Research map / Large language models / Language, safety and interpretability
DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders
Xu Wang, Bingqing Jiang, Yu Wan, Baosong Yang, Lingpeng Kong, Difan Zou
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 diffusion-time steering with Dream-Unmask SAE instead of Qwen-2.5-7B LLM steering | SAE feature steering · layer L27, L0=80 | steering score S: -0.02 -> 0.38 | Improved |
| use Dream-Mask SAE steering instead of Qwen SAE steering at deep layer | SAE feature steering · layer L23, L0=80 | steering score S: 0.15 -> 0.24 | Improved |
| use LLaDA-Unmask SAE steering instead of LLaMA-3-8B | SAE feature steering · layer L1, L0=80 | steering score S: -2.29 -> 0.13 | Improved |
| use LLaDA-Unmask instead of LLaMA-3-8B at layer L27 | SAE feature steering · layer L27, L0=80 | steering score S: 0.10 -> 0.08 | No clear effect |
| use Qwen-2.5-7B instead of Dream-Unmask at layer L14 | SAE feature steering · layer L14, L0=80 | steering score S: 0.14 -> 0.01 | Hurt |
| insert SAE into early layers of DLM instead of LLM | SAE reconstruction splicing · Dream-7B vs Qwen-2.5-7B, small L0 | delta LM loss | Improved |
| use confidence-based decoding order TopK-margin instead of Origin | Dream-7B decoding · GSM8K | accuracy: 8% -> 56% | Improved |
| use Entropy decoding order instead of Origin | Dream-7B decoding · GSM8K | accuracy: 8% -> 59% | Improved |
| apply Base SAE instead of SFT SAE on Dream SFT at layers L1-L23 | Dream SFT · Dream SFT SAE insertion | delta LM loss | No clear effect |
| apply Base SAE instead of SFT SAE on Dream SFT at layer L27 | Dream SFT · Dream SFT SAE insertion | delta LM loss | Hurt |
| steer only update (masked) tokens versus all tokens (LLaDA uses update, Dream all) (ablation) | SAE steering · Dream-7B Mask-SAE | qualitative steering | Mixed |
| increase Mask-SAE position selector from top-5 to top-15 (ablation) | Dream-7B steering · Dream-7B Mask-SAE | qualitative steering | Mixed |
Closest papers by meaning
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.90)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.90)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.90)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.89)
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.
