Beyond Gold Standards: Epistemic Ensemble of LLM Judges for Formal Mathematical Reasoning
Authors: Lan Zhang, Marco Valentino, Jordan Meadows, André Freitas
Organizations: Department of Computer Science, University of Manchester, United Kingdom · School of Computer Science, University of Sheffield, United Kingdom · Idiap Research Institute, Switzerland · National Biomarker Centre, CRUK Manchester Institute, United Kingdom
Statement autoformalization plays a crucial role in formal mathematical reasoning by enabling the automatic translation of natural language statements into formal languages. While recent advances using large language models (LLMs) have shown promising capabilities in autoformalization, methods for automatically evaluating autoformalization remain underexplored. LLM-as-a-judge presents a promising approach for automating such evaluation, however, existing methods typically employ coarse-grained and generic evaluation criteria, which limit their effectiveness for advanced formal mathematical reasoning, where quality hinges on nuanced, multi-granular dimensions. In this work, we take a step toward addressing this gap by introducing a systematic, automatic method to evaluate autoformalization tasks. The proposed method is based on an epistemically and formally grounded ensemble (EFG) of LLM judges, defined on criteria encompassing logical preservation (LP), mathematical consistency (MC), formal quality (FQ), and formal validity (FV), resulting in a transparent assessment that accounts for different contributing factors. We validate the proposed framework to serve as a proxy for autoformalization assessment within the domain of formal mathematics. Overall, our experiments demonstrate that the EFG ensemble of LLM judges is a more suitable emerging proxy for evaluation than a coarse-grained model. These findings suggest that LLM judges, especially when guided by a well-defined set of atomic properties, could offer a scalable, interpretable, and reliable support for evaluating formal mathematical reasoning.
Figures & tables
Figure 1: Overview of the proposed framework for automatically evaluating autoformalization. Our method leverages ensembles of LLM judges based on epistemic criteria organized into a taxonomy, while also providing explanations for their judgments. A linear evaluation model aggregates the criteria into a single overall assessment to serve as a proxy for human evaluation.
Case 1: miniF2F Isabelle Unpreserved Logic Find the minimum value of xsinx9x2sin2x+4 for 0<x<π . Show that it is 012. ⬇ theorem aime_1983_p9 : fixes x :: real assumes "0< x " " x < pi " shows "12 < le > ((9 * ( x ^2 * ( sin x )^2)) + 4) / ( x * sin x )" sorry
Case 2: Proofnet Lean4 Inconsistent Mathematics Let A be a nonempty set of real numbers which is bounded below. Let −A be the set of all numbers −x , where x∈A . Prove that infA=−sup(−A) . ⬇ theorem exercise_1_5 ( A minus_A : Set R ) ( hA : A . Nonempty ) ( hA_bdd_below : BddBelow A ) ( hminus_A : minus_A = { x | - x ∈ A }) : Inf A = Sup minus_A := sorry
Case 3: miniF2F Lean4 Redundant Hypothesis Simplify (2−2i)(5+5i) , where i2=−1. Show that it is 20. ⬇ theorem mathd_algebra_110 ( q e : C ) ( h0 : q = 2 - 2 * Complex . I ) ( h1 : e = 5 + 5 * Complex . I ) : q * e = 20 := sorry
Table 1: Typical challenges involved in formalization evaluation.
miniF2F
ProofNet
LLM
Setup
EM-H1
EM-H2
PM-H1
PM-H2
EM-H1
EM-H2
PM-H1
PM-H2
DeepSeek-Math-7B
OA-TP
18.00
16.00
39.33
39.33
6.00
16.00
51.33
46.67
DeepSeek-Math-7B
DA-T2
18.00
24.00
44.67
45.33
18.00
10.00
46.00
39.33
DeepSeek-Math-7B
OAP
20.00
16.00
47.33
49.33
20.00
8.00
60.00
48.00
Qwen2.5-Coder-7B
OA-TP
10.00
22.00
40.00
46.67
10.00
10.00
38.67
39.33
Qwen2.5-Coder-7B
DA-T2
22.00
22.00
50.67
52.00
12.00
20.00
48.00
56.00
Table 2: Results of using LLM judges to evaluate Lean4 formalizations from three sources, Human, GPT-4.1 and Qwen2.5-7B. The metrics include: ( EM ): the percentage of rankings that exactly match human rankings; ( PM ): the percentage of pairwise rankings that match human pairwise rankings. H1 and H2 denote two human annotators. The evaluation setups are: ( OA-TP ): overall assessment by an LLM judge with theorem prover feedback injected into the prompt; ( DA-T2 ): directly assessing the Tier-2 aspects LP, MC, and FQ using LLM judges, and FV using the theorem prover, then averaging the results; ( OAP ): independently assessing the OAPs using LLM judges and synthesizing the Tier-2 aspects LP, MC, and FQ using a weighted average, where each OAP is weighted equally within each Tier-2 aspect, and FV using the theorem prover, then averaging the results. Bold represents the setup with the highest score within the same LLM and Underline represents the setup with the highest score overall.
