Research map / Large language models / Retrieval and knowledge
From Static to Dynamic: Knowledge Metabolism for Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DynaMind, a memory-augmented LLM framework with a credibility-updated long/short-term memory and operator pipeline that delivers knowledge to LLMs without parameter updates, evaluated on two self-constructed benchmarks.
Problem
LLMs have static parameter knowledge that is costly to update, prone to catastrophic forgetting, and leads to hallucinations from obsolete or inaccurate knowledge.
Approach
A standalone memory manager with knowledge metabolism (time decay and persistence, credibility scores updated via discriminator feedback) plus an inference engine that issues operator commands (coordinator, searcher, browser, responder, discriminator), letting LLMs update knowledge without changing parameters.
Key results
- On Knowledge-driven Reasoning, DynaMind raises GPT-4 accuracy from 16.0 to 92.5 and GPT-3.5 from 8.5 to 89.0.
- Falcon-40B accuracy goes from 17.5 to 85.0 and Llama-30B from 6.0 to 56.5 with DynaMind.
- Knowledge Manipulation create/update/delete accuracy for GPT-4 is 95.5/85.0/71.5 and for GPT-3.5 is 92.0/81.0/71.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DynaMind to OpenAI GPT-3.5 | OpenAI GPT-3.5 · Knowledge-driven Reasoning | accuracy: 8.5 -> 89.0 (+80.5) | Helped |
| Add DynaMind to OpenAI GPT-4 | OpenAI GPT-4 · Knowledge-driven Reasoning | accuracy: 16.0 -> 92.5 (+76.5) | Helped |
| Add DynaMind to Falcon-40B | Falcon-40B · Knowledge-driven Reasoning | accuracy: 17.5 -> 85.0 (+67.5) | Helped |
| Add DynaMind to Llama-30B | Llama-30B · Knowledge-driven Reasoning | accuracy: 6.0 -> 56.5 (+50.5) | Helped |
Limitations
- Evaluation is limited to two self-constructed benchmarks with little methodological detail in the paper body.
- Baselines are reported only as the raw model accuracy without comparison to AutoGPT, BabyAGI, or LangChain in results.
Experimental setup
Benchmarks: Knowledge-driven Complex Reasoning Knowledge Manipulation
Metrics: accuracy
Baselines: AutoGPT BabyAGI LangChain
Builds on: AutoGPT BabyAGI LangChain Multi-Armed Bandit
Closest papers by meaning
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.89)
- Knowledge Fusion of Large Language Models (ICLR 2024 · similarity 0.89)
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.89)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.88)
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.88)
- MEMORYLLM: Towards Self-Updatable Large Language Models (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.