arXiv：cs.LG· Jingyuan Fan, Purui Liu, Hengbo Xiao, Yuxuan Zheng, Jingzhao Zhang, Chao Lu, Guannan He·· 7 小时前AI 评分39
PiERN：用 Token 级路由整合高精度计算与推理
PiERN: Token-Level Routing for Integrating High-Precision Computation and Reasoning
AI 导读
研究者提出 Physically-isolated Experts Routing Network（PiERN），一种在 token 级调度计算与推理的架构，可在单条思维链内实现迭代交替。在 PDEBench 和电池管理任务上，PiERN 精度高于直接微调 LLM，并在响应延迟、token 用量、GPU 能耗和专家路由准确率上优于主流多智能体方案，同时在 MMLU 和 GLUE 上无显著性能下降。
正文
Abstract:Tasks on complex systems require high-precision numerical computation to support decisions. However, current large language models (LLMs), even with enhanced reasoning capabilities, cannot integrate such computations as an intrinsic and interpretable capability with existing architectures. To this end, we propose Physically-isolated Experts Routing Network (PiERN), an architecture that directs computation and reasoning at token level, thereby enabling iterative alternation within a single chain of thought. We systematically evaluate PiERN on representative computation-reasoning tasks, including PDEBench and battery management tasks. Results show that PiERN achieves not only higher accuracy than directly finetuning LLMs but also significant improvements in response latency, token usage, GPU energy consumption, and experts routing accuracy compared with mainstream multi-agent approaches, while exhibiting no significant degradation in performance on MMLU and GLUE benchmarks. PiERN offers an efficient, interpretable, and scalable paradigm for interfacing language models with scientific systems.
| Subjects: | Machine Learning (cs.LG); Computational Engineering, Finance, and Science (cs.CE); Computation and Language (cs.CL) |
| Cite as: | arXiv:2509.18169 [cs.LG] |
| (or arXiv:2509.18169v4 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2509.18169 arXiv-issued DOI via DataCite |
Submission history
From: Guannan He [view email]
[v1]
Wed, 17 Sep 2025 10:15:25 UTC (5,495 KB)
[v2]
Sat, 27 Sep 2025 06:44:30 UTC (7,803 KB)
[v3]
Mon, 20 Apr 2026 12:46:48 UTC (4,817 KB)
[v4]
Tue, 6 Oct 2026 09:14:08 UTC (3,492 KB)
来源：arXiv：cs.LG · arxiv.org