Reasoning Model
A type of large language model that performs chain-of-thought reasoning before producing an answer. Such models are well-suited to mathematical derivation, code generation, and complex analysis, and typically consume five to twenty times more tokens than standard models.
Core Facts
Reasoning models perform better in information synthesis tasks requiring multi-step logical deduction, but have higher latency and should be selected appropriately based on task characteristics.
High-intensity reasoning configurations typically entail higher costs and longer response times, trading response speed for logical rigor.
Chain of Thought能显著提升模型在数学推理、逻辑判断等需要多步推导的任务上的表现
Timeline (last 90 days)
Between 2024 and 2025, reasoning models and general-purpose models were clearly two product lines, but by 2026 mainstream flagship models had basically all shifted to a unified architecture plus adaptive reasoning.
For reasoning models, prompts should be more concise, focusing only on the goal and requirements, and role-playing is not recommended for excessive use; for general-purpose models, reasoning steps such as COT need to be explicitly guided, and role-playing is very effective.
One of Maia 300's key upgrades is greater capability to handle the latest inference models and agentic tasks.
Reasoning models and linters excel at local algorithmic optimization and syntactic compliance, but limited by the local context window they cannot derive long-term product strategy or user empathy, nor detect unstated upstream system changes.
Taking reasoning models as an example, a single rollout may produce a chain of thought spanning thousands of tokens, and the total token generation volume of 100 million rollouts can exceed the scale of pretraining corpora by an order of magnitude.
Reasoning models exhibit overthinking at excessively high effort levels, causing them to overturn originally correct answers, and a similar pattern has been observed in Anthropic's Opus 5.
OpenAI's Codex is built on a reasoning model architecture; reasoning models perform extensive self-verification and step-by-step derivation before output, consuming far more tokens than traditional conversational models
Setting the sampling temperature to 0 causes reasoning models to fall into a thinking loop and become stuck during the thinking phase.
Reasoning models are extremely costly to train, typically requiring a combination of large-scale reinforcement learning and carefully designed reward models to acquire the capability.
The mainstream approaches to reducing inference costs fall into two categories: switching to a smaller base model and truncating/disabling the thinking process, where the former sacrifices capability and the latter sacrifices stability.
21 more timeline events
All Facts (20)
Reasoning models perform better in information synthesis tasks requiring multi-step logical deduction, but have higher latency and should be selected appropriately based on task characteristics.
90%VerifiedHigh-intensity reasoning configurations typically entail higher costs and longer response times, trading response speed for logical rigor.
90%VerifiedChain of Thought能显著提升模型在数学推理、逻辑判断等需要多步推导的任务上的表现
90%Verified推理能力基于思维链（Chain-of-Thought, CoT）技术，模型在给出最终答案前生成内部逐步推导过程
80%VerifiedReasoning models perform internal Chain-of-Thought before returning final results, consuming large amounts of tokens and computation time in exchange for deeper understanding and lower error rates.
75%VerifiedThe reasoning paradigm is essentially the application of reinforcement learning to the post-training stage of language models, allowing the model to autonomously explore reasoning paths and receive reward feedback based on answer correctness, first publicly demonstrated in the o1 series of models.
75%VerifiedThe internal chain of thought in reasoning models generates a large number of output tokens, making the actual cost of completing complex tasks often far higher than what a per-token price comparison would suggest.
70%VerifiedResearch has found that the chain-of-thought reasoning generated by models does not fully correspond to their actual internal computation processes, and may provide post-hoc explanations unrelated to the true decision-making path.
70%VerifiedChain-of-Thought（思维链）提示技术通过强制模型先输出中间步骤来部分缓解自回归模型在逻辑推理中的缺陷
65%VerifiedReasoning models (such as the o1 series) can cost several times to more than ten times more than regular models because they need to generate a large number of intermediate thinking steps.
65%Verified思考模型在输出前会经历内部链式推理（Chain-of-Thought）过程，在数学证明、代码调试、复杂逻辑推理等任务上表现优于标准模型，代价是延迟更高、计算成本更大
65%VerifiedReasoning tasks require a strong logical chain while summarization tasks demand concise language ability, and using the same model for both not only wastes computing power but may also not yield the best results.
65%UnverifiedPaths for improving model intelligence include increasing training and inference computation, optimizing reasoning architectures, and strengthening chain-of-thought training.
70%UnverifiedReasoning models have surpassed a large number of human experts in math competitions, programming, and logical reasoning.
60%UnverifiedModern reasoning models generally use RLHF or its variant RLAIF for alignment training, where the model learns during training when to stop reasoning and when to output an answer.
60%UnverifiedFor reasoning models, prompts should be more concise, focusing only on the goal and requirements, and role-playing is not recommended for excessive use; for general-purpose models, reasoning steps such as COT need to be explicitly guided, and role-playing is very effective.
50%UnverifiedOne of Maia 300's key upgrades is greater capability to handle the latest inference models and agentic tasks.
50%UnverifiedReasoning models and linters excel at local algorithmic optimization and syntactic compliance, but limited by the local context window they cannot derive long-term product strategy or user empathy, nor detect unstated upstream system changes.
50%UnverifiedTaking reasoning models as an example, a single rollout may produce a chain of thought spanning thousands of tokens, and the total token generation volume of 100 million rollouts can exceed the scale of pretraining corpora by an order of magnitude.
50%UnverifiedReasoning models exhibit overthinking at excessively high effort levels, causing them to overturn originally correct answers, and a similar pattern has been observed in Anthropic's Opus 5.
50%