AI BriefingFuriosaAI Presents Papers on AI Performance at ICML and ACL
Key point
FuriosaAI presented 6 research papers at ICML and ACL that enhance model efficiency, reasoning capability, and flexibility.
Details
Engineers and researchers at FuriosaAI presented a total of 6 papers at ICML 2025 and ACL 2025. This research focuses on three core areas to maximize the performance of the high-performance inference accelerator RNGD and next-generation products.
The key research areas are as follows:
- More efficient: Accelerating large-scale model inference and automating model parallelization
- More capable: Strengthening complex reasoning ability across diverse domains beyond math, such as law
- More flexible: Efficient fine-tuning of new architectures such as SSM (State Space Models)
The key paper contents are as follows:
Parameter-Efficient Fine-Tuning of State Space Models proposes the SDT (Sparse Dimension Tuning) technique optimized for SSM-based models such as Mamba. This overcomes the limitations of existing LoRA and achieves state-of-the-art performance.
VersaPRM introduces a multi-domain PRM leveraging synthetic reasoning data to address the limitations of existing PRM (Process Reward Model), which underperform outside of math domains. This demonstrated meaningful performance improvements in fields such as Law.
TabFlex covers technology for scaling learning on large-scale Tabular Data using Linear Attention.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.