Recently, the AISpeech-Shanghai Jiao Tong University (SJTU) X-LANCE Lab announced that five research papers have been accepted by NeurIPS 2025, a premier global conference in machine learning and computational neuroscience. 

The accepted papers focus on frontier directions such as data engineering for LLM training and instruction tuning, interpretable representation learning, and robust generalization. These research findings provide key technical support for AISpeech's LLM human-machine dialogue products. Key applications include enhancing multi-agent collaboration, complex task decomposition and execution, improving agent autonomy and fault tolerance, achieving greater robustness and cross-modal consistency in end-to-end speech understanding and generation, and establishing an interpretable foundation for managing agent distribution and personalized dialogue memory. These advancements lead to more stable, intelligent, and natural interactions in scenarios like in-car voice assistants, smart home devices, and customer service platforms.

One highlighted paper, "Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations," introduces the MONA paradigm. MONA addresses the challenge of selecting the most relevant data for fine-tuning LLMs on specific tasks. Instead of relying on textual semantics or neural embeddings, MONA uses activations from the base model's internal neurons to represent samples. A sparse autoencoder then disentangles these polysemantic activations into interpretable monosemantic features. By performing task alignment and similarity measurement in this sparse space, MONA enables more robust and efficient data selection, creating a high-quality, minimal data foundation for enterprise-level instruction tuning.

AISpeech has a long-standing record of academic excellence, with numerous publications at top-tier conferences like ICASSP, INTERSPEECH, ACL, EMNLP, AAAI, and ICML. The company is committed to advancing the translation of research findings into practical applications. 

As a leading conversational AI platform company, AISpeech drives innovation from core technology development to industrial deployment. It is recognized as a key player in the national AI strategy, operating the "MOST's National Platform for Language Computing AI Innovation". AISpeech also leads the "Jiangsu Provincial Key Laboratory of Language Computing and Applications" in collaboration with SJTU and Soochow University. The company has undertaken over ten national and provincial-level R&D projects and boasts an extensive intellectual property portfolio, including 685 authorized invention patents. AISpeech adheres to an independent LLM strategy, building a "1+N Reliable Distributed LLM Agent System" to serve enterprise clients effectively.