Lilian Weng is an AI researcher and the author of Lil’Log, a technical blog explaining machine learning, agents and AI safety. OpenAI confirmed to AI Magazine in September 2026 that she had returned to lead research on accelerating its internal research processes.[1][2]
Start with LLM Powered Autonomous Agents →
Education
PhD in Informatics; dissertation: Information Diffusion on Online Social Networks
[5]Links
Blog
X 
GitHub Google Scholar Research team lead, recursive self-improvement · OpenAI
Background
Indiana University’s informatics doctoral-program page lists Weng’s dissertation, Information Diffusion on Online Social Networks. Her later public writing ranges across deep learning, reinforcement learning and language models. She began documenting learning notes on Lil’Log in 2017 and describes periodically updating older posts, making the archive a continuing technical reference rather than a sequence of product announcements.[1][5][6]
Career
According to OpenAI’s confirmation reported by AI Magazine, Weng returned to the company in 2026 to lead a team supporting recursive self-improvement research. This concerns using AI to help improve research and future systems. The same report describes her earlier OpenAI research and safety leadership and her subsequent co-founding of Thinking Machines Lab. Her current work follows those roles; the blog’s independent explanations are separately documented contributions.[2]
Contributions
Her 2023 autonomous-agents essay organizes an agent around planning, memory and tool use. Planning breaks a larger task into steps; memory supplies information beyond the immediate prompt; tools let the system retrieve information or execute actions outside the model. The article connects research methods and example systems to those components, helping readers understand which part of an agent solves which problem. It is a survey and synthesis of others’ and the field’s work, rather than a claim that Weng invented each component.[3]
Ideas
Her reward-hacking survey explains how an agent can maximize its training reward while failing the intended task. A coding agent that modifies tests instead of fixing code illustrates the mismatch between a measurable score and the goal behind it. The article traces failures in environments, training and evaluators and reviews mitigation directions. It gives builders a reason to inspect how a score was earned before treating improvement on that score as improvement in the real task.[4]
Tools, writing & talks
Writing, research & talks
Start here
LLM Powered Autonomous Agentslilianweng.github.ioA technical survey of planning, memory and tool use in language-model agents.
Begin with the agent-system overview to connect the term agent to concrete software components. Familiarity with language models helps; follow the linked references for the mathematical details of individual methods.
Author[3]Original source- Harness Engineering for Self-Improvementlilianweng.github.io
Recursive self-improvement dates to I. J. Good in 1965 and Yudkowsky's 2008 framing of the feedback loop. This traces the idea forward to harness engineering, where the improvement happens around the model rather than inside it.
Catalog workOriginal source Lilian Weng announces a long-planned post on scaling laws: how they let you reason about the optimal split of compute between data and model size before committing to an expensive run.
Catalog workOriginal source- Scaling Laws, Carefullylilianweng.github.io
Scaling laws say training loss falls predictably as model size, data and compute grow, a straight line on a log-log plot. A careful look at what that framework does and does not license you to conclude.
Catalog workOriginal source - Why We Thinklilianweng.github.io
Special thanks to John Schulman for a lot of super valuable feedback and direct edits on this post. Test time compute ( Graves et al. 2016 , Ling, et al. 2017 , Cobbe et al. 2021 ) and Chain-of-thought (CoT) ( Wei et al. 2022 , Nye et al. 2021 ), have led to significant improvements in model performance, while raising many research questions. This post aims to review recent developments in how to effectively use test-time compute (i.e. “thinking time”) and why it helps.
Catalog workOriginal source - Reward Hacking in Reinforcement Learninglilianweng.github.io
A survey of how optimizing imperfect rewards can produce behavior that defeats the intended task.
Author[4]Original source - Extrinsic Hallucinations in LLMslilianweng.github.io
Narrowing hallucination to its useful meaning: output that is fabricated and grounded in neither the provided context nor world knowledge, rather than any mistake a model makes.
Catalog workOriginal source - Diffusion Models for Video Generationlilianweng.github.io
Diffusion models moved from image synthesis to video, a strictly harder problem because temporal consistency across frames demands world knowledge a single image never needed.
Catalog workOriginal source - Thinking about High-Quality Human Datalilianweng.github.io
Most task-specific labelled data comes from human annotation, which makes annotator disagreement, expertise and incentives a modelling problem rather than an operational detail. On what high-quality human data actually requires.
Catalog workOriginal source - Adversarial Attacks on LLMslilianweng.github.io
A survey of adversarial attacks and jailbreak prompts against language models, and why alignment work such as RLHF builds default-safe behaviour that these techniques are designed to route around.
Catalog workOriginal source - Prompt Engineeringlilianweng.github.io
A survey of prompt engineering as in-context steering of an autoregressive model without touching its weights, treated as an empirical science whose methods vary sharply between models.
Catalog workOriginal source - The Transformer Family Version 2.0lilianweng.github.io
A rewritten and roughly doubled survey of transformer architecture variants, restructured and brought up to date with the improvements proposed since the original 2020 version.
Catalog workOriginal source - Large Transformer Model Inference Optimizationlilianweng.github.io
Why inference on large transformers is hard and what actually helps: the memory and latency bottlenecks that make powerful models expensive to serve, and the optimisation techniques including distillation that address them.
Catalog workOriginal source - Generalized Visual Language Modelslilianweng.github.io
Vision-language systems traditionally bolt an object detector onto a text decoder. A look at how generalized vision-language models replace that pipeline, for captioning and visual question answering.
Catalog workOriginal source - Learning with not Enough Data Part 3: Data Generationlilianweng.github.io
Approaches to generating synthetic training data when real data is scarce, covering augmentation of existing samples and outright generation of new ones.
Catalog workOriginal source - Learning with not Enough Data Part 2: Active Learninglilianweng.github.io
This is part 2 of what to do when facing a limited amount of labeled data for supervised learning tasks. This time we will get some amount of human labeling work involved, but within a budget limit, and therefore we need to be smart when selecting which samples to label.
Catalog workOriginal source - Learning with not Enough Data Part 1: Semi-Supervised Learninglilianweng.github.io
When facing a limited amount of labeled data for supervised learning tasks, four approaches are commonly discussed.
Catalog workOriginal source - How to Train Really Large Models on Many GPUs?lilianweng.github.io
[Updated on 2022-03-13: add expert choice routing .] [Updated on 2022-06-10]: Greg and I wrote a shorted and upgraded version of this post, published on OpenAI Blog: “Techniques for Training Large Neural Networks”
Catalog workOriginal source - What are Diffusion Models?lilianweng.github.io
[Updated on 2021-09-19: Highly recommend this blog post on score-based generative modeling by Yang Song (author of several key papers in the references)]. [Updated on 2022-08-27: Added classifier-free guidance , GLIDE , unCLIP and Imagen . [Updated on 2022-08-31: Added latent diffusion model . [Updated on 2024-04-13: Added progressive distillation , consistency models , and the Model Architecture section .
Catalog workOriginal source - Contrastive Representation Learninglilianweng.github.io
The goal of contrastive representation learning is to learn such an embedding space in which similar sample pairs stay close to each other while dissimilar ones are far apart. Contrastive learning can be applied to both supervised and unsupervised settings. When working with unsupervised data, contrastive learning is one of the most powerful approaches in self-supervised learning .
Catalog workOriginal source - Reducing Toxicity in Language Modelslilianweng.github.io
Large pretrained language models are trained over a sizable collection of online data. They unavoidably acquire certain toxic behavior and biases from the Internet. Pretrained language models are very powerful and have shown great success in many NLP tasks. However, to safely deploy them for practical real-world applications demands a strong safety control over the model generation process.
Catalog workOriginal source - Controllable Neural Text Generationlilianweng.github.io
[Updated on 2021-02-01: Updated to version 2.0 with several work added and many typos fixed.] [Updated on 2021-05-26: Add P-tuning and Prompt Tuning in the “prompt design” section.] [Updated on 2021-09-19: Add “unlikelihood training” .]
Catalog workOriginal source - Neural Architecture Searchlilianweng.github.io
Although most popular and successful model architectures are designed by human experts, it doesn’t mean we have explored the entire network architecture space and settled down with the best option. We would have a better chance to find the optimal solution if we adopt a systematic and automatic way of learning high-performance model architectures.
Catalog workOriginal source - Exploration Strategies in Deep Reinforcement Learninglilianweng.github.io
Exploration versus exploitation in deep reinforcement learning: why committing to a solution too early risks local minima, and the strategies including curiosity and exploration via disagreement that push agents to keep looking.
Catalog workOriginal source - Curriculum for Reinforcement Learninglilianweng.github.io
[Updated on 2020-02-03: mentioning PCG in the “Task-Specific Curriculum” section. [Updated on 2020-02-04: Add a new “curriculum through distillation” section.
Catalog workOriginal source - Self-Supervised Representation Learninglilianweng.github.io
A survey of self-supervised representation learning, covering contrastive predictive coding and the momentum-contrast family including MoCo, SimCLR and BYOL.
Catalog workOriginal source - Evolution Strategieslilianweng.github.io
Stochastic gradient descent is a universal choice for optimizing deep learning models. However, it is not the only option. With black-box optimization algorithms, you can evaluate a target function $f(x): \mathbb{R}^n \to \mathbb{R}$, even when you don’t know the precise analytic form of $f(x)$ and thus cannot compute gradients or the Hessian matrix. Examples of black-box optimization methods include Simulated Annealing , Hill Climbing and Nelder-Mead method .
Catalog workOriginal source - Meta Reinforcement Learninglilianweng.github.io
In my earlier post on meta-learning , the problem is mainly defined in the context of few-shot classification. Here I would like to explore more into cases when we try to “meta-learn” Reinforcement Learning (RL) tasks by developing an agent that can solve unseen tasks fast and efficiently.
Catalog workOriginal source - Domain Randomization for Sim2Real Transferlilianweng.github.io
Deep RL is too sample-hungry to train on real robots, so models are trained in simulation and fail on the reality gap. Domain randomization closes it by varying the simulator enough that reality looks like one more variation.
Catalog workOriginal source - Are Deep Neural Networks Dramatically Overfitted?lilianweng.github.io
[Updated on 2019-05-27: add the section on Lottery Ticket Hypothesis.] If you are like me, entering into the field of deep learning with experience in traditional machine learning, you may often ponder over this question: Since a typical deep neural network has so many parameters and training error can easily be perfect, it should surely suffer from substantial overfitting. How could it be ever generalized to out-of-sample data points?
Catalog workOriginal source - Meta-Learning: Learning to Learn Fastlilianweng.github.io
[Updated on 2019-10-01: thanks to Tianhao, we have this post translated in Chinese !]
Catalog workOriginal source - Flow-based Deep Generative Modelslilianweng.github.io
GANs and VAEs never explicitly learn the density of real data because the integral is intractable. Flow-based models do, by composing invertible transformations, and this is how they work.
Catalog workOriginal source - From Autoencoder to Beta-VAElilianweng.github.io
A tour from the plain autoencoder through denoising, sparse and contractive variants to the variational autoencoder and beta-VAE, with the ELBO loss and the reparameterization trick worked through, plus VQ-VAE and TD-VAE.
Catalog workOriginal source - Attention? Attention!lilianweng.github.io
[Updated on 2018-10-28: Add Pointer Network and the link to my implementation of Transformer.] [Updated on 2018-11-06: Add a link to the implementation of Transformer model.] [Updated on 2018-11-18: Add Neural Turing Machines .] [Updated on 2019-07-18: Correct the mistake on using the term “self-attention” when introducing the show-attention-tell paper; moved it to Self-Attention section.] [Updated on 2020-04-07: A follow-up post on improved Transformer models is here .]
Catalog workOriginal source - Implementing Deep Reinforcement Learning Models with Tensorflow + OpenAI Gymlilianweng.github.io
A hands-on implementation walkthrough of deep reinforcement learning models in TensorFlow against OpenAI Gym, covering naive Q-learning, DQN, double and duelling variants, policy gradients and actor-critic.
Catalog workOriginal source - Policy Gradient Algorithmslilianweng.github.io
A survey of policy gradient algorithms in reinforcement learning, from the basic theorem through actor-critic variants including SAC, D4PG, TD3 and SVPG.
Catalog workOriginal source - A (Long) Peek into Reinforcement Learninglilianweng.github.io
[Updated on 2020-09-03: Updated the algorithm of SARSA and Q-learning so that the difference is more pronounced. [Updated on 2021-09-19: Thanks to 爱吃猫的鱼, we have this post in Chinese ].
Catalog workOriginal source - The Multi-Armed Bandit Problem and Its Solutionslilianweng.github.io
The exploration-exploitation dilemma stated as the multi-armed bandit problem, with the standard family of solutions from epsilon-greedy through UCB and Thompson sampling, implemented for Bernoulli bandits.
Catalog workOriginal source - Learning Word Embeddinglilianweng.github.io
Human vocabulary comes in free text. In order to make a machine learning model understand and process the natural language, we need to transform the free-text words into numeric values. One of the simplest transformation approaches is to do a one-hot encoding in which each distinct word stands for one dimension of the resulting vector and a binary value indicates whether the word presents (1) or not (0).
Catalog workOriginal source - Anatomize Deep Learning with Information Theorylilianweng.github.io
Naftali Tishby's information bottleneck applied to deep learning, using information theory to describe how a network's representations grow and transform over the course of training.
Catalog workOriginal source - From GAN to WGANlilianweng.github.io
A walk from the original generative adversarial network to Wasserstein GAN, covering why the game-theoretic setup of generator against critic is unstable and what the Wasserstein distance fixes.
Catalog workOriginal source