Research map / Architectures and efficiency / Attention and transformers
Efficient Autoregressive Inference for Transformer Probabilistic Models
Conor Hassan, Nasrulloh Ratu Bagus Satrio Loka, Cen-You Li, Daolang Huang, Paul Edmund Chang, Yang Yang, Francesco Silvestrin, Samuel Kaski, Luigi Acerbi
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 |
|---|---|---|---|
| use buffer K=16 instead of full re-encoding AR (TNP-D-AR) | TNP-D-AR · GP functions | average predictive density: 2.57 -> 2.51 | No clear effect |
| use buffer K=16 instead of TNP-D-AR | TNP-D-AR · EEG forecasting | average predictive density: 1.07 -> 0.85 | Hurt |
| deploy with K=1 (empty buffer) (ablation) | TNP w/ buffer · EEG forecasting | average predictive density: 0.85 -> 1.21 | Improved |
| add causal buffer instead of independent decoding | TNP-D-Ind · EEG forecasting | average predictive density: -0.74 -> 0.85 | Improved |
| use buffer instead of TNP-A on sawtooth | TNP-A · Sawtooth | average predictive density: 0.98 -> 1.00 | No clear effect |
| use buffer instead of TNP-A on EEG forecasting | TNP-A · EEG forecasting | average predictive density: 1.23 -> 0.85 | Hurt |
| use buffer instead of TNP-A for model selection | TNP-A · Multisensory model selection | LML RMSE: 4.75 -> 3.56 | Improved |
| use buffer instead of TNP-D-AR for model selection | TNP-D-AR · Multisensory model selection | LML RMSE: 3.10 -> 3.56 | Hurt |
| add buffer instead of independent predictions with TabICL | Independent · TabICL Jena forecasting (N=16) | log-predictive density: -4.53 -> -2.15 | Improved |
| add buffer instead of standard AR with TabICL | Standard AR · TabICL Electric Consumption interpolation (N=16) | log-predictive density: 0.88 -> 0.78 | Hurt |
| add buffer with TabICL at high context | Independent · TabICL Electric Consumption interpolation (N=1024) | log-predictive density: 1.78 -> 1.79 | No clear effect |
| use buffer instead of TNP-D-AR for sampling speed | TNP-D-AR · autoregressive sampling | wall-clock time: 3-20x faster | Improved |
Closest papers by meaning
- Autoregressive Learning in Joint KL: Sharp Oracle Bounds and Lower Bounds (NeurIPS 2026 · similarity 0.86)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 2025 · similarity 0.86)
- Towards Understanding the Universality of Transformers for Next-Token Prediction (ICLR 2025 · similarity 0.85)
- TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · similarity 0.85)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.85)
- On Mesa-Optimization in Autoregressively Trained Transformers: Emergence and Capability (NeurIPS 2024 · similarity 0.85)
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.
