arXiv:2512.05430v1 Announce Type: cross 
Abstract: Recent advances in large language models (LLMs) have transformed open-domain question answering, yet their effectiveness in music-related reasoning remains limited due to sparse music knowledge in pretraining data. While music information retrieval and computational musicology have explored structured and multimodal understanding, few resources support factual and contextual music question answering (MQA) grounded in artist metadata or historical context. We introduce MusWikiDB, a vector database of 3.2M passages from 144K music-related Wikipedia pages, and ArtistMus, a benchmark of 1,000 questions on 500 diverse artists with metadata such as genre, debut year, and topic. These resources enable systematic evaluation of retrieval-augmented generation (RAG) for MQA. Experiments show that RAG markedly improves factual accuracy; open-source models gain up to +56.8 percentage points (for example, Qwen3 8B improves from 35.0 to 91.8), approaching proprietary model performance. RAG-style fine-tuning further boosts both factual recall and contextual reasoning, improving results on both in-domain and out-of-domain benchmarks. MusWikiDB also yields approximately 6 percentage points higher accuracy and 40% faster retrieval than a general-purpose Wikipedia corpus. We release MusWikiDB and ArtistMus to advance research in music information retrieval and domain-specific question answering, establishing a foundation for retrieval-augmented reasoning in culturally rich domains such as music.

أدت التطورات الأخيرة في نماذج اللغة الكبيرة إلى تقديم ArtistMus، وهو معيار مصمم للإجابة على الأسئلة الموسيقية المعززة بالاسترجاع، إلى جانب MusWikiDB، وهي قاعدة بيانات متجهة تحتوي على 3.2 مليون مقطع من 144,000 صفحة مرتبطة بالموسيقى على ويكيبيديا. تهدف هذه المبادرة إلى تحسين فعالية التفكير الموسيقي من خلال توفير موارد منظمة للإجابة على الأسئلة الموسيقية الواقعية والسياقية.

Los avances recientes en modelos de lenguaje han llevado a la introducción de ArtistMus, un estándar diseñado para la respuesta a preguntas musicales aumentada por recuperación, junto con MusWikiDB, una base de datos vectorial que contiene 3.2 millones de pasajes de 144,000 páginas de Wikipedia relacionadas con la música. Esta iniciativa tiene como objetivo mejorar la efectividad del razonamiento musical al proporcionar recursos estructurados para la respuesta a preguntas musicales fácticas y contextuales.

Des avancées récentes dans les modèles de langage ont conduit à l'introduction d'ArtistMus, une référence conçue pour le questionnement musical augmenté par récupération, ainsi que de MusWikiDB, une base de données vectorielle contenant 3,2 millions de passages provenant de 144 000 pages Wikipédia liées à la musique. Cette initiative vise à améliorer l'efficacité du raisonnement musical en fournissant des ressources structurées pour le questionnement musical factuel et contextuel.

Recent advancements in large language models have led to the introduction of ArtistMus, a benchmark designed for retrieval-augmented music question answering, alongside MusWikiDB, a vector database containing 3.2 million passages from 144,000 music-related Wikipedia pages. This initiative aims to enhance the effectiveness of music-related reasoning by providing structured resources for factual and contextual music question answering.

ArtistMus: A Globally Diverse, Artist-Centric Benchmark for Retrieval-Augmented Music Question Answering

arXiv:2510.17238v2 Announce Type: replace 
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities in chain of thought (CoT) reasoning. However, the current LLM reasoning paradigm initiates thinking only after the entire input is available, which introduces unnecessary latency and weakens attention to earlier information in dynamic scenarios. Inspired by human cognition of thinking while reading, we first design a \textit{\textbf{streaming thinking}} paradigm for LLMs, where reasoning unfolds in the order of input and further adjusts its depth once reading is complete. We instantiate this paradigm with \textit{StreamingThinker}, a framework that enables LLMs to think while reading through the integration of streaming CoT generation, streaming-constraint training, and streaming parallel inference. Specifically, StreamingThinker employs streaming reasoning units with quality control for CoT generation, enforces order-preserving reasoning through streaming attention masks and position encoding, and leverages parallel KV caches that decouple input encoding from reasoning generation, thereby ensuring alignment and enabling true concurrency. We evaluate StreamingThinker on the Qwen3 model family across math reasoning, logical reasoning, and context-based QA reasoning tasks. Experimental results show that the StreamingThinker preserves performance comparable to batch thinking, while yielding an 80\% reduction in token waiting before the onset of reasoning and a more than 60\% reduction in time-level latency for producing the final answer, demonstrating the effectiveness of the streaming paradigm for LLM reasoning. Code will be released at https://github.com/EIT-NLP/StreamingLLM/tree/main/StreamingThinker.

تم تحسين نماذج اللغة الكبيرة (LLMs) من خلال إطار جديد يسمى StreamingThinker، والذي يسمح لها بالتفكير أثناء قراءة المدخلات بشكل متسلسل. تهدف هذه الطريقة إلى تقليل زمن الانتظار وتحسين الانتباه للمعلومات السابقة، مما يعالج القيود في نماذج التفكير التقليدية التي لا تبدأ في العمل إلا بعد استلام المدخلات بالكامل.

Los Modelos de Lenguaje Grande (LLMs) han sido mejorados con un nuevo marco llamado StreamingThinker, que les permite razonar mientras leen la entrada de manera secuencial. Este enfoque busca reducir la latencia y mejorar la atención a la información anterior, abordando las limitaciones de los paradigmas de razonamiento LLM tradicionales que solo se activan después de recibir toda la entrada.

Les modèles de langage de grande taille (LLMs) ont été améliorés avec un nouveau cadre appelé StreamingThinker, qui leur permet de raisonner tout en lisant l'entrée de manière séquentielle. Cette approche vise à réduire la latence et à améliorer l'attention portée aux informations antérieures, répondant ainsi aux limitations des paradigmes de raisonnement LLM traditionnels qui ne s'activent qu'après la réception de l'ensemble de l'entrée.

Large Language Models (LLMs) have been enhanced with a new framework called StreamingThinker, which allows them to engage in reasoning while reading input sequentially. This approach aims to reduce latency and improve attention to earlier information, addressing limitations in traditional LLM reasoning paradigms that only activate after the entire input is received.

ArtistMus: A Globally Diverse, Artist-Centric Benchmark for Retrieval-Augmented Music Question Answering

Was this article worth reading? Share it

LucidQuery AI

Halmblog Music

Magicley AI