Research map / Large language models / Language, safety and interpretability
THE END OF MANUAL DECODING: TOWARDS TRULY END-TO-END LANGUAGE MODELS
Zhichao Wang, Dongyang Ma, Xinting Huang, Deng Cai, Tian Lan, Jiahao Xu, Haitao Mi, Xiaoying Tang, Yan Wang
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 AutoDeco temperature and top-p heads | Llama-Nemotron-8B · Math average | Pass@1: 42.59 -> 46.05 | Improved |
| add AutoDeco heads | R1-Distill-Qwen-7B · Math average | Pass@1: 34.76 -> 37.37 | Improved |
| add AutoDeco heads | Qwen3-30B-A3B-Instruct-2507 · Math average | Pass@1: 56.05 -> 56.54 | Improved |
| add AutoDeco heads trained only on math | R1-Distill-Qwen-7B · General average (out-of-domain) | Pass@1: 42.47 -> 46.88 | Improved |
| add AutoDeco heads | Llama-Nemotron-8B · GPQA-Diamond | Pass@1: 51.01 (greedy) vs 50.52 | Mixed |
| add AutoDeco heads | Llama-Nemotron-8B · BeyondAIME | Pass@1: greedy 35.00 vs AutoDeco 34.19 | Mixed |
| use only temperature head or only top-p head (ablation) | AutoDeco · AIME (R1-Distill-Qwen-7B) | Pass@1: about +3-3.5 points over default each | Improved |
| use both heads jointly instead of one head (ablation) | AutoDeco · AIME (R1-Distill-Qwen-7B) | Pass@1 | Improved |
| replace test-set oracle expert-tuned static decoding with AutoDeco | AutoDeco · math and general benchmarks | Pass@1: gap less than one point | No clear effect |
| add AutoDeco heads | AutoDeco · latency (1k prompt, R1-Distill-Qwen-7B) | seconds: 18.23 -> 18.84 | Hurt |
| add decoding-control commands with ranking loss | AutoDeco · avg predicted temperature | temperature: 0.72 -> 0.61 (low), 0.72 -> 0.82 (high) | Improved |
Closest papers by meaning
- AdaDecode: Accelerating LLM Decoding with Adaptive Layer Parallelism (ICML 2025 · similarity 0.89)
- Multi-Branch Self-Drafting for LLM Inference Acceleration (AAAI 2025 · similarity 0.89)
- Instructive Decoding: Instruction-Tuned Large Language Models are Self-Refiner from Noisy Instructions (ICLR 2024 · similarity 0.89)
- Reference Trustable Decoding: A Training-Free Augmentation Paradigm for Large Language Models (NeurIPS 2024 · similarity 0.88)
- AutoJudge: Judge Decoding Without Manual Annotation (NeurIPS 2025 · similarity 0.88)
- Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads (ICML 2024 · similarity 0.88)
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.
