LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding
Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to propose candidate tokens that are then verified in parallel by the target model. The speedup is significantly determined by the acceptance rate, yet standard training minimizes Kullback-Leibler (KL) divergence as a proxy objective. While KL divergence and acceptance rate share the same global optimum, small draft models, having limited capacity, typically converge to suboptimal solutions where minimizing KL does not guarantee maximizing acceptance rate. To address this issue, we propose LK losses, special training objectives that directly target acceptance rate. Compr...
| Authors | Alexander Samarin, Sergei Krutikov, Anton Shevtsov et al. |
|---|---|
| Primary category | cs.LG |
| Published | 27 February 2026 |
| Last updated at source | 1 June 2026 |
Models and datasets referencing this paper
- nebius/EAGLE3-gpt-oss-20bModel
- nebius/EAGLE3-Llama-3.1-8B-InstructModel
- nebius/EAGLE3-Llama-3.3-70B-InstructModel
- nebius/DeepSeek-V3-Infinity-Instruct-0625Dataset
- nebius/EAGLE3-Qwen3-235B-A22B-Instruct-2507Model
- nebius/gpt-oss-120b-Infinity-Instruct-0625Dataset
- EntityDeletr/EAGLE3-gpt-oss-20b-GGUFModel
- nebius/Qwen3-235B-Instruct-Infinity-Instruct-0625Dataset
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.