Implement CoT no-op for reasoning models

Author: FireMasterK

dspy/clients/lm.py

@@ -37,6 +37,7 @@ def __init__(
         callbacks: Optional[List[BaseCallback]] = None,
         num_retries: int = 3,
         provider=None,
+        reasoning_model: Optional[bool] = None,
         finetuning_model: Optional[str] = None,
         launch_kwargs: Optional[dict[str, Any]] = None,
         train_kwargs: Optional[dict[str, Any]] = None,
@@ -51,6 +52,7 @@ def __init__(
             model_type: The type of the model, either ``"chat"`` or ``"text"``.
             temperature: The sampling temperature to use when generating responses.
             max_tokens: The maximum number of tokens to generate per response.
+            reasoning_model: Whether the model is a reasoning model.
             cache: Whether to cache the model responses for reuse to improve performance
                    and reduce costs.
             cache_in_memory (deprecated): To enable additional caching with LRU in memory.
@@ -71,6 +73,7 @@ def __init__(
         self.callbacks = callbacks or []
         self.history = []
         self.num_retries = num_retries
+        self.reasoning_model = reasoning_model if reasoning_model is not None else litellm.supports_reasoning(model)
         self.finetuning_model = finetuning_model
         self.launch_kwargs = launch_kwargs or {}
         self.train_kwargs = train_kwargs or {}

dspy/predict/chain_of_thought.py

@@ -27,12 +27,15 @@ def __init__(
         """
         super().__init__()
         signature = ensure_signature(signature)
-        prefix = "Reasoning: Let's think step by step in order to"
-        desc = "${reasoning}"
-        rationale_field_type = rationale_field.annotation if rationale_field else rationale_field_type
-        rationale_field = rationale_field if rationale_field else dspy.OutputField(prefix=prefix, desc=desc)
-        extended_signature = signature.prepend(name="reasoning", field=rationale_field, type_=rationale_field_type)
-        self.predict = dspy.Predict(extended_signature, **config)
+        if dspy.settings.lm.reasoning_model:
+            self.predict = dspy.Predict(signature, **config)
+        else:
+            prefix = "Reasoning: Let's think step by step in order to"
+            desc = "${reasoning}"
+            rationale_field_type = rationale_field.annotation if rationale_field else rationale_field_type
+            rationale_field = rationale_field if rationale_field else dspy.OutputField(prefix=prefix, desc=desc)
+            extended_signature = signature.prepend(name="reasoning", field=rationale_field, type_=rationale_field_type)
+            self.predict = dspy.Predict(extended_signature, **config)
 
     def forward(self, **kwargs):
         return self.predict(**kwargs)

pyproject.toml

@@ -34,7 +34,7 @@ dependencies = [
     "optuna>=3.4.0",
     "pydantic>=2.0",
     "magicattr>=0.1.6",
-    "litellm>=1.60.3",
+    "litellm>=1.72.4",
     "diskcache>=5.6.0",
     "json-repair>=0.30.0",
     "tenacity>=8.2.3",
@@ -50,7 +50,7 @@ dependencies = [
 [project.optional-dependencies]
 anthropic = ["anthropic>=0.18.0,<1.0.0"]
 weaviate = ["weaviate-client~=4.5.4"]
-aws = ["boto3~=1.34.78"]
+aws = ["boto3~=1.34.34"]
 mcp = ["mcp; python_version >= '3.10'"]
 langchain = ["langchain_core"]
 dev = [
@@ -62,8 +62,8 @@ dev = [
     "pillow>=10.1.0",
     "datamodel_code_generator>=0.26.3",
     "build>=1.0.3",
-    "litellm>=1.60.3; sys_platform == 'win32'",
-    "litellm[proxy]>=1.60.3; sys_platform != 'win32'",
+    "litellm>=1.72.4; sys_platform == 'win32'",
+    "litellm[proxy]>=1.72.4; sys_platform != 'win32'",
 ]
 test_extras = [
     "mcp; python_version >= '3.10'",

tests/predict/test_chain_of_thought.py

@@ -23,3 +23,15 @@ async def test_async_chain_of_thought():
         program = ChainOfThought("question -> answer")
         result = await program.acall(question="What is 1+1?")
         assert result.answer == "2"
+
+
+def test_cot_skips_with_reasoning_model():
+    lm = DummyLM([{"answer": "2"}])
+    lm.reasoning_model = True
+    dspy.settings.configure(lm=lm)
+    signature = dspy.Signature("question -> answer")
+    predict = ChainOfThought(signature)
+    assert list(predict.predict.signature.output_fields.keys()) == [
+        "answer",
+    ]
+    assert predict(question="What is 1+1?").answer == "2"