Research map / Architectures and efficiency / Attention and transformers
Vegas: Self-Speculative Decoding with Verification-Guided Sparse Attention
Yikang Yue, Yuqi Xue, Jian Huang
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 |
|---|---|---|---|
| select KV by attention over all draft tokens instead of last accepted token (ablation) | Vegas · LongBench-v2 | acceptance probability across draft chain | Improved |
| include discarded draft tokens in KV selection (ablation) | Vegas · LongBench-v2, gamma=7 | avg accepted tokens per iteration: 3%--14% | Improved |
| use only first draft and bonus token for logits instead of all draft tokens (ablation) | Vegas · LongBench-v2, Qwen3-8B gamma=7 | avg accepted tokens: 6.13 -> 6.11 | No clear effect |
| use only first draft and bonus token for logits instead of all draft tokens (ablation) | Vegas · LongBench-v2, Qwen3-8B gamma=11 | avg accepted tokens: 8.91 -> 8.72 | No clear effect |
| use only first draft and bonus token for logits instead of all draft tokens (ablation) | Vegas · FA3 kernel, Qwen3-8B | logit collection overhead: 53% -> 5% | Improved |
| use only first draft and bonus token for logits instead of all draft tokens (ablation) | Vegas · FA3 kernel, gpt-oss-20b | logit collection overhead: 189% -> 37% | Improved |
| increase sparsity ratio (ablation) | Vegas · LongBench-v2 | accepted tokens per iteration | Mixed |
| use page size 1 instead of larger pages up to 16 (ablation) | Vegas sparse attention · FA3 kernel latency | latency: 197.6 vs 198.3 us (Qwen3-8B) | No clear effect |
| use verification-guided KV selection instead of MagicDec-Stream | Vegas · AIME25 and CodeElo | decoding throughput: 1.15x--1.23x | Improved |
| use Vegas instead of vanilla vLLM decoding | Vegas · AIME25 and CodeElo | decoding throughput: 1.25x--2.70x | Improved |
Closest papers by meaning
- QuantSpec: Self-Speculative Decoding with Hierarchical Quantized KV Cache (ICML 2025 · similarity 0.91)
- RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression (ICML 2025 · similarity 0.90)
- SEED: Self-Speculative Decoding via Implicit Encoder–Decoder (NeurIPS 2026 · similarity 0.89)
- ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference (ICML 2025 · similarity 0.89)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.89)
- Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing (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.
