MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization
MemFold trains fixed-size soft memory with on-policy rewards and leads PersonaMem personalization across Qwen backbones.
MemFold compresses a query-conditioned textual memory into a fixed budget of K continuous vectors that form the reader's memory interface. The reader is trained on its own rollouts with group-relative task rewards and confidence-gated on-policy distillation from a frozen textual-memory teacher, which is never sampled and is removed at inference. Across three Qwen backbones it reports the highest measured accuracy on PersonaMem-32K and PersonaMem-128K, with wider margins at 128K, and transfers to PrefEval and LongMemEval without target-domain training. Ablations attribute most of the gain to the reward term and show the reader depends on instance-specific soft-memory content.
- Query-conditioned text memory is compressed into K continuous vectors.
- Training combines group-relative rewards with confidence-gated on-policy distillation.
- Highest measured accuracy on PersonaMem-32K and PersonaMem-128K across three Qwen backbones.
- Transfers to PrefEval and LongMemEval without target-domain training.
- The frozen teacher is removed entirely at inference.
From the original article
1 min read at huggingface.co
An assistant that serves the same user over a long horizon has to answer from what that user has revealed: which preferences still hold, which were revised, and which constraints apply now. Retaining that information is not the same as acting on it, and the two are usually optimized as if they were.
By Jingxuan Wu, Yuzhe Yang, Yiqiao Huang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.