Research map / Large language models / Language, safety and interpretability
Less is Enough: Synthesizing Diverse Data in LLM Feature Space with Sparse Autoencoders
Zhongzhi Li, Xuansheng Wu, Yijiang Li, Lijie Hu, Ninghao Liu
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 FAC Synthesis instead of the strongest synthesis baseline (SynAlign) | Llama-3.1-8B-Instruct · Toxicity Detection | AUPRC (%): 58.83 -> 62.60 | Improved |
| use FAC Synthesis instead of the best baseline on instruction following | Llama-3.1-8B-Instruct · AlpacaEval 2 | LC win rate (%): 11.26 -> 20.27 | Improved |
| increase FAC (task-relevant feature coverage) of the synthetic data | synthesis methods · Toxicity Detection | AUPRC (%): Pearson r = 0.95, Spearman 0.90 | Improved |
| cover a larger share of the missing features (30% to 100%) (ablation) | FAC Synthesis · Toxicity Detection | AUPRC (%) | Improved |
| raise sample count to a fixed N=200 at the same feature coverage (ablation) | FAC Synthesis · Toxicity Detection | AUPRC (%): +0.50% | No clear effect |
| use two-step contrastive-pair synthesis instead of one-step direct prompting (ablation) | FAC Synthesis · Toxicity Detection | FAC (%) | Improved |
| share synthetic data built from Llama-3.1-8B-Instruct features to fine-tune other model families | Mistral-7B-Instruct · Toxicity Detection | AUPRC (%): 27.66 -> 47.23 | Improved |
| share synthetic data built from Llama-3.1-8B-Instruct features to fine-tune Qwen2-7B-Instruct | Qwen2-7B-Instruct · Toxicity Detection | AUPRC (%): 51.44 -> 68.20 | Improved |
| take SAE features from Llama-3.1-8B-Instruct instead of Qwen2's own SAE features (ablation) | Qwen2-7B-Instruct downstream · Toxicity Detection | AUPRC (%): +1.60% to +5.13% | Improved |
| use Llama-3.1-8B-Instruct as generator instead of GPT-4o mini (ablation) | FAC Synthesis · Toxicity Detection | AUPRC (%): 44.90 -> 49.12 at temperature 0.8 | Improved |
| move decoding temperature away from the intermediate value 0.8 (ablation) | FAC Synthesis · Toxicity Detection | AUPRC (%): 49.12 (T=0.8) vs 46.71 (T=0.4), 46.40 (T=1.2) | Hurt |
| raise SAE activation threshold from 1.0 to 2.0 (ablation) | FAC Synthesis · Toxicity Detection | AUPRC (%): 44.20 -> 45.98 | Improved |
Closest papers by meaning
- Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data (NeurIPS 2025 · similarity 0.90)
- Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.88)
- Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling (NeurIPS 2026 · similarity 0.87)
- It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.87)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.86)
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.
