AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Linguistic encoding in LLM representations · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use diffusion-time steering with Dream-Unmask SAE instead of Qwen-2.5-7B LLM steeringSAE feature steering · layer L27, L0=80steering score S: -0.02 -> 0.38Improved
use Dream-Mask SAE steering instead of Qwen SAE steering at deep layerSAE feature steering · layer L23, L0=80steering score S: 0.15 -> 0.24Improved
use LLaDA-Unmask SAE steering instead of LLaMA-3-8BSAE feature steering · layer L1, L0=80steering score S: -2.29 -> 0.13Improved
use LLaDA-Unmask instead of LLaMA-3-8B at layer L27SAE feature steering · layer L27, L0=80steering score S: 0.10 -> 0.08No clear effect
use Qwen-2.5-7B instead of Dream-Unmask at layer L14SAE feature steering · layer L14, L0=80steering score S: 0.14 -> 0.01Hurt
insert SAE into early layers of DLM instead of LLMSAE reconstruction splicing · Dream-7B vs Qwen-2.5-7B, small L0delta LM lossImproved
use confidence-based decoding order TopK-margin instead of OriginDream-7B decoding · GSM8Kaccuracy: 8% -> 56%Improved
use Entropy decoding order instead of OriginDream-7B decoding · GSM8Kaccuracy: 8% -> 59%Improved
apply Base SAE instead of SFT SAE on Dream SFT at layers L1-L23Dream SFT · Dream SFT SAE insertiondelta LM lossNo clear effect
apply Base SAE instead of SFT SAE on Dream SFT at layer L27Dream SFT · Dream SFT SAE insertiondelta LM lossHurt
steer only update (masked) tokens versus all tokens (LLaDA uses update, Dream all) (ablation)SAE steering · Dream-7B Mask-SAEqualitative steeringMixed
increase Mask-SAE position selector from top-5 to top-15 (ablation)Dream-7B steering · Dream-7B Mask-SAEqualitative steeringMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy