Research map / Large language models / Retrieval and knowledge
From Style to Facts: Mapping the Boundaries of Knowledge Injection with Finetuning
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.
Large grid-search study (~4,000 LoRA finetuning runs on Gemini v1.5 Pro/Flash) showing knowledge injection via finetuning depends on data format, information type, quantity and evaluation alignment rather than a fundamental gap with task customization.
Problem
Finetuning is commonly believed to work well for style/tone customization but poorly for knowledge injection; it is unclear which factors (not the task-vs-knowledge dichotomy itself) drive this gap.
Approach
Treats task customization and knowledge injection as points on a spectrum and systematically varies information quantity, information type (numerical/categorical/emotional), entity type (real/fictional/persona), training data format, and evaluation task, with controls for confounders.
Key results
- Question-answer training data gives much stronger knowledge generalization than Wikipedia-article training data; Wikipedia-article training yields low accuracy across evaluation tasks.
- Tone teaching reached 96.3% evaluator accuracy (exact match 94.7%), while Wikipedia 2024-event knowledge injection reached only 11% QA accuracy.
- Accuracy falls roughly with a power law as information quantity grows from 20 to 4,000 facts; numerical facts are consistently harder to finetune than categorical or emotional facts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Teach a target writing tone instead of new Wikipedia facts | Gemini v1.5 Pro LoRA finetuning · Tone teaching vs. Wikipedia fact questions | accuracy: 96.3% vs. 11% | Mixed |
| Use question-answer pairs instead of Wikipedia-style articles for training (ablation) | Gemini v1.5 Pro and Gemini v1.5 Flash · Knowledge-injection evaluation tasks | accuracy | Helped |
| Use reasoning examples instead of question-answer examples for training (ablation) | Gemini v1.5 Pro · Reasoning evaluation tasks | accuracy | Helped |
| Use reasoning examples instead of question-answer examples for training (ablation) | Gemini v1.5 Pro · Direct question-answer evaluation tasks | accuracy | Hurt |
| Train on numerical rather than categorical facts (ablation) | Gemini v1.5 Pro · 20-fact knowledge injection | accuracy | Hurt |
| Train on emotional rather than categorical facts (ablation) | Gemini v1.5 Pro · 20-fact knowledge injection | accuracy | No clear effect |
| Increase the number of facts to learn (ablation) | Gemini v1.5 Pro · Knowledge-injection evaluation tasks | accuracy | Hurt |
| Teach facts about a persona rather than real-world or fictional entities (ablation) | Gemini v1.5 models · 20-fact knowledge injection | accuracy | No clear effect |
| Align the training data format with the evaluation task (ablation) | Gemini v1.5 Pro and Gemini v1.5 Flash · Knowledge-injection evaluation tasks | accuracy | Helped |
| Use question-answer training data instead of Wikipedia articles (ablation) | Gemini v1.5 Pro · Wikipedia fill-in-the-blank evaluation | accuracy | Helped |
| Use question-answer training data instead of Wikipedia articles (ablation) | Gemini v1.5 Pro · Direct question-answer evaluation | accuracy | Helped |
| Apply finetuned knowledge in multi-step reasoning rather than direct recall | Gemini v1.5 models · Reasoning evaluation tasks | accuracy | Hurt |
Surprising findings
- Wikipedia-style articles are among the least effective training formats despite being the most common pretraining format, even after controlling for information density.
- Persona vs. real-world/fictional entity type made no significant difference to finetuning effectiveness once other factors were controlled.
Limitations
- Experiments use proprietary Gemini v1.5 models only, so generality to other model families is untested.
- Mechanisms behind the numerical-vs-categorical gap and Wikipedia-format weakness are not identified.
- Finetuning regime is limited to ~10k-100k token datasets, not continued pretraining.
Experimental setup
Datasets: Wikipedia stub articles (real entities) synthetic fictional entities synthetic personas 2024 Wikipedia articles (meteorology and sports) tone conversations generated from seed chats
Benchmarks: Direct QA Conversational QA Role-play QA Reasoning tasks Wikipedia fill-in-the-blank Wikipedia sentence completion tone matching 2024 Wikipedia exam questions
Metrics: accuracy exact match accuracy
Baselines: in-context learning Wikipedia-article training format
Builds on: LoRA reversal curse random access limitations of autoregressive LMs
Closest papers by meaning
- Understanding Finetuning for Factual Knowledge Extraction (ICML 2024 · similarity 0.89)
- A Closer Look at the Limitations of Instruction Tuning (ICML 2024 · similarity 0.89)
- Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.88)
- How new data permeates LLM knowledge and how to dilute it (ICLR 2025 · similarity 0.88)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.87)
- From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data (ICLR 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.