Isabelle (miniF2F)
Lean4 (miniF2F)
Lean4 (ProofNet)
Aspect
Judge
GT
ZS
GT
ZS
GT
ZS
LP
Human Expert
0.24±0.43
0.80±0.40
0.40±0.49
0.80±0.40
0.68±0.47
0.40±0.49
MC
Human Expert
0.34±0.47
0.74±0.44
0.40±0.49
0.52±0.50
0.68±0.48
0.40±0.49
FQ
Human Expert
0.76±0.43
0.46±0.50
0.92±0.27
0.56±0.50
0.92±0.27
0.60±0.49
FV
Theorem Prover
1.00±0.00
0.52±0.50
1.00±0.00
0.20±0.40
1.00±0.00
0.12±0.32
OA
Human Expert
4.06±0.58
4.00±0.77
4.04±0.53
3.64±0.93
4.48±0.64
3.56±0.57
Table 3: Human evaluation of formalizations on subsets of validation sets. Results are reported in mean ± standard deviation across samples. ( GT ): Ground-Truth formalization; ( ZS ): GPT-4.1 Zero-Shot formalization; ( LP ): Logical Preservation; ( MC ): Mathematical Consistency; ( FQ ): Formal Quality; ( FV ): Formal Validity; ( OA ): Overall Assessment.
Figure 2: Pearson correlation coefficients between evaluation aspects and overall assessments across judges for flawed formalizations. LLM judges and human evaluators assess the Tier-2 aspects logical preservation (LP), mathematical consistency (MC), and formal quality (FQ), while theorem provers assess formal validity (FV). The lower triangular matrix shows results for Lean4, and the upper triangular matrix shows results for Isabelle/HOL.
Setting
Model
LP
MC
FV
FQ
OA
miniF2F Isabelle/HOL
0-Shot
GPT-4.1
0.939
0.963
0.643
0.953
0.874
8-Shot
GPT-4.1
0.945
0.986
0.746
0.949
0.906
8-Shot
Mistral-7B
0.418
0.624
0.201
0.480
0.431
8-Shot
DS-Math-7B
0.492
0.698
0.324
0.512
0.506
8-Shot
Qwen2.5-7B
0.651
0.831
0.590
0.740
0.703
Table 4: Evaluation of autoformalizations on the test sets with GPT-4.1-mini in OAP setting. ( DS ): Deepseek; ( LP ): Logical Preservation; ( MC ): Mathematical Consistency; ( FQ ): Formal Quality; ( FV ): Formal Validity; ( OA ): Overall Assessment.
AF
Judge
LP
MC
FQ
OA
Lean4
G-4.1
G-4.1-Mini
0.924
0.985
0.957
0.891
Q-7B
G-4.1-Mini
0.627
0.842
0.758
0.703
G-4.1
Q-Coder-7B
0.703
0.757
0.328
0.621
Q-7B
Q-Coder-7B
0.473
0.616
0.170
0.461
Isabelle/HOL
Table 5: Using different judges for evaluation on miniF2F test set. LLM autoformalizer (AF) is used in few-shot setting. ( G ): GPT; ( Q ): Qwen2.5.
Figure 3: Error bars of GPT-4.1-mini with OAP setting running with different temperatures.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Logical Preservation (LP) From your point of view, does the formal code capture the logical structure and content of the natural language statement? 1 : Yes 0 : No
Mathematical Consistency (MC) From your point of view, does the formal code accurately represent mathematical objects and operations present in the natural language statement? 1 : Yes 0 : No
Formal Quality (FQ) From your point of view, does the formalized code alone demonstrate high quality in terms of structural clarity, readability, and usefulness? 1 : Yes 0 : No
Overall Assessment (OA) What would you rate this formal code as the formalization of the natural language statement? 5 : Perfect formalization – fully accurate and complete with no issues. 4 : Good formalization – minor flaws are present, but they are easy to fix. 3 : Acceptable formalization – captures the intent but may require some time and effort to correct. 2 : Flawed formalization – contains major issues that require significant time and effort to correct. 1 : Poor formalization – completely fails to reflect the intended meaning of the natural language statement.
Appendix
Table 6: Annotation criteria
Table 7: Example 1 of incorrect LLM in Isabelle/HOL.