Daily ML news, summarized.
The State of Reinforcement Learning for LLM Reasoning explores how reasoning capabilities in large language models are evolving through reinforcement learning. OpenAI's o3 reasoning model, trained with 10× more compute than o1, demonstrates significant progress using RL methods, highlighting that scaling alone is insufficient and targeted RL training is essential for improving model accuracy on complex tasks.
Ahead of AI · Source
A comprehensive 15-hour course teaches how to build large language models from scratch, covering topics from text data preparation to instruction fine-tuning. The content includes seven detailed videos and a bonus overview of LLM evolution from 2018 to 2025, originally developed as supplementary material for a book and now offered as standalone learning resources.
Ahead of AI · Source
The author shares a curated list of LLM research papers from January to June 2025, organized by topic such as reasoning models, reinforcement learning, and multimodal systems. The list includes 30 papers, with a strong focus on training and inference-time reasoning strategies, and is released in bi-annual updates to remain timely and digestible.
Ahead of AI · Source
A KV cache stores key and value vectors from previous token generations to avoid redundant computations during inference, significantly speeding up text generation in large language models. The article provides a from-scratch, human-readable implementation of the KV cache, explaining its conceptual basis and code structure, while noting it increases memory usage and is not applicable during training.
Ahead of AI · Source
A comprehensive overview of ten major open-weight large language model releases from January to February 2026 is presented, highlighting architectural features such as Mixture-of-Experts, sliding window attention, and parameter efficiency. Key models include Arcee AI's Trinity Large (400B params with 13B active), z.AI's GLM-5, and Cohere's Tiny Aya, with comparisons emphasizing design similarities and innovations in attention mechanisms and scalability.
Ahead of AI · Source
The article examines key architectural advancements in modern large language models, focusing on DeepSeek V3 and its subsequent reasoning model DeepSeek R1. It highlights two innovations: Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE), which improve computational efficiency and distinguish DeepSeek V3 from other LLMs.
Ahead of AI · Source
This article provides a hands-on, code-based implementation of Qwen3 models in pure PyTorch. It highlights Qwen3's popularity due to its developer-friendly Apache 2.0 license, strong performance on benchmarks like LMArena, and a wide range of model sizes from 0.6B to 480B parameters.
Ahead of AI · Source
OpenAI has released gpt-oss-120b and gpt-oss-20b, their first fully open-weight large language models since GPT-2 in 2019. The models are designed with architectural optimizations, including MXFP4, enabling local execution on GPUs, with the 20B model running on consumer hardware and the 120B model on an H100 or newer.
Ahead of AI · Source
The four main approaches to evaluating large language models are multiple-choice question answering, verifiers, leaderboards, and LLM judges. These methods fall into benchmark-based and judgment-based categories, with multiple-choice benchmarks like MMLU being historically prominent. The article provides a clear overview of each method and includes from-scratch code examples to illustrate their implementation and limitations.
Ahead of AI · Source
While most large language models remain autoregressive decoder-style transformers, recent years have seen the rise of alternatives such as text diffusion models and linear attention hybrids. These approaches aim to improve efficiency or modeling performance, with some like code world models targeting specific use cases. The article highlights a range of open-weight models and notes that non-traditional architectures deserve deeper exploration in future coverage.
Ahead of AI · Source
DeepSeek released V3.2, an updated flagship model featuring sparse attention and reinforcement learning improvements. The model demonstrates strong performance comparable to GPT-5 and Gemini 3.0 Pro, and is available as an open-weight model, expanding access to advanced language capabilities.
Ahead of AI · Source
A curated list of research papers on large language models from July to December 2025 is presented, categorized by topics such as reasoning models, reinforcement learning, inference strategies, model releases, architectures, and multimodal systems. The list was created as a separate post to improve readability and accessibility, alongside the publication of the annual "State of LLMs 2025" review article.
Ahead of AI · Source
The article categorizes inference-time scaling techniques to improve large language model reasoning without altering model weights. It covers methods like Chain-of-Thought Prompting, Self-Consistency, Best-of-N Ranking, Rejection Sampling, Self-Refinement, and Search Over Solution Paths, with a focus on practical implementations and recent academic work.
Ahead of AI · Source
DeepSeek R1, released in January 2025, introduced Reinforcement Learning with Verifiable Rewards (RLVR) using the GRPO algorithm to enable reasoning in large language models. The open-weight model performed comparably to top proprietary models and suggested training costs of around $5 million—significantly lower than previous estimates—though it does not account for research and development expenses.
Ahead of AI · Source
The article presents a visual gallery of 45 LLM architectures with model cards, covering recent attention variants used in open-source models. It includes a poster version available on Redbubble, designed for print with clear visuals, though smaller text elements may be hard to read at reduced sizes.
Ahead of AI · Source
The author outlines a workflow for reverse-engineering LLM architectures by starting with technical papers and then using available config files and reference implementations from Hugging Face and the transformers library. This approach is effective for open-weight models, as working code provides concrete, verifiable details that papers often lack.
Ahead of AI · Source
The article outlines six key components of a coding agent, emphasizing the role of agent harnesses in managing code context, tool use, memory, and iterative feedback. It distinguishes between LLMs, reasoning models, and agents, explaining that agents operate as control loops using models and tools, with coding harnesses being specialized for software engineering tasks.
Ahead of AI · Source
Newer large language model architectures emphasize long-context efficiency by reducing KV-cache size and attention costs. Key innovations include KV sharing in Gemma 4, layer-wise attention budgeting in Laguna XS.2, compressed convolutional attention in ZAYA1-8B, and mHC with compressed attention in DeepSeek V4.
Ahead of AI · Source
A curated list of research papers from January to May 2026 focuses on reasoning models, reinforcement learning, agent systems, and long-context capabilities. It highlights hybrid architectures like Nemotron 3 and Qwen3.6, with emphasis on efficiency and practical applications in real-world agent systems.
Ahead of AI · Source