Beyond Theorem Proving: Formulation, Framework and Benchmark for Formal Problem-Solving

Qi Liu, Xinhao Zheng, Renqiu Xia, Xingzhi Qi, Qinxiang Cao, Junchi Yan
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:80113-80145, 2026.

Abstract

Large language models (LLMs) have achieved remarkable progress in mathematical reasoning, yet persistently suffer from hallucinations and erroneous logic. While formal theorem proving (FTP) shows promise in process-level reliability, it is limited to verification (checking known propositions). This leaves constructive problem-solving (finding unknown terms that satisfy specific conditions) underexplored and disconnected from process-level verifiability. To bridge this gap, we introduce FPS (Formal Problem-Solving), a principled framework to encompass the end-to-end problem-solving process in Lean 4. In FPS, the answer is an unknown metavariable coupled with a proof obligation, forcing it to be mathematically derived and verified. We further present D-FPS (Deductive FPS), which structures solving into forward derivation and backward verification, aligning more closely with human reasoning steps. Three benchmarks of over 1,000 problems are constructed for evaluation: FormalMath500, MiniF2F-Solving, and PutnamBench-Solving. We further propose RPE (Restricted Propositional Equivalence), a symbolic metric that evaluates semantic correctness beyond brittle string matching. Extensive experiments with state-of-the-art provers reveal that solving is significantly harder than proving, highlighting the “alignment tax” required to transition from loose validity checking to constructive, human-aligned reasoning. Code and data are available at https://github.com/Purewhite2019/formal_problem_solving_main.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-liu26gl, title = {Beyond Theorem Proving: Formulation, Framework and Benchmark for Formal Problem-Solving}, author = {Liu, Qi and Zheng, Xinhao and Xia, Renqiu and Qi, Xingzhi and Cao, Qinxiang and Yan, Junchi}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {80113--80145}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/liu26gl/liu26gl.pdf}, url = {https://proceedings.mlr.press/v306/liu26gl.html}, abstract = {Large language models (LLMs) have achieved remarkable progress in mathematical reasoning, yet persistently suffer from hallucinations and erroneous logic. While formal theorem proving (FTP) shows promise in process-level reliability, it is limited to verification (checking known propositions). This leaves constructive problem-solving (finding unknown terms that satisfy specific conditions) underexplored and disconnected from process-level verifiability. To bridge this gap, we introduce FPS (Formal Problem-Solving), a principled framework to encompass the end-to-end problem-solving process in Lean 4. In FPS, the answer is an unknown metavariable coupled with a proof obligation, forcing it to be mathematically derived and verified. We further present D-FPS (Deductive FPS), which structures solving into forward derivation and backward verification, aligning more closely with human reasoning steps. Three benchmarks of over 1,000 problems are constructed for evaluation: FormalMath500, MiniF2F-Solving, and PutnamBench-Solving. We further propose RPE (Restricted Propositional Equivalence), a symbolic metric that evaluates semantic correctness beyond brittle string matching. Extensive experiments with state-of-the-art provers reveal that solving is significantly harder than proving, highlighting the “alignment tax” required to transition from loose validity checking to constructive, human-aligned reasoning. Code and data are available at https://github.com/Purewhite2019/formal_problem_solving_main.} }
Endnote
%0 Conference Paper %T Beyond Theorem Proving: Formulation, Framework and Benchmark for Formal Problem-Solving %A Qi Liu %A Xinhao Zheng %A Renqiu Xia %A Xingzhi Qi %A Qinxiang Cao %A Junchi Yan %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-liu26gl %I PMLR %P 80113--80145 %U https://proceedings.mlr.press/v306/liu26gl.html %V 306 %X Large language models (LLMs) have achieved remarkable progress in mathematical reasoning, yet persistently suffer from hallucinations and erroneous logic. While formal theorem proving (FTP) shows promise in process-level reliability, it is limited to verification (checking known propositions). This leaves constructive problem-solving (finding unknown terms that satisfy specific conditions) underexplored and disconnected from process-level verifiability. To bridge this gap, we introduce FPS (Formal Problem-Solving), a principled framework to encompass the end-to-end problem-solving process in Lean 4. In FPS, the answer is an unknown metavariable coupled with a proof obligation, forcing it to be mathematically derived and verified. We further present D-FPS (Deductive FPS), which structures solving into forward derivation and backward verification, aligning more closely with human reasoning steps. Three benchmarks of over 1,000 problems are constructed for evaluation: FormalMath500, MiniF2F-Solving, and PutnamBench-Solving. We further propose RPE (Restricted Propositional Equivalence), a symbolic metric that evaluates semantic correctness beyond brittle string matching. Extensive experiments with state-of-the-art provers reveal that solving is significantly harder than proving, highlighting the “alignment tax” required to transition from loose validity checking to constructive, human-aligned reasoning. Code and data are available at https://github.com/Purewhite2019/formal_problem_solving_main.
APA
Liu, Q., Zheng, X., Xia, R., Qi, X., Cao, Q. & Yan, J.. (2026). Beyond Theorem Proving: Formulation, Framework and Benchmark for Formal Problem-Solving. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:80113-80145 Available from https://proceedings.mlr.press/v306/liu26gl.html.

Related Material