Research map / Large language models / Agents, code and math
VERIFY: A Novel Multi-Domain Dataset Grounding LTL in Contextual Natural Language via Provable Intermediate Logic
Paapa Kwesi Quansah, Pablo Rivas, Ernest Bonnah
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 |
|---|---|---|---|
| use ITL instead of LTL as the source for NL generation | Llama-3-8B-Instruct (FT) · VERIFY | BERTScore F1: 0.91 -> 0.94 | Improved |
| use ITL instead of LTL as the source for NL generation | Llama-3-8B-Instruct (FT) · VERIFY | ROUGE-L F1: 0.67 -> 0.73 | Improved |
| use ITL instead of LTL as the source for NL generation | T5-large · VERIFY | BERTScore F1: 0.67 -> 0.89 | Improved |
| add domain and activity context to NL-to-LTL training (ablation) | Llama-3-8B-Instruct (FT) · VERIFY NL-to-LTL | semantic equivalence (%): 7.7 -> 28.2 | Improved |
| add domain and activity context to NL-to-LTL training (ablation) | Llama-3-8B-Instruct (FT) · VERIFY NL-to-LTL | exact match (%): 0.8 -> 4.1 | Improved |
| add domain and activity context to NL-to-ITL training (ablation) | Llama-3-8B-Instruct (FT) · VERIFY NL-to-ITL | semantic equivalence (%): 13.9 -> 41.5 | Improved |
| increase LTL formula AST depth from 1-4 to 13+ | Llama-3-8B-Instruct (FT) · VERIFY NL-to-LTL | semantic equivalence (%): 35.4 -> 15.2 | Hurt |
| hold out a domain from training (Aerospace) (ablation) | Llama-3-8B-Instruct (FT) · VERIFY NL-to-LTL | semantic equivalence (%): 27.5 -> 19.2 | Hurt |
| hold out a domain from training (Home Automation) (ablation) | Llama-3-8B-Instruct (FT) · VERIFY NL-to-LTL | semantic equivalence (%): 30.1 -> 22.5 | Hurt |
| hold out a domain from training (Financial) (ablation) | Llama-3-8B-Instruct (FT) · VERIFY NL-to-LTL | semantic equivalence (%): 25.8 -> 16.7 | Hurt |
| use code-focused DeepSeek-Coder instead of Llama-3-8B | fine-tuned LLM · VERIFY NL-to-LTL | semantic equivalence (%): 28.2 -> 31.5 | Improved |
| use fine-tuned LLM instead of T5-large | T5-large · VERIFY NL-to-LTL | semantic equivalence (%): 22.3 -> 28.2 | Improved |
Closest papers by meaning
- RvLLM: LLM Runtime Verification with Domain Knowledge (NeurIPS 2025 · similarity 0.89)
- Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks (NeurIPS 2025 · similarity 0.87)
- Grammar-Forced Translation of Natural Language to Temporal Logic using LLMs (ICML 2025 · similarity 0.87)
- ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation (AAAI 2026 · similarity 0.86)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.86)
- Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus (NeurIPS 2024 · similarity 0.86)
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.
