Research map / Large language models / Language, safety and interpretability
Do LLMs Really Struggle at NL-FOL Translation? Revealing Their Strengths via a Novel Benchmarking Strategy
Andrea Brunello, Luca Geatti, Michele Mignani, Angelo Montanari, Nicola Saccomanno
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add task-specific instruction to the embedding input (ablation) | Qwen3-Embedding-8B · D_FOLIO FOL most similar | accuracy: .44 -> .76 | Improved |
| add task-specific instruction to the embedding input (ablation) | Qwen3-Embedding-8B · D_FOLIO FOL ranking-both | score: .13 -> .37 | Improved |
| use o3-mini instead of GPT-4o-mini | logical translation · D_Stanford | equivalence accuracy: .84 -> .94 | Improved |
| use Qwen3-30B instead of Qwen3-8B | logical translation · D_FOLIO | equivalence accuracy: .72 -> .74 | No clear effect |
| enable thinking mode (ablation) | Qwen3-30B · Ranking-Equivalence and Ranking-Both | accuracy: up to a 6 point increment in accuracy | Improved |
| enable thinking mode (ablation) | Qwen3-8B · Ranking-Equivalence and Ranking-Both | accuracy: performance differences are marginal | No clear effect |
| enable thinking mode (ablation) | Qwen3-8B and Qwen3-30B · Logical Translation | accuracy: identical average accuracy | No clear effect |
| use FOL formulas instead of NL renderings as candidates | GPT-4o-mini · D_FOLIO ranking-equivalence | score: .52 -> .64 | Improved |
| use FOL formulas instead of NL renderings as candidates | Qwen-Emb-plain · D_Stanford most similar | accuracy: .62 -> .46 | Hurt |
| use ranking-equivalence task instead of most similar | GPT-4o-mini · D_FOLIO FOL | success rate: fails ranking-equivalence in 31% while opposite occurs in 4% | Hurt |
| evaluate on D_Stanford instead of D_FOLIO | dialogue models · logical translation | accuracy: .11--.14 difference | Improved |
| use dialogue-oriented o3-mini instead of Qwen-Emb-plain | ranking task · D_Stanford FOL ranking-both | score: .89 vs .09 | Improved |
Closest papers by meaning
- Grammar-Forced Translation of Natural Language to Temporal Logic using LLMs (ICML 2025 · similarity 0.88)
- Autonomous Evaluation of LLMs for Truth Maintenance and Reasoning Tasks (ICLR 2025 · similarity 0.87)
- MindMerger: Efficiently Boosting LLM Reasoning in non-English Languages (NeurIPS 2024 · similarity 0.87)
- It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.87)
- Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? (ICLR 2025 · similarity 0.87)
- LLM Augmented LLMs: Expanding Capabilities through Composition (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
