Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Scheduling LLM Inference with Uncertainty-Aware Output Length Predictions
Haoyu Zheng, Yongqiang Zhang, Fangcheng Fu, Xiaokai Zhou, Hao Luo, Hongchao Zhu, Yuanyuan Zhu, Hao Wang, Xiao Yan, Jiawei Jiang
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 |
|---|---|---|---|
| model output length as a log-t distribution and schedule by Tail Inflated Expectation instead of point-estimate SJF (LTR) | LLM request scheduler · LMSYS-Chat-1M online, 100 RPS | average per-token latency: 2.31x reduction over LTR; 4.73x over FCFS | Improved |
| use TIE instead of LTR for offline synthetic data generation | LLM request scheduler · Alpaca SDG, 8B | time for 3k samples: 139.53 -> 98.12 | Improved |
| replace log-t with log-normal distribution (ablation) | TIE · LMSYS-Chat-1M online | average per-token latency: 0.67 -> 1.63 | Hurt |
| learn dynamic nu in log-t instead of fixing nu=3.5 (ablation) | TIE · LMSYS-Chat-1M online | average per-token latency: 0.67 -> 0.69 | No clear effect |
| add CVaR tail term to expected length score (ablation) | TIE · LMSYS-Chat-1M online | average per-token latency: 0.75 -> 0.72 (beta=0.1) | Improved |
| use adaptive beta instead of fixed beta (ablation) | TIE · LMSYS-Chat-1M online | average per-token latency: 0.71 (beta=0.3) -> 0.67 | Improved |
| use CVaR tail term in SDG (ablation) | TIE · Alpaca SDG | throughput within 3 min: 4257 -> 4762 | Improved |
| train predictor on Alpaca instead of LMSYS-Chat-1M (ablation) | TIE · Alpaca SDG, 8B | time for 3k samples: 98.12 -> 95.08 | Improved |
| transfer TIE to different datasets and a 70B model without retraining | TIE · ShareGPT/Alpaca, 70B | per-token latency: 3.77x over FCFS on 70B | Improved |
| risk-adaptive beta under rising load from 30 to 100 RPS | TIE · LMSYS-Chat-1M online | per-token latency growth: 7.42x/8.55x/6.17x -> 3.68x | Improved |
Closest papers by meaning
- Beyond Prediction: Tail-Aware Scheduling for LLM Inference (ICML 2026 · similarity 0.92)
- Efficient LLM Scheduling by Learning to Rank (NeurIPS 2024 · similarity 0.91)
- Geometry-Aware Online Scheduling for LLM Serving: From Theoretical Bound to System Practice (NeurIPS 2026 · similarity 0.89)
- Predicting LLM Output Length via Entropy-Guided Representations (ICLR 2026 · similarity 0.89)
- DON’T STOP ME NOW: EMBEDDING BASED SCHEDULING FOR LLMS (ICLR 2025 · similarity 0.86)
- Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios (AAAI 2026 · similarity 0.86)
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.
