AI Paper Atlas

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

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Transformer block and depth design · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
use buffer K=16 instead of full re-encoding AR (TNP-D-AR)TNP-D-AR · GP functionsaverage predictive density: 2.57 -> 2.51No clear effect
use buffer K=16 instead of TNP-D-ARTNP-D-AR · EEG forecastingaverage predictive density: 1.07 -> 0.85Hurt
deploy with K=1 (empty buffer) (ablation)TNP w/ buffer · EEG forecastingaverage predictive density: 0.85 -> 1.21Improved
add causal buffer instead of independent decodingTNP-D-Ind · EEG forecastingaverage predictive density: -0.74 -> 0.85Improved
use buffer instead of TNP-A on sawtoothTNP-A · Sawtoothaverage predictive density: 0.98 -> 1.00No clear effect
use buffer instead of TNP-A on EEG forecastingTNP-A · EEG forecastingaverage predictive density: 1.23 -> 0.85Hurt
use buffer instead of TNP-A for model selectionTNP-A · Multisensory model selectionLML RMSE: 4.75 -> 3.56Improved
use buffer instead of TNP-D-AR for model selectionTNP-D-AR · Multisensory model selectionLML RMSE: 3.10 -> 3.56Hurt
add buffer instead of independent predictions with TabICLIndependent · TabICL Jena forecasting (N=16)log-predictive density: -4.53 -> -2.15Improved
add buffer instead of standard AR with TabICLStandard AR · TabICL Electric Consumption interpolation (N=16)log-predictive density: 0.88 -> 0.78Hurt
add buffer with TabICL at high contextIndependent · TabICL Electric Consumption interpolation (N=1024)log-predictive density: 1.78 -> 1.79No clear effect
use buffer instead of TNP-D-AR for sampling speedTNP-D-AR · autoregressive samplingwall-clock time: 3-20x fasterImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy