-to-FL translation; poorly formulated existence theorems can lead the reasoner astray. Finally, certain domains like probability remain difficult to formalize effectively, leading to more modest gains compared to algebra or geometry.
Summary updated