AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding

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 changedWhereMeasuredEffect
add AutoDeco temperature and top-p headsLlama-Nemotron-8B · Math averagePass@1: 42.59 -> 46.05Improved
add AutoDeco headsR1-Distill-Qwen-7B · Math averagePass@1: 34.76 -> 37.37Improved
add AutoDeco headsQwen3-30B-A3B-Instruct-2507 · Math averagePass@1: 56.05 -> 56.54Improved
add AutoDeco heads trained only on mathR1-Distill-Qwen-7B · General average (out-of-domain)Pass@1: 42.47 -> 46.88Improved
add AutoDeco headsLlama-Nemotron-8B · GPQA-DiamondPass@1: 51.01 (greedy) vs 50.52Mixed
add AutoDeco headsLlama-Nemotron-8B · BeyondAIMEPass@1: greedy 35.00 vs AutoDeco 34.19Mixed
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 eachImproved
use both heads jointly instead of one head (ablation)AutoDeco · AIME (R1-Distill-Qwen-7B)Pass@1Improved
replace test-set oracle expert-tuned static decoding with AutoDecoAutoDeco · math and general benchmarksPass@1: gap less than one pointNo clear effect
add AutoDeco headsAutoDeco · latency (1k prompt, R1-Distill-Qwen-7B)seconds: 18.23 -> 18.84Hurt
add decoding-control commands with ranking lossAutoDeco · avg predicted temperaturetemperature: 0.72 -> 0.61 (low), 0.72 -> 0.82 (high)Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy