MemTrain: Self-Supervised Context Memory Training

In three linesMemTrain introduces a self-supervised training framework to enhance context-memory capabilities of LLM agents. Two coupled proxy tasks on Wikipedia (masked entity reconstruction and intermediate memory recall) are jointly optimized using GRPO. Achieves gains up to 17.67 points on long-text QA and search-based QA benchmarks.

Summary generated by Claude — human-verified