Research map / Architectures and efficiency / RNNs and normalization
WAVE: Window-Aware Vocabulary-Efficient Early-Exit for Training-Free LLM Acceleration
Seonggeun Kim, Gilha lee, Hyun Kim
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 |
|---|---|---|---|
| restrict exit decisions to a calibrated window and use a proxy top-K LM head | LLaMA-2 7B Chat · MMLU/CSQA/SST2 | end-to-end speedup: up to 1.57x on SST2; 1.4x average | Improved |
| use WAVE instead of AdaInfer | AdaInfer · SST2 exit overhead | fraction of inference time (%): 34.1 -> 2.8 | Improved |
| use proxy LM head with K=4096 instead of full LM head (ablation) | WAVE · exit-decision cost | FLOPs: about 87% reduction | Improved |
| calibrate on GSM8K but evaluate on other domains | WAVE · MMLU, CSQA, SST2, BoolQ, MRPC, RTE | accuracy: -0.64 to +1.72 %p | No clear effect |
| increase proxy LM head size K from 64 to 4096 (ablation) | WAVE · CSQA | accuracy | Improved |
| vary exit window size W among 4, 8, 12 (ablation) | WAVE · MMLU/CSQA/SST2 | accuracy | No clear effect |
| apply WAVE to batched vLLM serving | WAVE · vLLM batched inference | throughput: average 1.13x speedup | Improved |
| apply WAVE to deeper or different models | WAVE · LLaMA-2-13B MMLU | accuracy (%): 53.58 -> 52.53 | Hurt |
| apply WAVE to Mistral-7B | WAVE · Mistral-7B BoolQ | accuracy (%): 86.02 -> 81.65 | Hurt |
| apply WAVE under AWQ quantization instead of SpecEE | SpecEE · SST2 AWQ W4A16 | end-to-end speedup: 1.28x -> 1.49x | Improved |
Closest papers by meaning
- EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism (ICML 2024 · similarity 0.92)
- ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization (NeurIPS 2024 · similarity 0.90)
- La RoSA: Enhancing LLM Efficiency via Layerwise Rotated Sparse Activation (ICML 2025 · similarity 0.90)
- Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads (ICML 2024 · similarity 0.90)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.89)
- Progressive Mixed-Precision Decoding for Efficient LLM Inference (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.
