Natural Language Processing (NLP)

Understand and generate human language.

3 papers

Written by Junkun Yuan.

Click here to go back to main contents.

Table of contents:

Papers are displayed in reverse chronological order. High-impact or inspiring works are highlighted in red.

Foundation Algorithms & Models

Large Language Diffusion Models

Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, Chongxuan Li

Renmin University of China, Ant Group

arXiv, 2025

Feb 14, 2025   |   LLaDA   |   code

It introduces a masked diffusion language model (8B) that matches strong autoregressive LLMs while inherently enabling bidirectional reasoning.

  • It argues that generative modeling is to learn \(\max_{\theta}\mathbb{E}_{p_{data}(x)}\log p_{\theta}(x)\), which need not be auto-regressive.
  • Instruction-following and in-context learning are also not exclusive advantages of autoregressive models.
  • Auto-regressive models have disadvantages such as high computational costs due to token-by-token generation, and limitations in reversal reasoning due to left-to-right generation.
  • LLaDA (8B) with 4096 tokens is pre-trained from scratch on 2.3T tokens using 0.13M H800 GPU hours, followed by SFT on 4.5M pairs.
  • LLaDA does not use a causal mask, as it sees the entire context.
  • LLaDA uses vanilla multi-head attention, as it is incompatible with KV cache.
Figure 1. Pre-training. Tokens are independently randomly masked by probability of \(t\sim U[0,1]\), the model predicts the masked tokens by minimizing the cross-entropy loss. SFT. Only response tokens are possibly masked. Inference. Stimulate a diffusion process from \(t=1\) to \(t=0\).

Attention Is All You Need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin

Google Brain, Google Research, University of Toronto

Advances in Neural Information Processing Systems (NeurIPS), 2017

Jun 12, 2017   |   Transformer

It revolutionized deep learning by introducing the Transformer architecture, which replaced recurrence with self-attention, enabling massively parallel training and becoming the foundational model for virtually all modern large-scale language systems. It has 192,000 citations (as of Sep 2025).

It introduces sequence transduction architecture relying solely on multi-head self-attention, dramatically reducing training time.

(see notes in jupyter)

  • Details to be added

Reinforcement Learning

Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, Chelsea Finn

Stanford University, CZ Biohub

Advances in Neural Information Processing Systems (NeurIPS), 2023

May 29, 2023   |   DPO

It offers a simple, RL-free recipe to turn human preference data into aligned language models with equal or better performance than RLHF while eliminating reward-model training and heavy hyper-parameter tuning overhead. It has over 5,000 citations (as of Sep 2025).

It introduces DPO, a single-stage, RL-free algorithm that directly optimizes a language model on preference data by reparameterizing the Bradley-Terry objective into a simple classification loss.

Last updated on May 18, 2026 at 10:47 (UTC-7).