Research map / Image generation / Fast diffusion sampling
Self-Augmenting Retrieval for Diffusion Language Models
Paul Jünger, Justin Lovelace, Linxi Zhao, Dongyoung Go, Kilian Q Weinberger
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 SARDI self-augmenting retrieval to DREAM-7B | DREAM-7B · 2WikiMultihopQA | EM: 43.7 -> 59.1 | Improved |
| add SARDI self-augmenting retrieval to DREAM-7B | DREAM-7B · MuSiQue | EM: 11.1 -> 20.6 | Improved |
| add SARDI to DREAM-7B | DREAM-7B · HotpotQA | EM: 39.9 -> 48.7 | Improved |
| add SARDI to DREAM-7B | DREAM-7B · CofCA | EM: 43.4 -> 45.3 | Improved |
| lower query threshold tau_q to 0 (expose all speculative tokens to retriever) (ablation) | SARDI · 2WikiMultihopQA | EM | Improved |
| add SARDI to DREAM-7B on compositional questions | DREAM-7B · 2Wiki compositional questions | EM: 16.6 -> 45.3 | Improved |
| add SARDI to DREAM-7B on inference questions | DREAM-7B · 2Wiki inference questions | EM: 14.0 -> 37.5 | Improved |
| add SARDI on comparison (single-hop) questions | DREAM-7B · 2Wiki comparison questions | EM: 85.6 -> 84.6 | No clear effect |
| remove gold documents from context (ablation) | DREAM-7B · 2WikiMultiHopQA entity token pairs | conditional mutual information: 0.060 -> 0.588 | Hurt |
| refresh retrieval every 2 steps instead of every step (ablation) | SARDI · 2WikiMultihopQA | EM: 58 -> 56 | Hurt |
| refresh retrieval every 4 steps instead of every step (ablation) | SARDI · 2WikiMultihopQA | EM: 58 -> 52 | Hurt |
| replace BM25 with dense E5-base-v2 retriever (ablation) | SARDI · 2WikiMultihopQA | EM: 59.1 -> 50 | Mixed |
Closest papers by meaning
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Support Before Frequency in Discrete Diffusion (NeurIPS 2026 · similarity 0.90)
- InstructRAG: Instructing Retrieval-Augmented Generation via Self-Synthesized Rationales (ICLR 2025 · similarity 0.89)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.89)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.89)
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · 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.
