超越提示词工程:基于对数几率空间融合的语音大模型高效鲁棒上下文偏置 (LOGIC)
Beyond Prompting: Efficient and Robust Contextual Biasing for Speech LLMs via Logit-Space Integration (LOGIC)
- arXiv ID: 2601.15397 [cs.AI]
- Primary Subject: Artificial Intelligence (
cs.AI)- Secondary Subjects: Computation and Language (
cs.CL), Sound (cs.SD)- Authors: Peidong Wang, Jian Xue, Jinyu Li
- Submission History:
[v1]21 Jan 2026;[v3]10 Sep 2026 (latest revision)
文章背景与核心概要
语音大语言模型 (Speech Large Language Models, Speech LLMs) 在通用人机对话中展现出色,但由于训练数据固化,难以准确识别快速涌现的专有名词、联系人姓名或个性化歌单等领域实体。传统的提示词 (Prompting) 方案在实体列表增加时面临上下文窗口膨胀、推理延迟飙升与“迷失在中间”等问题;而生成式纠错 (Generative Error Correction, GEC) 则常导致过度纠正与虚假实体幻觉。为此,作者提出了 LOGIC 框架,直接在模型的解码层对数几率 (Logit) 空间中进行上下文信息融合,彻底将上下文注入与输入处理解耦,实现了相对提示词长度的常数级时间复杂度。在多模态模型 Phi-4-MM 跨越 11 种跨国语言环境的测试中,LOGIC 将实体词错误率 (WER) 相对降低了 9%,而虚警率增幅仅为微乎其微的 0.30%。
📌 内容概要
📌 Summary
语音大语言模型 (Speech LLMs) 在日常通用对话任务中表现优异,但受限于静态的预训练知识库,往往难以有效识别迅速更迭的新词或特定领域实体 (例如个人联系人姓名、专业术语或个性化音乐播放列表) 。
Speech Large Language Models (Speech LLMs) excel at general conversational tasks, but struggle to recognize rapidly emerging or domain-specific entities (such as contact names, technical jargon, or playlists) due to their static training knowledge.
传统的解决思路通常依赖提示词工程 (Prompting) ,但其扩展能力极差——随着待识别实体列表的扩充,很容易遭遇上下文窗口饱和、推理延迟成倍上升以及“迷失在中间” (Lost-in-the-Middle) 现象。与此同时,另一种生成式纠错 (GEC) 方案则经常面临“过度纠错”的问题,甚至会凭空凭造出音频中完全不存在的实体幻觉。
Traditional solutions like prompting fail to scale efficiently, leading to context window limitations, high inference latency, and the "lost-in-the-middle" phenomenon. Meanwhile, Generative Error Correction (GEC) suffers from over-correction and entity hallucinations.
为破解上述难题,研究团队提出了 LOGIC (Logit-Space Integration for Contextual Biasing,即对数几率空间上下文偏置融合) 框架,该方法直接在神经网络的最终解码层实施干预。通过将上下文偏置的注入过程与输入端表征处理完全解耦,LOGIC 实现了相对于提示词长度的常数级时间复杂度 \(\mathcal{O}(1)\)。在搭载 Phi-4-MM 模型并覆盖 11 种多语言区域设置的实验中,LOGIC 取得了实体词错误率 (Word Error Rate, WER) 平均相对下降 9% 的显著提升,且虚警率 (False Alarm Rate) 仅轻微上升 0.30%。
To solve this, the authors introduce LOGIC (Logit-Space Integration for Contextual Biasing), a framework operating directly in the decoding layer. By decoupling context injection from input processing, LOGIC achieves constant-time complexity relative to prompt length. Experiments using the
Phi-4-MMmodel across 11 multilingual locales show an average 9% relative reduction in Entity Word Error Rate (WER) with only a 0.30% increase in the False Alarm Rate.
📄 论文元数据
📄 Paper Metadata
- arXiv 编号: 2601.15397
- 主学科领域: 人工智能 (
cs.AI) - 次要学科领域: 计算与语言 (
cs.CL),声音音频 (cs.SD) - 论文作者: Peidong Wang, Jian Xue, Jinyu Li
- 提交历史:
[v1]2026年1月21日[v3]2026年9月10日 (最新修订版本)
- arXiv ID: 2601.15397
- Primary Subject: Artificial Intelligence (
cs.AI)- Secondary Subjects: Computation and Language (
cs.CL), Sound (cs.SD)- Authors: Peidong Wang, Jian Xue, Jinyu Li
- Submission History:
[v1]21 Jan 2026[v3]10 Sep 2026 (latest revision)
🔍 论文摘要
🔍 Abstract
受文化演变、网络热点变迁以及用户高度定制化数据驱动,新实体的快速涌现给现有的语音大语言模型 (Speech LLMs) 带来了严峻挑战。尽管这些模型在宽泛的通用对话任务中表现卓越,但其静态的训练期知识库极大地限制了其识别特定领域词汇的能力,例如通讯录人名、歌单名称或行业技术行话。现存主流方案主要依赖提示词,但其可扩展性极差:随着实体清单长度的增长,提示词方法面临着上下文窗口受限、推理延迟增加以及“迷失在中间”现象。另一类替代方法即生成式纠错 (GEC),试图通过后处理机制重写转写文本,但又极易陷入“过度纠正”的陷阱,凭空臆造出说话人根本未提及的虚假实体。
在本工作中,我们提出了 LOGIC (基于对数几率空间融合的上下文偏置机制) ,这是一个直接在解码层操作的高效且鲁棒的框架。与提示词机制截然不同,LOGIC 将上下文注入与输入处理深度解耦,从而保证了关于提示词长度的常数级时间复杂度。利用 Phi-4-MM 模型在横跨 11 种多语言本地化场景下开展的大规模实验表明,LOGIC 实现了实体词错误率 (Entity WER) 平均 9% 的相对下降,且虚警率增量仅为可以忽略不计的 0.30%。
The rapid emergence of new entities -- driven by cultural shifts, evolving trends, and personalized user data -- poses a significant challenge for existing Speech Large Language Models (Speech LLMs). While these models excel at general conversational tasks, their static training knowledge limits their ability to recognize domain-specific terms such as contact names, playlists, or technical jargon. Existing solutions primarily rely on prompting, which suffers from poor scalability: as the entity list grows, prompting encounters context window limitations, increased inference latency, and the "lost-in-the-middle" phenomenon. An alternative approach, Generative Error Correction (GEC), attempts to rewrite transcripts via post-processing but frequently suffers from "over-correction", introducing hallucinations of entities that were never spoken.
In this work, we introduce LOGIC (Logit-Space Integration for Contextual Biasing), an efficient and robust framework that operates directly in the decoding layer. Unlike prompting, LOGIC decouples context injection from input processing, ensuring constant-time complexity relative to prompt length. Extensive experiments using the Phi-4-MM model across 11 multilingual locales demonstrate that LOGIC achieves an average 9% relative reduction in Entity WER with a negligible 0.30% increase in False Alarm Rate.
🔗 全文获取与参考文献
🔗 Full-Text & References
- 论文研读: 查看 PDF | HTML 网页版
- 源码与学术指标:
- Google Scholar
- Semantic Scholar
- NASA ADS
- Read the Paper: View PDF | HTML Version
- Source Code & Scholarly Citations:
- Google Scholar
- Semantic Scholar
- NASA ADS