Research map / Learning theory and algorithms / Combinatorial and search optimization
Teaching Temporal Logics to Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train a Transformer to predict satisfying traces from LTL formulas | LTLPattern126 test formulas · LTLPattern126 | semantic accuracy: 96.8% | Helped |
| Use the trained Transformer to generate traces for formulas that timed out in spot | LTLUnsolved254 · LTLUnsolved254 | semantic accuracy: 83% | Helped |
| Train on random LTL formulas and predict satisfying traces | LTLRandom35 · LTLRandom35 | semantic accuracy: 98.5% | Helped |
| Use tree positional encoding to generalize to longer formulas | Transformer trained on LTLRandom35 · LTLRandom50 | semantic accuracy: 92.2% | Helped |
| Replace tree positional encoding with standard positional encoding (ablation) | Transformer trained on LTLRandom35 · LTLRandom50 | accuracy | Hurt |
| Train a Transformer to predict satisfying assignments | PropRandom35 · PropRandom35 | semantic accuracy: 96.5% | Helped |
| Use tree positional encoding instead of standard positional encoding (ablation) | Transformer trained on PropRandom35 · PropRandom35–50 | accuracy: 75.7% -> 86.1% | Helped |
| Increase the encoder and decoder stacks from 8 to 10 layers (ablation) | Transformer on LTLRandom35 · LTLRandom35 | semantic accuracy: 98.5% -> 97.6% | Hurt |
| Use 4 heads and FC size 512 rather than 8 heads and FC size 1024 (ablation) | Transformer on LTLRandom35 · LTLRandom35 | semantic accuracy: 98.5% -> 98.3% | Hurt |
| Increase embedding size from 128 to 256 (ablation) | Transformer on LTLRandom35 · LTLRandom35 | semantic accuracy: 98.5% -> 97.9% | Hurt |
| Train on a different formula distribution (random formulas rather than patterns) | Transformer evaluated on LTLPattern126 · LTLPattern126 | semantic accuracy: 24.7% | Hurt |
| Train on a different formula distribution (pattern formulas rather than random formulas) | Transformer evaluated on LTLRandom126 · LTLRandom126 | semantic accuracy: 38.5% | Hurt |
Closest papers by meaning
- The Logical Expressiveness of Temporal GNNs via Two-Dimensional Product Logics (NeurIPS 2025 · similarity 0.88)
- End-to-End Learning of LTLf Formulae by Faithful LTLf Encoding (AAAI 2024 · similarity 0.87)
- Neuro-Symbolic Temporal Point Processes (ICML 2024 · similarity 0.87)
- Neural Model Checking (NeurIPS 2024 · similarity 0.87)
- DeepLTL: Learning to Efficiently Satisfy Complex LTL Specifications for Multi-Task RL (ICLR 2025 · similarity 0.86)
- Transformers Provably Learn Chain-of-Thought Reasoning with Length Generalization (NeurIPS 2025 · 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.