Researchers from MIT, NVIDIA and Zhejiang University have created TriAttention, a KV cache compression method that match...

Researchers from MIT, NVIDIA and Zhejiang University have created TriAttention, a KV cache compression method that matches full attention accuracy while achieving 2.5x higher throughput. The method addresses a key limitation in existing approaches by operating in pre-RoPE space, where Query and Key vectors cluster around stable centres. On the AIME25 mathematical reasoning benchmark, TriAttention achieved 32.9% accuracy versus just 17.5% for the best existing method at the same memory budget. https://www.marktechpost.com/2026/04/11/researchers-from-mit-nvidia-and-zhejiang-university-propose-triattention-a-kv-cache-compression-method-that-matches-full-attention-at-2-5x-higher-throughput/ #AIagent #AI #GenAI #AIResearch