LightThinker++: From Reasoning Compression to Memory Management

Researchers developed LightThinker++, a new framework that enables large language models to compress intermediate reasoning thoughts and manage memory more efficiently. The system reduces peak token usage by up to 70% while improving accuracy by 2.42% and maintaining performance over extended reasoning tasks.

• →LightThinker++ reduces LLM peak token usage by 69.9% while improving accuracy by 2.42% in standard reasoning tasks.
• →The framework introduces Explicit Adaptive Memory Management to prevent logical bottlenecks from irreversible compression.
• →Inference time is reduced by 26% with minimal accuracy loss compared to traditional approaches.
• →In long-horizon tasks, the system maintains stable performance beyond 80 rounds with 60-70% footprint reduction.
• →The approach achieves an average 14.8% performance gain across complex agentic scenarios.