Research map / Large language models / Instruction and fine-tuning
LLM DNA: Tracing Model Evolution via Functional Representations
Zhaomin Wu, Haodong Zhao, Ziyang Wang, Jizhou Guo, Qian Wang, Bingsheng He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DNA (BGE) + SVM instead of PhyloLM + SVM for relation detection | LLM DNA · LLM relation detection (305 LLMs) | AUC: 0.788 -> 0.992 | Improved |
| replace EmbedLLM learned embeddings with frozen training-free DNA | EmbedLLM matrix factorisation routing · Model routing (EmbedLLM dataset) | Routing accuracy: 0.665 -> 0.672 | Improved |
| extract DNA from random word prompts instead of benchmark data (ablation) | DNA (Qwen3) + SVM · LLM relation detection | Accuracy: 0.919 -> 0.949 | Improved |
| bypass chat template (completion-style prompts) for DNA extraction (ablation) | DNA (Qwen3) · LLM relation detection | Accuracy: 0.919 -> 0.930 | Improved |
| bypass chat template for DNA extraction (ablation) | DNA (MPNet) · LLM relation detection | Accuracy: 0.932 -> 0.958 | Improved |
| bypass chat template for DNA extraction (ablation) | DNA (BGE) · LLM relation detection | Accuracy: 0.934 -> 0.945 | Improved |
| use a smaller sentence-embedding model (MPNet 0.1B, BGE 0.3B) instead of Qwen3-8B (ablation) | DNA · DNA structure (Mantel test) | Correlation: r>0.75, p<0.0001 | No clear effect |
| vary embedding output dimension p in {4,64,128,1024,4096} (ablation) | DNA (Qwen3-8B-Embedding) · LLM relation detection | Accuracy/AUC | No clear effect |
| increase DNA dimension L from 4 up to 128 (ablation) | DNA · LLM relation detection | Accuracy/AUC | Improved |
| restrict to Qwen2Tokenizer models (harder same-tokenizer setting) | DNA (MPNet) · LLM relation detection | Accuracy: 0.93 -> 0.86 | Hurt |
| same-tokenizer setting for PhyloLM | PhyloLM · LLM relation detection | Accuracy: 0.742 -> 0.755 | Improved |
| vary model parameter size of test pairs | DNA classifier · Size-vs-correctness analysis | Point-biserial correlation: -0.006 to 0.023, p 0.63 to 0.69 | No clear effect |
Closest papers by meaning
- PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks (ICLR 2025 · similarity 0.91)
- Language Modeling by Language Models (NeurIPS 2025 · similarity 0.89)
- FunctionEvolve: Structure-Guided Symbolic Regression with LLMs (NeurIPS 2026 · similarity 0.88)
- Measuring Scientific Capabilities of Language Models with a Systems Biology Dry Lab (NeurIPS 2025 · similarity 0.88)
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.87)
- Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks (ICML 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.
