VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
u_Unrolled

Research papers

Read the papers that built AI.

A fresh paper every week, the 100 Days 100 Research Paper Challenge, and 42 latest and trending papers sorted by domain. Tick them off as you go.
Week of 9 November 2026

VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Chou et al.NVIDIA2026

Most-upvoted paper on Hugging Face Daily Papers this week

Robot and driving models plan with points, boxes and trajectories. VFlash predicts each shape’s coordinates together with a small drafter, instead of one token at a time, so visual planning gets much faster without losing accuracy.

Your progress
Saved in this browser
0 / 141 read0%

100 Days 100 Research Paper Challenge to Learn AI Engineer from Scratch

One landmark paper a day, from transformers and ResNets to agents and SWE-bench.

0 / 99 read0%
Up next: Day 1
  1. DAY 1Attention Is All You Need
  2. DAY 2Deep Residual Learning for Image Recognition
  3. DAY 3BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
  4. DAY 4Language Models are Few-Shot Learners
  5. DAY 5Highly accurate protein structure prediction with AlphaFold
  6. DAY 6Denoising Diffusion Probabilistic Models
  7. DAY 7Training language models to follow instructions with human feedback
  8. DAY 8An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
  9. DAY 9Learning Transferable Visual Models From Natural Language Supervision
  10. DAY 10Evaluating Large Language Models Trained on Code
  11. Days 11–100: 89 more papers with Unrolled Pro.Unlock the challenge
By domain7 domains

Latest and Trending Research Papers, Categorised by Domain

Each paper is linked to the lesson that teaches it, where there is one.

0 / 42 read0%
DOMAIN 01 Foundations: from words to the transformer
  1. 01
    Efficient Estimation of Word Representations in Vector Space
    Mikolov et al. · Google · 2013

    Words as vectors learned from context: the idea embeddings still rest on.

  2. 02
    Attention Is All You Need
    Vaswani et al. · Google · 2017

    The transformer. Read it twice: once now, once after building one.

  3. 03
    Improving Language Understanding by Generative Pre-Training
    Radford et al. · OpenAI · 2018

    GPT-1: pretrain a decoder on raw text, then fine-tune. The recipe begins here.

  4. 04
    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
    Devlin et al. · Google · 2018

    The encoder side: masked-language modelling and why it suits understanding tasks.

  5. 05
    Language Models are Unsupervised Multitask Learners
    Radford et al. · OpenAI · 2019

    GPT-2: scale plus diverse data gives zero-shot task ability.

  6. 06
    Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
    Raffel et al. · Google · 2019

    T5: a careful, controlled study of what actually matters in pretraining.

  7. 07
    RoFormer: Enhanced Transformer with Rotary Position Embedding
    Su et al. · Academia · 2021

    RoPE, the position encoding used by most open models today.

DOMAIN 02 Scaling and pretraining
  1. 08
    Language Models are Few-Shot Learners
    Brown et al. · OpenAI · 2020

    GPT-3 and in-context learning: prompting instead of fine-tuning.

  2. 09
    Scaling Laws for Neural Language Models
    Kaplan et al. · OpenAI · 2020

    Loss falls as a smooth power law in parameters, data and compute.

  3. 10
    Training Compute-Optimal Large Language Models
    Hoffmann et al. · Google DeepMind · 2022

    Chinchilla: most models were under-trained; about 20 tokens per parameter.

  4. 11
    LLaMA: Open and Efficient Foundation Language Models
    Touvron et al. · Meta · 2023

    A readable, complete recipe for a strong open model on public data.

  5. 12
    Llama 2: Open Foundation and Fine-Tuned Chat Models
    Touvron et al. · Meta · 2023

    Pretraining plus the full chat fine-tuning and RLHF pipeline, in detail.

  6. 13
    The Llama 3 Herd of Models
    Grattafiori et al. · Meta · 2024

    A frontier-scale training report: data, infrastructure, failures and evals.

DOMAIN 03 Post-training and alignment
  1. 14
    Fine-Tuning Language Models from Human Preferences
    Ziegler et al. · OpenAI · 2019

    The first RLHF on language models: reward model plus PPO.

  2. 15
    Learning to summarize from human feedback
    Stiennon et al. · OpenAI · 2020

    RLHF beats supervised fine-tuning, and over-optimizing the reward hurts.

