Research map / Architectures and efficiency / Attention and transformers
FOCUS: DLLMs Know How to Tame Their Compute Bound
Kaihua Liang, Xin Tan, An Zhong, Hong Xu, Marco Canini
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 FOCUS token eviction (alpha 1.5) at confidence 0.7 | SDAR-8B-Chat · Math500 | accuracy: 54.60 -> 62.20 | Improved |
| add FOCUS token eviction at confidence 0.9 | SDAR-8B-Chat · Math500 | accuracy: 64.70 -> 64.30 | No clear effect |
| add FOCUS token eviction at confidence 0.9 | SDAR-8B-Chat · GSM8K | accuracy: 89.20 -> 90.15 | Improved |
| add FOCUS token eviction vs LMDeploy | SDAR-8B-Chat · ShareGPT batch 256 | throughput (tokens/s): 979 -> 2272 | Improved |
| use block size 64 instead of 32 | FOCUS · ShareGPT batch 256 | speedup over LMDeploy: 2.32x -> 3.52x | Improved |
| add FOCUS to LLaDA2.0-mini (MoE) | LLaDA2.0-mini · ShareGPT batch 256 | throughput (tokens/s): 2181 -> 3369 | Improved |
| add FOCUS on LLaDA2.0 MATH at batch 64 | LLaDA2.0-mini · MATH batch 64 | throughput | Hurt |
| select tokens by top importance delta instead of random (ablation) | FOCUS · GSM8K (K=2) | accuracy: 40.38 -> 90.33 | Improved |
| select tokens by top importance delta instead of bottom (ablation) | FOCUS · Math500 (K=2) | accuracy: 4.40 -> 65.40 | Improved |
| add FOCUS eviction on top of delayed intra-block cache (ablation) | SDAR-8B-Chat · MBPP | accuracy: 52.14 -> 59.73 | Improved |
| use standard delayed cache for intra-block KV (ablation) | SDAR-8B-Chat · GSM8K | accuracy: 89.20 -> 84.92 | Hurt |
| use neighbor-aware stability criterion instead of standard delayed cache (ablation) | SDAR-8B-Chat · GSM8K | accuracy: 84.92 -> 88.02 | Improved |
Closest papers by meaning
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.92)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.91)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.90)
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.
