Temperature
A parameter that controls the randomness of an AI model’s outputs. A lower temperature setting produces more consistent and deterministic responses, while a higher setting produces more varied and creative outputs. In commercial real estate applications, lower temperatures are generally preferred for analytical tasks such as underwriting, lease review, and financial modeling, where precision and repeatability are more valuable than creativity.
Putting Temperature in Context
A CRE technology team building an AI assistant for lease abstraction sets the model temperature to a low value, ensuring that when the tool extracts rent escalation clauses, expiration dates, and tenant improvement allowances from a lease document, it returns consistent and repeatable results rather than paraphrasing key terms differently on each run, which would introduce noise into the downstream data pipeline feeding the asset management platform.
Frequently Asked Questions about Temperature
What temperature setting should I use for CRE analytical tasks?
For tasks like underwriting support, lease review, financial summary generation, or data extraction from offering memoranda, a temperature of 0 or close to 0 is generally appropriate. At that setting the model selects the highest-probability response each time, which means two identical inputs will produce nearly identical outputs and reviewers can treat the tool’s behavior as consistent and auditable rather than unpredictable.
Are there CRE tasks where a higher temperature is actually useful?
Higher temperature settings can be useful for tasks that benefit from varied phrasing or creative output, such as drafting multiple versions of a property marketing description, brainstorming deal narrative angles for an offering memorandum, or generating diverse investor update language for A/B testing. In those cases, a moderate temperature in the range of 0.5 to 0.8 introduces enough variation to be useful without producing outputs that are incoherent or factually unstable.
How is temperature different from other parameters like top-p or max tokens?
Temperature controls how the model samples from its probability distribution when selecting each word, while top-p controls which portion of that distribution is even considered as a candidate. Max tokens is a separate parameter entirely and simply caps the length of the output. In practice, temperature is the most commonly adjusted parameter for CRE use cases because it has the most direct effect on whether outputs feel precise and repeatable or varied and exploratory.
What are the risks of using a high temperature setting in a CRE workflow?
At higher temperature settings, the model is more likely to introduce phrasing, figures, or interpretations that deviate from the source material, which creates meaningful risk when the output is feeding into a rent roll summary, a DSCR calculation narrative, or a lender package. Inconsistency across runs is also harder to quality-control at scale, since the same input may produce a subtly different extraction or analysis each time, complicating any comparison or audit process.
Can I change the temperature setting when using off-the-shelf AI tools, or only when building custom tools?
Most consumer-facing AI tools do not expose temperature as a user-adjustable setting, though some platforms allow it in advanced or developer modes. Temperature control is most accessible when accessing a model through an API, which is typical when a CRE firm is building a custom internal tool, an automated workflow, or an integration with a platform like Yardi or ARGUS. Teams using off-the-shelf tools should assume the vendor has set a default temperature and evaluate whether that default is appropriate for the precision their workflows require.
Click here to get this CRE Glossary in an eBook (PDF) format.