This blog post introduces the first chapter of a new book on reasoning in Large Language Models (LLMs). It defines reasoning in the context of LLMs as the ability to produce intermediate steps before providing a final answer, distinguishing it from pattern matching. The chapter also reviews the conventional pre-training and post-training stages of LLMs. It further explains how LLMs learn from data through statistical associations and pattern recognition. The book will focus on practical, hands-on coding examples to directly implement reasoning techniques for LLMs.
Inference-Time Compute Scaling Methods to Improve Reasoning Models Part 1: Inference-Time Compute Scaling Methods
The latest research in 2025 focuses on improving LLM reasoning abilities through inference-time compute scaling. This approach enhances performance by increasing computational resources during inference, without modifying the underlying model weights. Methods range from simple token-based interventions to sophisticated search strategies, effectively trading compute for enhanced reasoning. A key trend is that even smaller models can achieve substantial improvements with the right inference strategy, narrowing the gap with larger models. This article summarizes recent papers and highlights the industry's move towards 'thinking on demand,' making reasoning an integral part of LLMs.
On the Biology of a Large Language Model Lindsey et al., 2025 We investigate the internal mechanisms used by Claude 3.5 Haiku — Anthropic\'s lightweight production model — in a variety of contexts.
The study investigates Claude 3.5 Haiku's internal mechanisms using circuit tracing to understand how it performs tasks like multi-hop reasoning and poetry writing. It uncovers language-specific and independent circuits, and explores generalization across different contexts, and identifies mechanisms for distinguishing familiar entities and handling harmful requests. The findings reveal sophisticated strategies, including planning, goal-directed reasoning, and metacognitive circuits. It provides insight into model's behavior and suggests a path towards safety auditing applications.
Circuit Tracing: Revealing Computational Graphs in Language Models Ameisen et al., 2025 We describe an approach to tracing the \"step-by-step\" computation involved when a model responds to a single prompt.
This blog post introduces a new method for understanding how language models work by tracing the computational steps they take. The method involves creating a "replacement model" that uses more interpretable components to mimic the original model. Attribution graphs are used to visualize the flow of information within the model, and perturbation experiments are conducted to validate the findings. This method is applied to Claude 3.5 Haiku, and the findings are used to investigate a diverse range of behaviors.
The "think" tool: Enabling Claude to stop and think in complex tool use situations
Anthropic introduces the 'think' tool, a simple yet effective method that significantly improves Claude's ability to handle complex tasks by creating a dedicated space for structured thinking. This tool enhances Claude's agentic tool use, policy adherence, decision-making consistency, and multi-step problem-solving with minimal overhead. Evaluations using \(τ\)-bench demonstrate remarkable improvements in customer service scenarios, particularly in navigating conversations, following policy guidelines, and utilizing various tools, with the best performance achieved by pairing the 'think' tool with optimized prompts that provide reasoning examples. The tool is most beneficial when Claude needs to process tool outputs carefully, follow detailed guidelines, and make sequential decisions, improving performance by up to 54% in certain domains.