PI-Mem: Parallel-Iterative Memory Extends LLM Context to 3.6M Tokens
A novel approach known as PI-Mem (Parallel-Iterative Memory) tackles the challenges of long-context reasoning in large language models by allowing simultaneous processing of all data segments while iteratively refining a collective memory. This technique addresses the drawbacks of traditional sequential recurrent-memory approaches that risk overwriting initial information. As outlined in a preprint on arXiv (ID: 2608.03048), PI-Mem evaluates chunks based on existing memory, identifies new information, and integrates it into a streamlined shared memory. Enhanced through reinforcement learning, this method facilitates an adaptive early exit once adequate data is collected. It reportedly expands context reasoning capabilities to 3.6 million tokens, improving efficiency for tasks that demand a thorough comprehension of extensive materials, including legal texts and scientific articles.
Key facts
- PI-Mem processes all chunks in parallel and iteratively refines shared memory.
- It addresses sequential chunk-wise updates that can overwrite early evidence.
- It overcomes serial inter-chunk dependencies that limit parallelism.
- The mechanism uses reinforcement learning with a turn-efficiency reward.
- It enables adaptive early exit once sufficient evidence is gathered.
- The approach extends context reasoning to 3.6 million tokens.
- The paper is available on arXiv with ID 2608.03048.
- The announcement type is cross.
Entities
Institutions
- arXiv