Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.22454

Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss

Zhiyuan Wang·Han Wang·Shuo Zhang
View original source ↗Read source document ↗Export metadata ↓

Abstract

Running prompt tokens through fewer routed experts during prefill makes mixture-of-experts (MoE) serving cheaper, but what it damages has not been measured. We locate the damage with token-, layer- and attention-path interventions in four MoE families on short-answer question answering. At half the experts the F1 loss is mostly answer form: the model still finds the answer but no longer answers briefly. The loss is made at a few tokens: the two chat-template tokens before the answer in Qwen3-30B-A3B, and the first 8-32 tokens, where the instruction sits, in DeepSeek-V2-Lite. Keeping them at full experts removes most of the F1 loss and about half and a fifth of the increase in answer NLL. These tokens are not the familiar suspects: reduction leaves the first-token sink within 4%, and restoring super experts or massive activations recovers at most a quarter of the loss. On real prompts the localisation becomes a placement rule: with about 4% of tokens at full experts, putting them on the task interface keeps F1 within one point of the unmodified model, while the same budget scattered in the documents is about 3 points worse in both prespecified primary tests of this comparison. Time to first token falls about 14% on 11K-token prompts.

Explore related subjects

cs.CLcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zhiyuan Wang, Han Wang, Shuo Zhang. 2026-10-25. Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss. https://arxiv.org/abs/2610.22454

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Semantic Self-Consistency: Enhancing Language Model Reasoning via Semantic Weighting

While large language models (LLMs) have rapidly improved their performance on a broad number of tasks, they still often fall short on reasoning tasks. As LLMs become more integrated in diverse real-world tasks, advancing their reasoning capabilities is crucial to their effectiveness in nuanced, complex problems. Wang et al.'s self-consistency framework reveals that sampling multiple rationales before taking a majority vote reliably improves model performance across various closed-answer reasoning tasks. Standard methods based on this framework aggregate the final decisions of these rationales but fail to utilize the semantic information detailed in the step-by-step reasoning paths. Our work introduces semantic self-consistency, enhancing this approach by incorporating and analyzing both the reasoning paths of these rationales in addition to their final decisions before taking a majority vote. These methods not only improve the reliability of reasoning paths but also cause more robust performance on complex reasoning tasks.

cs.CL↗
arXiv2026

Beyond Isolated Dots: Benchmarking Structured Table Construction as Deep Knowledge Extraction

Large language models (LLMs) are increasingly expected to extract and synthesize structured information from complex real-world documents, yet they typically produce unstructured, hard-to-verify paragraph-style outputs. We introduce AOE (Arranged and Organized Extraction), a bilingual benchmark that evaluates LLMs' ability to distill long-form, multi-document inputs (64k to 1.8M tokens) into structured tables. AOE comprises 12 tasks across four domains (academic, financial, legal, and medical), each grounded in genuine analytical needs, with schema guidance provided to isolate extraction quality from formatting variance. We evaluate 12 frontier LLMs (both open-weight and proprietary) under four generation settings of increasing autonomy: direct full-document input, RAG-based retrieval, offline document-browsing agent, and online web-search agent. Even the best-performing model achieves only 42.7% Cell F1, and retrieval-augmented approaches often degrade rather than improve performance, revealing that structured extraction from long contexts remains a fundamental challenge - an "Area of Effect" (AOE) that impacts all tested systems.

cs.CL↗
arXiv2026

PersonaFuse: A Situation-Aware Post-Training Framework for Adaptive LLM Behavior in Enterprise Tasks

Large language models (LLMs) are increasingly deployed in enterprise settings to support task execution and workflow automation. Despite the strong performance of general-purpose models on standard benchmarks, organizations often need to adapt them to domain-specific tasks and operational constraints in practice. However, existing post-training approaches introduce capability trade-offs and lack mechanisms for situational adaptation. In this work, we propose PersonaFuse, a post-training framework that enables situation-aware behavioral adaptation in LLMs. Grounded in Trait Activation Theory and operationalized through the Five-Factor Model, PersonaFuse dynamically adjusts model behavior based on task context using a mixture-of-experts architecture with trait-level experts and a context-aware routing mechanism. We evaluate PersonaFuse at three levels. At the capability level, it achieves balanced improvements across general reasoning, interaction quality, and safety without introducing trade-offs. At the workflow level, it improves task completion in enterprise scenarios involving tool use and structured execution across both B2B and B2C tasks. At the deployment level, we analyze model selection under cost and safety constraints using a capability-cost frontier framework, showing that models adapted with PersonaFuse achieve more favorable trade-offs and are consistently selected as optimal across deployment scenarios. PersonaFuse provides a practical framework for developing LLMs that are not only capable, but also adaptable and deployable in real-world enterprise environments.

cs.CL↗