arXiv cs.CL· Zhiyuan Wang·· 12 小时前AI 评分31
研究揭示 MoE 模型预填充阶段减少专家对短答案 F1 的损害机制
Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss
AI 导读
在四种 MoE 模型家族上，预填充阶段将路由专家减半导致的短答案 F1 损失主要体现为答案形式受损——模型仍能找到答案但不再简短作答。Qwen3-30B-A3B 的损失集中在答案前的两个 chat-template token，DeepSeek-V2-Lite 则集中在开头 8-32 个 prompt token 的指令位置。
正文 · 原文
Abstract:Running prompt tokens through fewer routed experts during prefill is a simple way to make mixture-of-experts (MoE) serving cheaper, yet it is unclear what this damages. We locate the damage with token-, layer- and attention-path interventions on short-answer question answering in four MoE families. At half the experts the F1 loss of three families is mostly answer form: the model still finds the answer but stops answering briefly. In Qwen3-30B-A3B this loss is made at the two chat-template tokens before the answer, and in DeepSeek-V2-Lite at the first 8-32 prompt tokens, where its instruction sits; by answer likelihood these positions account for about half and a fifth of the damage. Reduction changes the first token's attention mass, norm and largest activation by less than 4%, and restoring super experts or massive activations recovers at most about a quarter of the loss, well below keeping the sensitive positions full. On real prompts with about 4% of tokens at full experts, placing them on the task interface keeps F1 within one point of the unmodified model, while placing them at random in the documents is about 3 points worse (two of five prespecified primary comparisons; the others were undecided or not run). Because the answer boundary always ends a single-turn prompt, and DeepSeek's loss was no longer detectable once its instruction moved away from the start, the sensitive tokens lay at the prompt ends in every sensitive layout we found, so rules that keep the ends are not distinguished from rules that read application marks. Time to first token falls about 14% on 11K-token prompts and not at 2K tokens.
| Comments: | 8 pages main text, 26 pages with appendices; supplementary material in ancillary files |
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.22454 [cs.CL] |
| (or arXiv:2610.22454v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.22454 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Zhiyuan Wang [view email]
[v1]
Thu, 15 Oct 2026 16:26:07 UTC (509 KB)
来源：arXiv cs.CL · arxiv.org