Reasoning LLMs (2026) | Chain-of-Thought, RL with Verifiable Rewards, GRPO, Test-Time Compute, Open Models & Benchmarks

What Are Reasoning LLMs?

6 entries

Reasoning LLMs generate a long, explicit chain of intermediate steps before giving a final answer, and spend more computation at inference time on harder problems. The modern recipe combines chain-of-thought prompting and its descendants, training signals from verifiers or process reward models, reinforcement learning on tasks with checkable answers such as maths and code, and test-time compute scaling through longer reasoning, sampling or search. They excel at maths, coding and multi-step planning, but cost more, can overthink simple questions, and do not always report their real reasoning. The readable guides below are the fastest way in.

01Understanding Reasoning LLMs (Sebastian Raschka)Clear overview of the main ways reasoning models are built. 02A Visual Guide to Reasoning LLMs (Maarten Grootendorst)Illustrated walk-through of test-time compute and training. 03Why We Think (Lilian Weng)Survey of thinking time, test-time compute and chain-of-thought. 04DeepSeek R1's recipe to replicate o1 (Interconnects)Analysis of the R1 training recipe. 05Chain-of-Thought Prompting (Prompt Engineering Guide)Practical guide to chain-of-thought prompting. 06Large Reasoning Models (this site)Side-by-side comparison of 17 reasoning models, methods and benchmarks.

Foundations: Prompting for Reasoning

8 entries

The ideas that preceded and underpin reasoning models.

07Show Your Work: Scratchpads for Intermediate ComputationEarly work on writing out intermediate steps. 08Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsThe paper that started chain-of-thought prompting. 09Large Language Models are Zero-Shot Reasoners"Let's think step by step" without examples. 10Self-Consistency Improves Chain of Thought ReasoningSample many reasoning paths and take the majority answer. 11Least-to-Most PromptingBreak problems into easier subproblems. 12Tree of ThoughtsSearch over branching reasoning paths. 13ReAct: Synergizing Reasoning and ActingInterleave reasoning with tool use. 14Towards Reasoning in Large Language Models: A SurveyEarly survey of reasoning techniques.

Training Reasoning: Verifiers, Rewards & Reinforcement Learning

15 entries

How models learn to reason rather than merely imitate.

15STaR: Bootstrapping Reasoning With ReasoningLearn from self-generated rationales that reach correct answers. 16Training Verifiers to Solve Math Word Problems (GSM8K)Outcome verifiers and the GSM8K dataset. 17Let's Verify Step by StepProcess supervision beats outcome supervision for maths. 18Math-ShepherdProcess reward models without human step labels. 19The Lessons of Developing Process Reward Models in Mathematical ReasoningPractical lessons and pitfalls with PRMs. 20DeepSeekMathIntroduced GRPO, a critic-free RL algorithm widely reused. 21DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningOpen-weight model trained with large-scale RL on verifiable rewards. 22Kimi k1.5: Scaling Reinforcement Learning with LLMsRL recipe with long-context scaling. 23Tulu 3: Pushing Frontiers in Open Language Model Post-TrainingOpen post-training recipe including RL with verifiable rewards. 24DAPO: An Open-Source LLM Reinforcement Learning System at ScaleOpen RL system with training-stability techniques. 25Demystifying Long Chain-of-Thought Reasoning in LLMsWhat makes long reasoning emerge and how to shape it. 26LLM Post-Training: A Deep Dive into Reasoning Large Language ModelsSurvey of post-training methods for reasoning. 27Quiet-STaRModels learn to think before speaking at every token. 28Training Large Language Models to Reason in a Continuous Latent SpaceReasoning in hidden states instead of words (Coconut). 29Let's Think Dot by DotHidden computation in filler tokens.

Test-Time Compute & Efficient Reasoning

8 entries

Spending inference compute where it helps.

30Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model ParametersShows how allocating inference compute can beat bigger models. 31s1: Simple test-time scalingBudget forcing and a small curated reasoning dataset. 32s1 on GitHubCode, data and models for s1. 33From System 1 to System 2: A Survey of Reasoning Large Language ModelsSurvey of the shift to deliberate reasoning. 34Towards Reasoning Era: A Survey of Long Chain-of-ThoughtSurvey of long chain-of-thought reasoning. 35Stop Overthinking: A Survey on Efficient Reasoning for LLMsMethods to shorten reasoning without losing accuracy. 36vLLM reasoning outputsServing reasoning models and parsing their thinking content. 37SGLang documentationServing engine with reasoning-model support.

Reasoning Models & Provider Documentation

14 entries

Official guides for the main model families.

