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
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 causal attention over clean tokens too (Eso-LMs, alpha0=1) instead of MDLM bidirectional attention (ablation) | MDLM-style masked diffusion · LM1B NELBO PPL | perplexity: 31.78 -> 36.12 | Hurt |
| keep bidirectional attention among clean tokens, causal only on masks (Eso-LMs (A)) (ablation) | Eso-LMs alpha0=1 · LM1B and OWT perplexity | perplexity | Improved |
| lower alpha0 (more tokens generated autoregressively) during training (ablation) | Eso-LMs · LM1B exact PPL | perplexity: 31.65 -> 22.39 | Improved |
| add AdaLN to the AR baseline (ablation) | AR Transformer · LM1B PPL | perplexity: 22.83 -> 21.86 | Improved |
| estimate NELBO with single-sample any-order estimator L_AO instead of L_MDM | Eso-LMs NELBO estimation · OWT test sequence | standard deviation: 0.56 -> 0.03 | Improved |
| use causal attention with KV caching and restricted forward pass for sampling (Eso-LMs vs MDLM) | Eso-LMs alpha0=1 · OWT sampling latency, L=8192 | speedup: ~14x at L=2048, ~65x at L=8192 | Improved |
| use Eso-LMs instead of BD3-LMs for long-context sampling | Eso-LMs · OWT sampling latency, L=8192 | speedup: ~3.2x vs L'=16, ~3.8x vs L'=4 | Improved |
| reduce sampling steps for BD3-LMs (T=1024 to T=256) (ablation) | BD3-LM L'=16 · OWT unconditional generation | Gen. PPL: 72.80 -> 356.02 | Hurt |
| reduce sampling steps for MDLM (T=1024 to T=256) (ablation) | MDLM · OWT unconditional generation | Gen. PPL: 41.92 -> 45.07 | No clear effect |
| raise num_tries (quality-control resampling) from 1 to 10 (ablation) | BD3-LM L'=16 · OWT unconditional generation | entropy: 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 frontier | MAUVE vs sampling time | No clear effect |
| use the block sampler (parallel decode only far-apart tokens) instead of ancestral sampler (ablation) | Eso-LMs · OWT low-NFE generation | generation quality | Improved |
Closest papers by meaning
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.92)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.92)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.92)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.92)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.92)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
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.
