Reporting and perspectives from Lilian Weng. Headlines and excerpts link to the original articles.
Lilian Weng published a post on harness engineering for recursive self-improvement, covering harness design patterns, context engineering methods such as ACE and MCE, automated workflow search approaches including ADAS and AFlow, and Meta-Harness. The post also traces the concept of recursive self-improvement to I. J. Good (1965) and Yudkowsky (2008).
A technical article by Lilian Weng reviews the history and methodology of neural scaling laws, tracing them from early learning-curve work through Kaplan et al. (2020), the Chinchilla scaling laws, and efforts to reconcile the two. It also covers scaling in data-limited regimes and practical difficulties in fitting scaling laws.
Lilian Weng published a review of test-time compute and chain-of-thought reasoning, covering parallel sampling, sequential revision, reinforcement learning, and continuous-space thinking. She credits John Schulman with feedback and edits.
In a new post, Lilian Weng surveys reward hacking in reinforcement learning, where agents exploit flaws or ambiguities in reward functions to score highly without completing the intended task. She catalogues examples across RL, language model and real-world settings, and calls for more research on practical mitigations, especially for RLHF and LLMs.
Lilian Weng published an article on extrinsic hallucinations in large language models, defining them as fabricated outputs not grounded in pre-training data or world knowledge. The post covers causes, detection methods, and anti-hallucination techniques, citing research including Gekhman et al. 2024, FactualityPrompt, FActScore, SAFE, SelfCheckGPT, TruthfulQA, and SelfAware.
Lilian Weng's blog post explains how diffusion models are being extended from image synthesis to video generation. It covers training video diffusion models from scratch, including parameterization, sampling, 3D U-Net and DiT architectures, and adapting pre-trained image models to video.
Lilian Weng published an article surveying methods for collecting and evaluating high-quality human-annotated data for machine learning, covering rater agreement, disagreement, and annotation aggregation techniques.
Lilian Weng published an overview post on adversarial attacks against large language models, covering threat models, attack classification, and specific techniques such as token manipulation, gradient-based attacks, and jailbreak prompting.
Lilian Weng published an overview of LLM-powered autonomous agent systems, describing planning, memory, and tool use as the key components, with proof-of-concept demos including AutoGPT, GPT-Engineer, and BabyAGI as examples.
Lilian Weng published a post explaining prompt engineering, also called in-context prompting, as methods to steer LLM behavior without updating model weights. The post covers zero-shot and few-shot prompting, example selection and ordering, instruction prompting, self-consistency sampling, chain-of-thought prompting, and automatic prompt design.
Lilian Weng published version 2.0 of her "The Transformer Family" post, a refactored and expanded update to her 2020 article. The new version restructures the section hierarchy, adds more recent papers, and is described as a superset of the old version at about twice its length.
A technical blog post by Lilian Weng provides an overview of methods for optimizing inference of large transformer models, covering distillation, quantization, pruning, sparsity, mixture-of-experts, and architecture-specific improvements. It explains challenges such as high memory usage, including KV cache memory, and reviews techniques to reduce memory footprint, computation, and latency.
Lilian Weng published a technical blog post titled "Some Math behind Neural Tangent Kernel," providing a deep dive into the motivation, definition, and proofs behind neural tangent kernel (NTK) theory. The post covers NTK's connection to Gaussian processes and the proof of deterministic convergence for infinite-width networks.
Lilian Weng's post surveys one approach to vision language models: extending pre-trained language models to consume visual signals. It groups methods into four buckets and describes techniques such as jointly training image and text, learned image embeddings as frozen LM prefixes, cross-attention fusion, and training-free decoding guided by vision-based scores.
Lilian Weng published Part 3 of her "Learning with not Enough Data" series, covering synthetic data generation for training. The post details data augmentation methods for images, text, audio, and architecture, plus data synthesis using pretrained language models as noisy annotators or data generators.
This article, part two of a series on learning with limited labeled data, covers active learning methods that select samples for human labeling under a budget. Topics include acquisition functions, uncertainty and diversity sampling, deep acquisition functions, and representativeness measures.
Lilian Weng published Part 1 of a series on learning with insufficient data, covering semi-supervised learning, which combines labeled and unlabeled data. The post surveys four general approaches and specific consistency regularization, pseudo labeling and self-training techniques, drawing mostly on vision research.
Lilian Weng's article surveys techniques for training very large neural networks across many GPUs, covering data, model, pipeline, and tensor parallelism, mixture-of-experts layers, and memory-saving designs. It notes updates adding expert choice routing and a related OpenAI Blog post.
Lilian Weng's blog post "What are Diffusion Models?" explains diffusion-based generative models, covering forward and reverse diffusion processes, connections to score networks and Langevin dynamics, guidance methods, and later updates adding latent diffusion, progressive distillation, and consistency models.
Lilian Weng published a technical overview of contrastive representation learning, covering training objectives such as contrastive loss, triplet loss, N-pair loss, NCE and InfoNCE, plus common setups and key ingredients including heavy data augmentation, large batch sizes and hard negative mining.