38Learning to reason with LLMs (OpenAI)OpenAI's announcement of its first reasoning model; the site may block automated checks. 39Reasoning models (OpenAI API)Using OpenAI reasoning models and effort settings. 40Extended thinking (Anthropic)Claude's thinking budget and thinking blocks. 41Gemini thinkingGemini thinking models and budgets. 42DeepSeek-R1 on GitHubModel release and distilled variants. 43DeepSeek-R1 on Hugging FaceModel weights. 44Qwen3: Think Deeper, Act FasterHybrid thinking and non-thinking models. 45Qwen3 Technical ReportTraining details for the Qwen3 family. 46Qwen3 on GitHubCode and usage. 47gpt-oss model cardOpenAI open-weight reasoning models. 48gpt-oss on GitHubReference implementations. 49gpt-oss-120b on Hugging FaceModel weights. 50Kimi K2Open-weight mixture-of-experts model from Moonshot AI. 51OlmoFully open models from Ai2 with open data and training code.

Open Training Stacks, Libraries & Datasets

13 entries

Reproduce and fine-tune reasoning models.

52Open-R1Open reproduction of the DeepSeek-R1 pipeline. 53Open-R1 announcementGoals and plan of the Open-R1 project. 54Open R1 for Students (Hugging Face course)Course chapter on reasoning-model training. 55TRLHugging Face library for post-training with RL. 56TRL GRPO trainerGuide to training with GRPO. 57verlFlexible RL training library for LLMs. 58OpenRLHFScalable RLHF and reasoning RL framework. 59open-instruct (Ai2)Ai2's post-training code, including RL with verifiable rewards. 60OpenThoughtsOpen reasoning datasets and models. 61OpenThoughts on GitHubData recipes and training code. 62OpenR1-Math-220kReasoning traces for mathematics. 63NuminaMath 1.5Competition-level maths problems and solutions. 64PRM800KStep-level correctness labels for maths solutions.

Benchmarks & Evaluation

12 entries

Measuring reasoning, and its limits.

65MATH datasetCompetition mathematics benchmark. 66AIME problems (Art of Problem Solving)Source of the AIME problems used as a benchmark; the site may block automated checks. 67GPQAGraduate-level science questions. 68MMLU-ProHarder multi-task knowledge and reasoning benchmark. 69LiveCodeBenchContamination-free code reasoning benchmark. 70Humanity's Last ExamExpert-level multi-domain benchmark. 71FrontierMathResearch-level mathematics problems. 72SWE-benchReal GitHub issues for coding agents. 73ARC PrizeAbstraction and reasoning challenge. 74Epoch AI benchmarksIndependent benchmark tracking. 75simple-evals (OpenAI)Lightweight evaluation scripts. 76lm-evaluation-harnessStandard open evaluation framework.

Limits, Faithfulness & Safety

5 entries

Where reasoning models fall short.

77Do NOT Think That Much for 2+3=?Documents overthinking on simple questions. 78The Illusion of ThinkingStudies how reasoning models behave as puzzle complexity rises; its conclusions have been debated. 79Reasoning Models Don't Always Say What They ThinkChain-of-thought is often not a faithful account of the model's reasoning. 80Reasoning models don't always say what they think (Anthropic)Summary of the faithfulness research. 81Chain of Thought MonitorabilityWhy monitoring reasoning traces is useful for safety and why it may be fragile.

9 entries

Neighbouring pages on this site.

82Large Language ModelsLLM hub. 83LLM Inference EnginesServing engines such as vLLM and SGLang. 84Causal LLMsCausal reasoning in LLMs and agents. 85LLM-as-a-Judge SolutionsEvaluating model outputs. 86Reinforcement Learning AlgorithmsRL algorithms behind reasoning training. 87Agentic AIAgentic AI hub. 88Loop vs Graph EngineeringAgent control flow. 89Language ModelsLanguage model resources. 90Artificial IntelligenceAI hub.

Frequently Asked Questions

What is a reasoning LLM?

A reasoning LLM generates an explicit chain of intermediate steps, often long, before its final answer, and is typically trained with reinforcement learning on tasks whose answers can be checked, such as maths and code. It spends more compute at inference time on harder problems.

How are reasoning models trained?

Common ingredients are supervised fine-tuning on reasoning traces, reward signals from outcome verifiers or process reward models, and reinforcement learning with verifiable rewards using algorithms such as GRPO. DeepSeek-R1, Kimi k1.5, Tulu 3 and Qwen3 describe open recipes.

What is test-time compute scaling?

It means improving answers by using more computation at inference, for example longer chains of thought, sampling many answers and voting, or search guided by a verifier. Research such as Snell et al. and s1 shows it can rival scaling model size on some tasks.

What are the main limitations?

Reasoning models cost more tokens and time, can overthink simple questions, may produce chains of thought that do not reflect their actual computation, and still fail on some complex problems. Benchmark scores also date quickly, so test on your own tasks.

How current is this page?

Links were checked in October 2026. Reasoning models change quickly: model names, prices and benchmark scores date fast, and many papers are arXiv preprints. For a side-by-side model comparison see our Large Reasoning Models page. Sites that block automated checks are kept and may need a manual click.

Reasoning LLMs: Chain-of-Thought, RL Training & Test-Time Compute