Research map / Large language models / Language, safety and interpretability
Masks Can Be Distracting: On Context Comprehension in Diffusion Language Models
Julianna Piskorz, Cristina Pinneri, Alvaro Correia, Motasem Alfarra, Risheek Garrepalli, Christos Louizos
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 |
|---|---|---|---|
| append many extra mask tokens to the prompt | LLaDA-Base and LLaDA-Instruct · few-shot ICL tasks | accuracy: 23 and 27 percentage points drop for base and instruct | Hurt |
| append about 20 extra masks | Dream-Base and Dream-Instruct · few-shot ICL tasks | accuracy: 6 and 8 percentage points drop | Hurt |
| replace extra masks with repeated ' .' tokens (ablation) | LLaDA-Base and LLaDA-Instruct · few-shot ICL tasks | accuracy: up to 3 and 10 percentage points drop vs 23 and 27 for masks | Hurt |
| add iterative unmasking (40 steps, high confidence) at inference (ablation) | LLaDA with extra masks · few-shot ICL tasks | accuracy | Improved |
| use high-confidence instead of random unmasking order (ablation) | LLaDA with 40 decoding steps · few-shot ICL tasks | accuracy | Improved |
| add CE+TV mask-agnostic fine-tuning (LoRA) | LLaDA-Base-8B · few-shot ICL, 2/4/6 decoding steps | relative degradation (%): 15.5% -> 7.9% | Improved |
| add CE+TV mask-agnostic fine-tuning (LoRA) | LLaDA-Instruct-8B · few-shot ICL, 2/4/6 decoding steps | relative degradation (%): 27.5% -> 17.0% | Improved |
| add the TV loss term to the CE loss (ablation) | mask-agnostic fine-tuning of LLaDA · few-shot ICL tasks | accuracy robustness to extra masks | Improved |
| add mask-agnostic fine-tuning (CE+TV) | LLaDA · few-shot ICL tasks | locality bias | Improved |
| add MA-loss LoRA fine-tuning | LLaDA-Base-8B · wikitext infilling | ROUGE-L: 0.497 -> 0.486 | No clear effect |
| add MA-loss LoRA fine-tuning | LLaDA-Base-8B · lm1b infilling | Generative PPL: 6.058 -> 6.252 | No clear effect |
| add MA-loss LoRA fine-tuning | LLaDA-MoE-Base · lm1b infilling | Generative PPL: 6.241 -> 5.828 | No clear effect |
Closest papers by meaning
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.91)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.90)
- UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models (ICLR 2026 · similarity 0.89)
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- CORE: Context-Robust Remasking for Diffusion Language Models (ICML 2026 · similarity 0.89)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 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.
