Research map / Image generation / Fast diffusion sampling
Stop the Flip-Flop: Context-Preserving Verification for Fast Revocable Diffusion Decoding
Yanzheng Xiang, Lan Wei, Yizhen Yao, Qinglin Zhu, Hanqi Yan, Chen Jin, Philip Alexander Teare, Dandan Zhang, Lin Gui, Amrutha Saseendran, Yulan He
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 KV cache override with diagonal correction (ablation) | COVER · GSM8K (LLaDA-Ins-8B, len 256) | decoding steps: 123.28 -> 51.65 | Improved |
| add KV cache override with diagonal correction (ablation) | COVER · GSM8K (LLaDA-Ins-8B, len 256) | relative speed: 0.63x -> 1.00x | Improved |
| add KV cache override with diagonal correction (ablation) | COVER · HumanEval (LLaDA-Ins-8B, len 256) | accuracy: 39.63 -> 40.24 | Improved |
| add KV cache override with diagonal correction (ablation) | COVER · MATH500 (LLaDA-Ins-8B, len 256) | accuracy: 30.60 -> 32.80 | Improved |
| add stability-aware seed selection (ablation) | COVER · GSM8K (LLaDA-Ins-8B, len 256) | decoding steps: 65.10 -> 51.65 | Improved |
| add stability-aware seed selection (ablation) | COVER · HumanEval (LLaDA-Ins-8B, len 256) | decoding steps: 102.14 -> 96.16 | Improved |
| add stability-aware seed selection (ablation) | COVER · HumanEval (LLaDA-Ins-8B, len 256) | accuracy: 39.02 -> 40.24 | Improved |
| use COVER instead of greedy baseline | LLaDA-Ins-8B · GSM8K (len 512) | relative speed: 1.00x -> 5.52x | Improved |
| use COVER instead of greedy baseline | Dream-Ins-7B · MBPP (len 512) | relative speed: 1.00x -> 11.64x | Improved |
| use COVER instead of greedy baseline | LLaDA-Base-8B · HumanEval (len 512) | accuracy: 32.93 -> 35.37 | Improved |
| use COVER instead of WINO | LLaDA-Ins-8B · MBPP (len 256) | accuracy: 36.20 -> 39.00 | Improved |
| add Fast-dLLM prefix cache | COVER · GSM8K (LLaDA-Ins-8B) | relative speed: 1.00x -> 1.23x | Improved |
Closest papers by meaning
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- CORE: Context-Robust Remasking for Diffusion Language Models (ICML 2026 · similarity 0.90)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.89)
- dKV-Cache: The Cache for Diffusion Language Models (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.
