Skip to content

Chain-of-Thought Reasoning

Enable <think> step-by-step reasoning before the final answer. Critical for complex physical reasoning tasks.

See It In Action

Chain-of-thought driving analysis on Jetson

📺 Can't see the animation? Download MP4

How It Works

graph LR
    Q["Question"] --> T["&lt;think&gt;<br/>Step-by-step<br/>reasoning<br/>&lt;/think&gt;"]
    T --> A["Final Answer"]

    style T fill:#fff3cd,stroke:#ffc107,color:#333

When reasoning=True, Cosmos generates a reasoning trace wrapped in <think>...</think> tags, then provides a concise final answer.

Enable Chain-of-Thought

from strands import Agent
from strands_cosmos import CosmosVisionModel

model = CosmosVisionModel(
    model_id="nvidia/Cosmos-Reason2-2B",
    reasoning=True,  # ← This enables CoT
    params={"temperature": 0.6},
)
agent = Agent(model=model)

→ Full driving analysis example with CoT

Example Output

<think>
I can see a residential street with parked cars on both sides.
There is a crosswalk ahead with a pedestrian waiting to cross.
The traffic light appears to be green for the ego vehicle.
However, the pedestrian might step onto the road.
The recommended action is to slow down and prepare to yield.
</think>

The driver should slow down approaching the crosswalk and be
prepared to yield to the pedestrian on the right.

When to Use Reasoning

Scenario reasoning= Why
Quick captions False Speed matters
Safety analysis True Need thorough analysis
Robot planning True Multi-step planning
Video descriptions False Direct description
Physics reasoning True Logical chain needed

Performance Impact

Reasoning increases output length and inference time:

Mode Typical Tokens Inference Time (2B)
Direct (reasoning=False) 100–500 5–15s
CoT (reasoning=True) 500–2000 15–55s

Temperature for reasoning

Use temperature=0.6 with reasoning enabled. Lower values produce more focused chains, higher values explore more possibilities.

Robot Embodied Reasoning with CoT

Robot embodied reasoning with chain-of-thought

model = CosmosVisionModel(
    model_id="nvidia/Cosmos-Reason2-2B",
    reasoning=True,
)
agent = Agent(model=model)

agent("<image>robot_workspace.png</image> What is the next action?")

→ Full embodied reasoning example

What's Next