AI BriefingFuriosaAI Presents Full-Stack Software Efficiency at ICML 2026
Key point
FuriosaAI presented four papers at ICML 2026 addressing efficiency in LLM inference, training, and serving.
Details
FuriosaAI presented four research papers on software optimization at ICML 2026, covering everything from AI model architecture and serving pipelines to hardware execution. All papers were authored by the AI Research Group within Furiosa's Algorithm Team.
- ReJump: Represents the LLM reasoning process as trees and discontinuous transitions called 'jumps' to analyze execution patterns such as backtracking, verification, and computation. This improved Best-of-N selection performance by up to 9.1% on reasoning tasks.
- LoSA: Addresses the KV cache memory growth issue in block-wise diffusion language models. It reuses prefix attention for stable tokens and applies sparse attention only to active tokens, enhancing scalability for long-form generation.
- AsyncOPD: Proposes an asynchronous on-policy distillation pipeline that decouples rollout generation from learner updates. It increased training throughput by 1.6–3.8x compared to synchronous baselines.
- EfficientRollout: Combines a 4-bit self-drafting model with system-aware speculative decoding and dynamic draft length adjustment. It accelerated RL rollouts by up to 19.6% and end-to-end training by 12.7% without loss.
During the ICML conference, FuriosaAI hosted a community event at its Gangnam headquarters in Seoul, attended by over 100 researchers, system architects, engineers, and partners, to share its direction for building a full-stack ecosystem.
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.