Research map / Large language models / Language, safety and interpretability
COLD Decoding: Energy-based Constrained Text Generation with Langevin Dynamics
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 |
|---|---|---|---|
| add n-gram similarity constraint (ablation) | COLD · alphaNLG human eval | overall coherence: 2.69 -> 2.83 | Helped |
| add reverse LM fluency constraint (ablation) | COLD · alphaNLG human eval | overall coherence: 2.80 -> 2.83 | Helped |
| add future-token prediction constraint (ablation) | COLD · alphaNLG human eval | overall coherence: 2.77 -> 2.83 | Helped |
| add n-gram similarity constraint (ablation) | COLD · alphaNLG human eval | grammar: 4.54 -> 4.17 | Hurt |
| add future-token prediction constraint (ablation) | COLD · alphaNLG human eval | grammar: 4.61 -> 4.17 | Hurt |
| COLD instead of DeLorean decoding | GPT2-XL · alphaNLG abductive generation | CIDEr: 7.88 -> 10.68 | Helped |
| COLD instead of DeLorean decoding | GPT2-XL · Counterfactual story rewriting | Edit similarity: 52.90 -> 56.84 | Helped |
| COLD instead of DeLorean decoding | GPT2-XL · Counterfactual story rewriting | BERTScore: 73.66 -> 73.47 | No clear effect |
| COLD instead of NeuroLogic decoding | GPT2-XL · Lexically constrained generation | keyword coverage percent: 91.00 -> 94.50 | Helped |
| COLD instead of NeuroLogic decoding | GPT2-XL · Lexically constrained generation | perplexity: 28.61 -> 54.98 | Hurt |
Closest papers by meaning
- Controlled LLM Decoding via Discrete Auto-regressive Biasing (ICLR 2025 · similarity 0.89)
- Improving Open-Ended Text Generation via Adaptive Decoding (ICML 2024 · similarity 0.87)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.87)
- Decoding Game: On Minimax Optimality of Heuristic Text Generation Strategies (ICLR 2025 · similarity 0.87)
- Top-H Decoding: Adapting the Creativity and Coherence with Bounded Entropy in Text Generation (NeurIPS 2025 · similarity 0.87)
- Language Model Decoding as Direct Metrics Optimization (ICLR 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.