Research map / Image generation / Fast diffusion sampling
Dual Language Models: Balancing sample-efficiency and overfitting resilience
David Samuel, Lucas Georges Gabriel Charpentier
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 |
|---|---|---|---|
| add masked-diffusion objective (α=1/8 instead of pure AR) (ablation) | autoregressive LM · 128 repetitions, avg 9 tasks | normalized AR score: 9.4 -> 19.1 | Improved |
| add masked-diffusion objective (α=3/4) (ablation) | autoregressive LM · 32 repetitions, avg 9 tasks | normalized AR score: 22.0 -> 23.9 | Improved |
| add masked-diffusion objective (α=63/64) (ablation) | autoregressive LM · 1 repetition, avg 9 tasks | normalized AR score: 26.1 -> 26.9 | Improved |
| add autoregressive objective (α=3/4) (ablation) | masked-diffusion LM · 32 repetitions, PLL eval | normalized PLL score: 21.4 -> 25.3 | Improved |
| add autoregressive objective (α=1/8) (ablation) | masked-diffusion LM · 128 repetitions, PLL eval | normalized PLL score: 21.0 -> 23.0 | Improved |
| evaluate dual model with prefix attention mask instead of causal | dual-objective LM · 1 repetition | normalized score: 26.9 -> 27.9 | Improved |
| evaluate dual model with prefix attention mask instead of causal | dual-objective LM · 32 repetitions | normalized score: 23.9 -> 25.0 | Improved |
| evaluate dual model with prefix attention mask instead of causal | dual-objective LM · 128 repetitions | normalized score: 19.1 -> 20.5 | Improved |
| set α so autoregressive objective sees about 16 repetitions in data-constrained training | dual-objective LM · more than 32 repetitions | downstream performance | Improved |
| use pure autoregressive training in regular-data regime (16 repetitions or fewer) (ablation) | autoregressive LM · regular data | AR performance | No clear effect |
| use 6 mask tokens instead of 1 in PLL (ablation) | PLL evaluation · BLiMP and HellaSwag | normalized PLL score: 65.2 -> 60.3 | Hurt |
| use 6 mask tokens instead of 1 in PLL (ablation) | PLL evaluation · ARC-Easy and PIQA | normalized PLL score: 18.7 -> 26.3 | Improved |
Closest papers by meaning
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.91)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.91)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.90)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (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.
