Research map / Large language models / Language, safety and interpretability
Diverse Text Decoding via Iterative Reweighting
Ruiqi Shi, Sinno Jialin Pan
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 OverRIDE iterative reweighting to greedy decoding | Qwen-2.5-7B-Instruct · HumanEval | PASS@10: 64.9 -> 86.0 | Improved |
| add OverRIDE to greedy decoding | Mistral-7B-Instruct-v0.3 · GSM8K | PASS@10: 39.5 -> 74.2 | Improved |
| add OverRIDE to top-p sampling (tau=1.0) | Qwen-2.5-7B-Instruct · MATH500 | PASS@10: 87.2 -> 87.6 | Improved |
| add OverRIDE to greedy decoding | Qwen-2.5-7B-Instruct · HumanEval | CodeBLEU (pairwise): 1.000 -> 0.626 | Improved |
| add OverRIDE to greedy decoding | Qwen-2.5-7B-Instruct · CCNews | MAUVE: 0.803 -> 0.753 | Hurt |
| add OverRIDE to top-p sampling (tau=0.6) | Qwen-2.5-7B-Instruct · CCNews | MAUVE: 0.901 -> 0.938 | Improved |
| add OverRIDE to larger model (14B vs 3B) | Qwen-2.5 · HumanEval PASS@10 (top-p tau=0.6) | PASS@10: 3B: 82.9 -> 83.2; 14B: 71.1 -> 77.2 | Improved |
| use low-rank output head adapter rank 16 instead of full output heads (ablation) | OverRIDE on Qwen-2.5-7B · MATH500 | throughput: 1984.1 -> 3435.8 | Improved |
| raise adapter rank from 16 to 256 (ablation) | OverRIDE on Qwen-2.5-7B · MATH500 | PASS@10: 87.5 -> 87.1 | No clear effect |
| use rank-4 adapter instead of rank 16 (ablation) | OverRIDE on Qwen-2.5-7B · MATH500 | PASS@10: 86.9 -> 87.5 | Improved |
| OverRIDE instead of Contrastive Decoding | GPT2-XL · CCNews | MAUVE: 0.752 -> 0.846 | Improved |
| OverRIDE instead of SimCTG | Qwen-2.5-7B-Instruct · HumanEval | PASS@10: 77.2% -> 86.4% | Improved |
Closest papers by meaning
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.88)
- Multi-Branch Self-Drafting for LLM Inference Acceleration (AAAI 2025 · similarity 0.88)
- The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation (ICLR 2025 · similarity 0.88)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.88)
- Adaptive Draft-Verification for Efficient Large Language Model Decoding (AAAI 2025 · similarity 0.88)
- Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment (ICLR 2025 · similarity 0.87)
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.
