Research map / Learning theory and algorithms / Combinatorial and search optimization
A Non-monotonic Self-terminating Language Model
Eugene Choi, Kyunghyun Cho, Cheolhyoung Lee
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 |
|---|---|---|---|
| Replace vanilla softmax with NMST parametrization | RNN and LSTM · WikiText-2, greedy search | non-termination ratio | Improved |
| Replace ST parametrization with NMST parametrization | RNN · WikiText-2 validation | perplexity: 180.4\pm(3.3) -> 177.4\pm(4.5) | Improved |
| Replace ST parametrization with NMST parametrization | LSTM · WikiText-2 validation | perplexity: 104.5\pm(1.4) -> 101.5\pm(0.8) | Improved |
| Use NMST parametrization with ε=1.0×10^-5 instead of vanilla softmax | RNN · WikiText-2 validation | perplexity: 178.6\pm(6.3) -> 177.4\pm(4.5) | No clear effect |
| Use NMST parametrization with ε=1.0×10^-5 instead of vanilla softmax | LSTM · WikiText-2 validation | perplexity: 101.6\pm(1.0) -> 101.5\pm(0.8) | No clear effect |
| Replace ST parametrization with NMST parametrization | GPT-2 · WikiText-103 validation | perplexity: 21.16\pm(0.03) -> 20.69\pm(0.03) | Improved |
| Use NMST parametrization with ε=1.0×10^-5 instead of vanilla softmax | GPT-2 · WikiText-103 validation | perplexity: 20.72\pm(0.03) -> 20.69\pm(0.03) | No clear effect |
| Replace vanilla softmax with NMST parametrization | GPT-2 · WikiText-103, greedy search, L=1,000 | non-termination ratio: 0.27\pm(0.08) -> 0.23\pm(0.10) | Improved |
| Replace ST parametrization with NMST parametrization | GPT-2 · WikiText-103, greedy search, L=1,000 | non-termination ratio: 0.75\pm(0.10) -> 0.23\pm(0.10) | Improved |
| Replace vanilla softmax with NMST parametrization | GPT-2 · WikiText-103, top-2 sampling, L=1,000 | non-termination ratio: 0.0\pm(0.0) -> 0.0\pm(0.0) | No clear effect |
| Replace vanilla softmax with NMST parametrization | GPT-2 · WikiText-103, nucleus-0.2 sampling, L=1,000 | non-termination ratio: 0.25\pm(0.08) -> 0.21\pm(0.10) | Improved |
| Replace vanilla softmax with NMST parametrization | GPT-2 · WikiText-103, nucleus-0.4 sampling, L=1,000 | non-termination ratio: 0.14\pm(0.05) -> 0.10\pm(0.06) | Improved |
Closest papers by meaning
- Closing the Curious Case of Neural Text Degeneration (ICLR 2024 · similarity 0.86)
- When Linear Attention Meets Autoregressive Decoding: Towards More Effective and Efficient Linearized Large Language Models (ICML 2024 · similarity 0.85)
- Top-H Decoding: Adapting the Creativity and Coherence with Bounded Entropy in Text Generation (NeurIPS 2025 · similarity 0.85)
- Multi-Branch Self-Drafting for LLM Inference Acceleration (AAAI 2025 · similarity 0.85)
- Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding (ICML 2025 · similarity 0.85)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.85)
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.
