Research map / Large language models / Language, safety and interpretability
LongLLaDA: Unlocking Long Context Capabilities in Diffusion LLMs
Xiaoran Liu, Yuerong Song, Zhigeng Liu, Zengfeng Huang, Qipeng Guo, Ziwei He, Xipeng Qiu
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 |
|---|---|---|---|
| apply NTK RoPE scaling (λ=4) to LLaDA at 8k LongBench | LLaDA-8B-Instruct · LongBench 8k | Avg score: 36.8 -> 40.6 | Improved |
| apply NTK scaling (λ=4) to LLaDA at 4k LongBench | LLaDA-8B-Base · LongBench 4k | Avg score: 34.1 -> 34.0 | No clear effect |
| raise NTK factor from λ=4 to λ=14 at 8k (ablation) | LLaDA-8B-Base · RULER 8k | Avg score: 84.7 -> 86.8 | Improved |
| raise NTK factor from λ=14 to λ=31 at 16k (ablation) | LLaDA-8B-Base · RULER 16k | Avg score: 72.0 -> 78.0 | Improved |
| use NTK extrapolation instead of direct extrapolation at 16k | LLaDA-8B-Base · RULER 16k | Avg score: 19.5 -> 44.1 (λ=4) | Improved |
| set λ=31 (up to 24k) and λ=55 (32k) in NIAH (ablation) | LongLLaDA · NIAH | retrieval | Mixed |
| increase diffusion sampling steps from 1 or 4 to 8 or more (ablation) | LLaDA-8B-Base · NIAH | retrieval depth: fail beyond 8k at s=1,4; 25% depth at 16k and almost 10% at 24k for s>=8 | Improved |
| replace autoregressive LLaMA3 with diffusion LLaDA (bidirectional training) under direct extrapolation | 8B models · NIAH beyond pretrain length | retrieval: 54.25 vs 51.71 overall score (Base) | Improved |
| use dLLM instead of AR LLM on RULER aggregation tasks | LLaDA-8B vs LLaMA3-8B · RULER 4k AGG | AGG score: LLaDA-8B-Base 65.2 vs LLaMA3-8B-Base 98.1 | Hurt |
| use dLLM instead of AR LLM on RULER QA tasks | LLaDA-8B-Instruct vs LLaMA3-8B-Instruct · RULER 4k QA | QA score: 90.5 vs 68.5 | Improved |
Closest papers by meaning
- UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models (ICLR 2026 · similarity 0.95)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- Large Language Diffusion Models (NeurIPS 2025 · similarity 0.90)
- SemanticDLM+: Improving Diffusion LLMs through Bias-variance Trade-off in Transition Kernel Design (NeurIPS 2026 · similarity 0.90)
- What is Wrong with Perplexity for Long-context Language Modeling? (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.