  3. 16
    Training language models to follow instructions with human feedback
    Ouyang et al. · OpenAI · 2022

    InstructGPT: SFT, reward model, PPO. The ChatGPT recipe.

  4. 17
    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
    Bai et al. · Anthropic · 2022

    Helpfulness vs harmlessness trade-offs, measured.

  5. 18
    Constitutional AI: Harmlessness from AI Feedback
    Bai et al. · Anthropic · 2022

    Replace most human labels with a written constitution and AI feedback.

  6. 19
    LoRA: Low-Rank Adaptation of Large Language Models
    Hu et al. · Microsoft · 2021

    Fine-tune a huge model by training two small matrices.

  7. 20
    QLoRA: Efficient Finetuning of Quantized LLMs
    Dettmers et al. · Academia · 2023

    LoRA on a 4-bit base model: fine-tuning on a single GPU.

  8. 21
    Direct Preference Optimization: Your Language Model is Secretly a Reward Model
    Rafailov et al. · Academia · 2023

    Preference tuning with one simple loss, no reward model or RL loop.

DOMAIN 04 Reasoning, tools and agents
  1. 22
    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
    Wei et al. · Google · 2022

    Writing out intermediate steps unlocks multi-step reasoning.

  2. 23
    ReAct: Synergizing Reasoning and Acting in Language Models
    Yao et al. · Academia · 2022

    Interleave thoughts and tool calls: the core agent loop.

  3. 24
    Toolformer: Language Models Can Teach Themselves to Use Tools
    Schick et al. · Meta · 2023

    A model learns when and how to call APIs from self-labelled data.

  4. 25
    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
    Lewis et al. · Meta · 2020

    The original RAG: retrieve passages, then generate from them.

  5. 26
    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
    Snell et al. · Google DeepMind · 2024

    When thinking longer beats a bigger model, and when it does not.

  6. 27
    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
    Shao et al. · DeepSeek · 2024

    Introduces GRPO: PPO without a value model.

  7. 28
    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
    DeepSeek-AI · DeepSeek · 2025

    Reasoning that emerges from RL with verifiable rewards, then distilled.

DOMAIN 05 Efficient architectures and inference
  1. 29
    FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
    Dao et al. · Academia · 2022

    Exact attention, much faster, by respecting the GPU memory hierarchy.

  2. 30
    Efficient Memory Management for Large Language Model Serving with PagedAttention
    Kwon et al. · Academia · 2023

    vLLM: OS-style paging for the KV cache.

  3. 31
    GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
    Frantar et al. · Academia · 2022

    3–4-bit weights with little accuracy loss, in hours.

  4. 32
    Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
    Fedus et al. · Google · 2021

    Mixture-of-experts made simple: route each token to one expert.

  5. 33
    Mixtral of Experts
    Jiang et al. · Mistral AI · 2024

    A strong open sparse MoE: big capacity, small active compute.

  6. 34
    DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
    DeepSeek-AI · DeepSeek · 2024

    Multi-head latent attention: a much smaller KV cache.

  7. 35
    DeepSeek-V3 Technical Report
    DeepSeek-AI · DeepSeek · 2024

    Frontier quality on a modest budget: FP8 training, MoE, multi-token prediction.

  8. 36
    Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
    Yuan et al. · DeepSeek · 2025

    Sparse attention designed for the hardware and trained end to end.

  9. 37
    Mamba: Linear-Time Sequence Modeling with Selective State Spaces
    Gu & Dao · Academia · 2023

    A serious alternative to attention that scales linearly with length.

DOMAIN 06 World models and JEPA
  1. 38
    A Path Towards Autonomous Machine Intelligence
    LeCun · Meta · 2022

    The position paper behind JEPA: world models, energy-based learning, planning.

  2. 39
    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture
    Assran et al. · Meta · 2023

    I-JEPA: predict the representation of hidden image regions, not their pixels.

  3. 40
    Revisiting Feature Prediction for Learning Visual Representations from Video
    Bardes et al. · Meta · 2024

    V-JEPA: the same idea on video, learning motion and physics.

  4. 41
    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
    Assran et al. · Meta · 2025

    A video world model that robots can plan with.

DOMAIN 07 Looking inside the model
  1. 42
    Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
    Templeton et al. · Anthropic · 2024

    Sparse autoencoders find millions of human-readable features in a production model.

AI research papers: paper of the week and 100-day challenge · Unrolled