The Curious Case of the Unreliable Decipherer: What LLMs struggle with while explaining Formal Proofs

ACL ARR 2025 July Submission26 Authors

Abstract: Formal methods—such as symbolic proofs—offer a principled way to force Large Language Models (LLMs) to reason more reliably. However it is unclear whether LLMs can actually use these proofs to generate faithful, and yet human understandable explanations? We introduce ProofTeller, a new benchmark to evaluate this capability. On a new dataset with over 68,000 human-annotated tokens, we evaluate several LLMs on three tasks: identifying key proof steps, summarizing the proofs, and creating a user-targeted message. Across tasks and settings, both automated metrics and human evaluation reveal a critical reliability gap: LLMs over‑emphasize steps near the final conclusion, while humans draw on evidence distributed throughout the proof. These findings expose a fundamental mismatch between the reasoning strategies of current LLMs and those of humans, underscoring the need for approaches that enable LLMs to employ formal proof methods reliably while faithfully generating a comprehensible reasoning chain.

Paper Type: Long

Research Area: Resources and Evaluation

Research Area Keywords: benchmark, LLM evaluation, reliability, faithfulness

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English

Reassignment Request Area Chair: This is not a resubmission

Reassignment Request Reviewers: This is not a resubmission

A1 Limitations Section: This paper has a limitations section.

A2 Potential Risks: N/A

B Use Or Create Scientific Artifacts: Yes

B1 Cite Creators Of Artifacts: Yes

B1 Elaboration: 3

B2 Discuss The License For Artifacts: Yes

B2 Elaboration: 3

B3 Artifact Use Consistent With Intended Use: Yes

B3 Elaboration: 4

B4 Data Contains Personally Identifying Info Or Offensive Content: No

B4 Elaboration: There is no personally identifying info or offensive content in any of the data used or created

B5 Documentation Of Artifacts: Yes

B5 Elaboration: 3

B6 Statistics For Data: Yes

B6 Elaboration: 3

C Computational Experiments: Yes

C1 Model Size And Budget: Yes

C1 Elaboration: 4

C2 Experimental Setup And Hyperparameters: Yes

C2 Elaboration: 4

C3 Descriptive Statistics: Yes

C3 Elaboration: 5

C4 Parameters For Packages: Yes

C4 Elaboration: 4

D Human Subjects Including Annotators: Yes

D1 Instructions Given To Participants: Yes

D1 Elaboration: 3/6/G

D2 Recruitment And Payment: Yes

D2 Elaboration: 3/6

D3 Data Consent: Yes

D3 Elaboration: G

D4 Ethics Review Board Approval: Yes

D4 Elaboration: Ethics statement

D5 Characteristics Of Annotators: Yes

D5 Elaboration: 3/6

E Ai Assistants In Research Or Writing: No

E1 Information About Use Of Ai Assistants: N/A

E1 Elaboration: N/A - We didn't use any.

Author Submission Checklist: yes

Submission Number: 26