SDK: Is there a generic way to control thinking/reasoning?

[johann-petrak]

With "reasoning"/"thinking" models it may be possible to control if or how intensive the CoT process should be and sometimes also some other specifications controlling the effort.

But I was unable to find a good overview of which parameters can be used (or must not be used!) with which models in the SDK.

This page: https://docs.litellm.ai/docs/reasoning_content is more focused on how the resoning content is returned but only gives examples of how to pass parameters, focusing on two different ones:

• reasoning_effort which can be "low", "medium" , "high", supposedly
• thinking: dict which apparently takes "type" (="enabled", what else is possible) and "budget_tokens"

Is there an overview over which parameters actually work with which models? If a reasoning-related parameter is passed to a model which is not a reasoning model or which does not support that parameter will it be silently dropped or does the developer have to make sure not to pass the parameter?

I think it would be useful for a library which tries to unify access to the LLM models to make sure that:

• all parameters can be specified for any model to allow easy switching between models, but parameters or their values get ignored if the model does not support them
• there is some documentation of which parameters work with which models

For example, apparently OpenAI supports reasoning_effort low, medium, high, but Gemini not only alsy supports those, but also "none", but only for some models.
This is messy and complex, but it would be good to have some documentation or perhaps a community-maintained page to maintain and overview, because otherwise every developer has to figure this out for each provider and model by themselves again?

[krrishdholakia]

• if you pass reasoning_content or thinking to an unsupported model, and have drop_params enabled - it will be dropped

Dropping unsupported parameters - https://docs.litellm.ai/docs/completion/drop_params