Reasoning

Reasoning is an LLM capability involving multi-step deduction, analysis, or planning to arrive at an answer - for example breaking a problem into parts, evaluating constraints, or sequencing actions.

Details

Some models and inference APIs expose reasoning as a separate output stream (a reasoning trace or "thinking" block) produced before the final answer, while in other models reasoning occurs implicitly within the generation process. Exposed reasoning traces are not necessarily faithful representations of the model's internal process and may be filtered or summarized by the provider. Reasoning capabilities can be developed through reinforcement learning with verifiable rewards, where models learn chain-of-thought-style strategies by being rewarded for producing correct answers.

Reasoning tokens add to inference cost and latency, particularly for models that produce extended reasoning traces before the final answer. In agent systems, reasoning is the capability exercised at each step of the agent loop - evaluating the current context, selecting a tool call, and determining when a final response is ready. Planning is an application of reasoning focused specifically on structuring multi-step execution: decomposing a task into steps and determining their order. Because reasoning governs an agent's planning and action selection, it is a target for goal manipulation attacks that attempt to redirect the agent's objectives.

Examples

• Multi-step arithmetic or unit conversion (for example "If A costs X and B costs Y, what is the total?")
• Following a set of constraints to derive a consistent answer (logic puzzles, scheduling)
• Choosing and sequencing tool calls in an agent
• Returning a separate reasoning trace alongside a final answer in an inference API

Synonyms

thinking