Research map / Large language models / Retrieval and knowledge
Knowledge Card: Filling LLMs' Knowledge Gaps with Plug-in Specialized 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 Knowledge Card, a modular framework where many small domain-specialized LMs (OPT-1.3B based) generate background knowledge that is filtered by relevance, pruning, and factuality selectors and plugged into black-box LLMs via bottom-up or top-down integration.
Problem
Large LLMs are static and expensive to update, and fail on factual, long-tail, and up-to-date knowledge-intensive tasks; existing retrieval and generated-knowledge methods rely on a single fixed knowledge source and do not support modular, community-contributed knowledge for black-box LLMs.
Approach
Train many small specialized LMs (knowledge cards) on domain corpora, use them as parametric knowledge generators, filter their outputs with relevance, pruning (summarization), and factuality selectors (summarization factuality plus retrieval-augmented fact-checking with top-k sampling), and integrate them into a black-box LLM either by activating all cards at once (bottom-up) or by letting the LL
Key results
- On MMLU, Top-Down exp reaches 72.8% overall vs. 68.3% for Codex and 72.2% for Flan-PaLM, a 6.6% improvement over Codex.
- On LUN misinformation detection, Bottom-Up reaches 89.8 two-way BAcc vs. 65.6 for Codex, and Top-Down exp 91.3.
- On MidtermQA open-book, Top-Down auto reaches 87.5 EM vs. 55.1 for Codex, using a single 1.3B knowledge card added to a 175B LLM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use top-down explicit selection with Knowledge Card | Codex · MMLU | accuracy: 68.3 -> 72.8 | Helped |
| Use bottom-up integration of Knowledge Cards | Codex · LUN misinformation detection two-way | balanced accuracy: 65.6 -> 89.8 | Helped |
| Use bottom-up integration of Knowledge Cards | Codex · LUN misinformation detection four-way | balanced accuracy: 52.8 -> 70.6 | Helped |
| Add a midterm-election News knowledge card with top-down automatic selection | Codex · MidtermQA open-book | exact match: 55.1 -> 87.5 | Helped |
| Add PubMed, IMDB, and BookCorpus knowledge cards (ablation) | ChatGPT · LUN misinformation detection two-way | balanced accuracy: 80.1 -> 80.7; 80.6; 82.3 | Helped |
| Add an in-domain News knowledge card (ablation) | ChatGPT · LUN misinformation detection two-way | balanced accuracy: 82.3 -> 85.7 | Helped |
| Add a Wikipedia knowledge card after News (ablation) | ChatGPT · LUN misinformation detection two-way | balanced accuracy: 85.7 -> 76.5 | Hurt |
| Remove the relevance selector (ablation) | Bottom-Up Knowledge Card · LUN misinformation detection | model performance | Hurt |
| Remove the pruning selector (ablation) | Bottom-Up Knowledge Card · LUN misinformation detection | model performance | Hurt |
| Remove the factuality selector (ablation) | Bottom-Up Knowledge Card · LUN misinformation detection | model performance | Hurt |
| Increase the number of generated documents per knowledge card (n1) (ablation) | Bottom-Up Knowledge Card · Knowledge stream analysis | model performance | No clear effect |
| Retain more documents after the knowledge selectors (n2) (ablation) | Bottom-Up Knowledge Card · Knowledge stream analysis | model performance | Hurt |
Surprising findings
- Generated-knowledge baselines (GKP, Recitation, GRTR) underperform vanilla Codex on MidtermQA, suggesting probing LLM parametric knowledge is counterproductive when that knowledge is outdated.
- Adding the Wikipedia knowledge card alone to a 4-card setup reduced two-way BAcc in the accumulation study (85.7 to 76.5), an unexplained drop.
Limitations
- Knowledge cards built on a small 1.3B OPT model occasionally produce low-quality, degenerate, or off-topic documents.
- The factuality selector is biased toward information-rich domains with Wikipedia coverage and may score new knowledge poorly.
- Yes/no prompting for knowledge need is imperfect, with LLMs sometimes overconfident (no-incorrect cases).
Experimental setup
Datasets: MMLU LUN MidtermQA Wikipedia PubMed IMDB BookCorpus ATOMIC ConceptNet UMLS Wikidata YAGO
Benchmarks: MMLU LUN misinformation dataset (two-way and four-way) MidtermQA (open-book, two-way, four-way)
Metrics: accuracy balanced accuracy macro F1 exact match F1
Baselines: Codex PaLM Flan-PaLM Atlas RePlug RePlug LSR GKP Recitation GRTR Si et al.
Builds on: generated knowledge prompting RePlug recitation-augmented language models OPT Codex
Closest papers by meaning
- Knowledge Fusion of Large Language Models (ICLR 2024 · similarity 0.89)
- KoLA: Carefully Benchmarking World Knowledge of Large Language Models (ICLR 2024 · similarity 0.88)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.88)
- The Rise of Parameter Specialization for Knowledge Storage in Large Language Models (NeurIPS 2025 · similarity 0.88)
- BenchmarkCards: Standardized Documentation for Large Language Model Benchmarks (NeurIPS 2025 · similarity 0.88)
- KScope: A Framework for Characterizing the Knowledge Status of Language Models (NeurIPS 2025 · similarity 0.87)
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.