arXiv cs.CL· Mingyu Zhao, Shuhang Lin, Ligong Han, Xiaojing Yu, Zhenting Wang, Yifei Pang, Yang Zhou, Dimitris N. Metaxas·· 13 小时前AI 评分26
Metacognitive Decoding：为扩散语言智能体打造高效推理
Metacognitive Decoding: Efficient Inference for Diffusion Language Agents
AI 导读
研究者提出 Metacognitive Decoding，一种轻量级插件式元认知控制器，引导扩散语言模型按当前回答需求和智能体与环境交互情况决定何时定稿 token、生成延伸多远。在五个基准上，该方法相比最高效基线平均减少 20.6% 推理 FLOPs，同时保持有竞争力的任务准确率。控制器基于智能体轨迹训练，以平衡正确性与生成成本。
正文 · 原文
Abstract:Diffusion language models offer a promising path to efficient agents through parallel token generation. However, efficient interaction requires adapting both where computation is spent and how much text is generated. We introduce a metacognitive controller that guides diffusion generation according to the needs of the current response and the agent's interaction with the environment. It learns when to finalize tokens and how far generation should extend, reducing overall computation within responses. Our controller, a lightweight plug-in for the diffusion backbone, is trained on agent trajectories to balance correctness and generation cost. This approach connects the agent's generation process to the interaction the agent serves. Experiments across five benchmarks demonstrate that our approach reduces inference FLOPs by 20.6% on average compared with the most efficient baseline, while maintaining competitive task accuracy. Ablation studies further validate the effectiveness of our approach.
| Comments: | 10 pages, 4 figures |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.37642 [cs.CL] |
| (or arXiv:2610.37642v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.37642 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Mingyu Zhao [view email]
[v1]
Tue, 27 Oct 2026 17:55:24 UTC (257 KB)
来源：arXiv cs.CL · arxiv.org