ExplicitLM: Decoupling Knowledge from Parameters via Explicit Memory Banks

2025-11-03

AI Summary

Large language models (LLMs) implicitly encode knowledge in their parameters, resulting in high knowledge staleness, poor interpretability, and difficulty in updating. Method: We propose an explicit memory bank architecture that decouples knowledge from model parameters by storing factual and pattern knowledge as human-readable token sequences in an external memory repository. Inspired by cognitive dual-system theory, the memory distinguishes between frozen facts and learnable patterns. A differentiable two-stage retrieval mechanism is introduced: product quantization–based key decomposition for coarse-grained filtering, followed by Gumbel-Softmax sampling with exponential moving average for fine-grained matching and end-to-end training. Contribution/Results: On knowledge-intensive tasks, our approach achieves up to a 43.67% improvement; under few-shot settings (10K examples), it attains 3.62× the baseline performance, increases memory hit rate by 49%, and significantly enhances reasoning transparency and knowledge maintainability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large language models suffer from knowledge staleness and lack of interpretability due to implicit knowledge storage across entangled network parameters, preventing targeted updates and reasoning transparency. We propose ExplicitLM, a novel architecture featuring a million-scale external memory bank storing human-readable knowledge as token sequences, enabling direct inspection and modification. We design a differentiable two-stage retrieval mechanism with efficient coarse-grained filtering via product key decomposition (reducing complexity from $mathcal{O}(N cdot |I|)$ to $mathcal{O}(sqrt{N} cdot |I|)$) and fine-grained Gumbel-Softmax matching for end-to-end training. Inspired by dual-system cognitive theory, we partition knowledge into frozen explicit facts (20%) and learnable implicit patterns (80%), maintained through Exponential Moving Average updates for stability. ExplicitLM achieves up to 43.67% improvement on knowledge-intensive tasks versus standard Transformers, with 3.62$ imes$ gains in low-data regimes (10k samples). Analysis shows strong correlations between memory retrieval and performance, with correct predictions achieving 49% higher hit rates. Unlike RAG systems with frozen retrieval, our jointly optimized architecture demonstrates that interpretable, updatable models can maintain competitive performance while providing unprecedented knowledge transparency.

Problem

• Decouples knowledge from parameters using explicit memory banks
• Solves knowledge staleness and interpretability issues in LLMs
• Enables direct knowledge inspection and targeted updates

Innovation

• External memory bank stores human-readable knowledge tokens
• Differentiable two-stage retrieval with product key decomposition
• Partitioned knowledge into frozen facts and learnable patterns