Research map / Large language models / Language, safety and interpretability
Neural Rule-Execution Tracking Machine For Transformer-Based Text Generation
Yufei Wang, Can Xu, Huang Hu, Chongyang Tao, Stephen Wan, Mark Dras, Mark Johnson, Daxin Jiang
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 dynamic rule-execution tracking to T5-Base | T5-Base on ROCStories with sentence-index and ordering constraints · ROCStories Test Split | CSR: 98.7 -> 98.8; 97.0 -> 97.4 | Improved |
| Add dynamic rule-execution tracking to T5-Base | T5-Base on ROCStories with sentence-length constraints · ROCStories Test Split | CSR: 18.5 -> 84.5 | Improved |
| Add dynamic rule-execution tracking to T5-Base | T5-Base on ROCStories with two length-or-stop-word constraints · ROCStories Test Split | CSR: 11.0 -> 35.9 | Improved |
| Add dynamic rule-execution tracking to T5-Base | T5-Base on ROCStories · ROCStories Test Split, simple sentence-index control | R-L: 33.1 -> 33.3; 49.6 -> 49.5 | No clear effect |
| Use morphological inflection alternatives rather than exact-form copying | T5-Large + NRETM on CommonGen · CommonGen test | CIDEr: 167.3 -> 175.0 | Improved |
| Add exact-copy constraints with NRETM | T5-Large on CommonGen · CommonGen test | CIDEr: 163.6 -> 167.3 | Improved |
| Use inflectional alternatives in copy constraints | T5-Large + NRETM on CommonGen · CommonGen test | Constraint satisfaction ALL: 93.9 -> 99.4 | Improved |
| Apply Grid Beam Search to enforce lexical constraints at inference | T5-Base on CommonGen · CommonGen test | CIDEr: 159.5 -> 110.7 | Hurt |
| Add a translated-once sentence alignment constraint | Doc-mBART25 on TED15 Zh-En · TED15 Zh-En test | SAR: 98.7 % -> 100 % | Improved |
| Add a translated-once sentence alignment constraint | Doc-mBART25 on TED15 Zh-En · TED15 Zh-En test | s-BLEU: + 1.1 BLEU | Improved |
| Train on storyline positions 3, 4, and 5, then test positions 1 and 2 | T5 versus NRETM on ROCStories · ROCStories novel sentence-index test | CSA: 19.7 -> 97.7 | Improved |
| Add NRETM to T5-Base | T5-Base on CommonGen · CommonGen test | CIDEr: 159.5 -> 169.8 | Improved |
Closest papers by meaning
- Controllable Text Generation with Neurally-Decomposed Oracle (NeurIPS 2022 · similarity 0.86)
- Copy That! Editing Sequences by Copying Spans (AAAI 2021 · similarity 0.84)
- Controlled Text Generation via Language Model Arithmetic (ICLR 2024 · similarity 0.83)
- Successor Features for Efficient Multi-Subject Controlled Text Generation (ICML 2024 · similarity 0.83)
- LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs (ICLR 2025 · similarity 0.82)
- LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (ICLR 2025 · similarity 0.82)
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.
