AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Esoteric Language Models: A Family of Any-Order Diffusion LLMs

Subham Sekhar Sahoo, Zhihan Yang, Yash Akhauri, Johnna Liu, Deepansha Singh, Zhoujun Cheng, Zhengzhong Liu, Eric P. Xing, John Thickstun, Arash Vahdat

ICML 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 0 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 causal attention over clean tokens too (Eso-LMs, alpha0=1) instead of MDLM bidirectional attention (ablation)MDLM-style masked diffusion · LM1B NELBO PPLperplexity: 31.78 -> 36.12Hurt
keep bidirectional attention among clean tokens, causal only on masks (Eso-LMs (A)) (ablation)Eso-LMs alpha0=1 · LM1B and OWT perplexityperplexityImproved
lower alpha0 (more tokens generated autoregressively) during training (ablation)Eso-LMs · LM1B exact PPLperplexity: 31.65 -> 22.39Improved
add AdaLN to the AR baseline (ablation)AR Transformer · LM1B PPLperplexity: 22.83 -> 21.86Improved
estimate NELBO with single-sample any-order estimator L_AO instead of L_MDMEso-LMs NELBO estimation · OWT test sequencestandard deviation: 0.56 -> 0.03Improved
use causal attention with KV caching and restricted forward pass for sampling (Eso-LMs vs MDLM)Eso-LMs alpha0=1 · OWT sampling latency, L=8192speedup: ~14x at L=2048, ~65x at L=8192Improved
use Eso-LMs instead of BD3-LMs for long-context samplingEso-LMs · OWT sampling latency, L=8192speedup: ~3.2x vs L'=16, ~3.8x vs L'=4Improved
reduce sampling steps for BD3-LMs (T=1024 to T=256) (ablation)BD3-LM L'=16 · OWT unconditional generationGen. PPL: 72.80 -> 356.02Hurt
reduce sampling steps for MDLM (T=1024 to T=256) (ablation)MDLM · OWT unconditional generationGen. PPL: 41.92 -> 45.07No clear effect
raise num_tries (quality-control resampling) from 1 to 10 (ablation)BD3-LM L'=16 · OWT unconditional generationentropy: 5.11 -> 5.28 (T=256)Improved
train with alpha0=1 only and vary alpha0 at evaluation instead of training several alphas (ablation)Eso-LMs · OWT speed-quality Pareto frontierMAUVE vs sampling timeNo clear effect
use the block sampler (parallel decode only far-apart tokens) instead of ancestral sampler (ablation)Eso-LMs · OWT low-NFE generationgeneration qualityImproved

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