Reproduce “Learning to Forget: Sleep-Inspired Memory Consolidation for Resolving Proactive Interference in Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ying Xie

SleepGate is a biologically inspired framework augmenting Large Language Models with a learned sleep cycle to manage proactive interference within their working memory. It integrates conflict-aware tagging, a forgetting gate, and consolidation over the KV cache, achieving near-perfect retrieval accuracy for current information on a synthetic benchmark while significantly suppressing stale associations.

Key insights

  • Proactive interference in LLMs, analogous to human cognitive psychology, can be systematically addressed by integrating biologically inspired 'sleep-like' memory consolidation mechanisms into transformer architectures.
  • Content-dependent semantic signatures and conflict detection enable selective identification and suppression of superseded information, offering a targeted approach beyond indiscriminate cache pruning.
  • A soft attention biasing mechanism, driven by a learned forgetting gate, provides a differentiable and robust method to downweight stale memories, allowing for graceful degradation rather than abrupt hard evictions, which is crucial for training stability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.