checking-reasoning-steps.md

description
Where did we go wrong?

Checking reasoning steps

Let’s change the interface of the verifier so that it doesn’t just take an answer, but also a sequence of reasoning steps leading up to it. This way, we can check each step independently and get a probability that it’s correct.

Representing and rendering reasoning steps

First, let’s represent reasoning steps as a list (so that we can more easily manipulate them programmatically) and make a function to render them as a string (so that we can use them in prompts):

{% code title="verify/utils.py" overflow="wrap" %}

from fvalues import F

DEFAULT_QUESTION = "Beth bakes 4x 2 dozen batches of cookies in a week. If these cookies are shared amongst 16 people equally, how many cookies does each person consume?"

DEFAULT_STEPS = [
    "Beth bakes 4x 2 dozen batches of cookies for a total of 4*2 = 8 dozen cookies",
    "There are 12 cookies in a dozen and she makes 8 dozen cookies for a total of 12*8 = 96 cookies",
    "She splits the 96 cookies equally amongst 16 people so they each eat 96/16 = 6 cookies",
    "So, the final answer is 6 cookies per person.",
]


def render_steps(steps: list[str]) -> str:
    return F("\n").join(F(f"{i}. {step}") for (i, step) in enumerate(steps, start=1))

{% endcode %}

If we run render_steps(DEFAULT_STEPS), we get back the original numbered list:

{% code overflow="wrap" %}

1. Beth bakes 4x 2 dozen batches of cookies for a total of 4*2 = 8 dozen cookies
2. There are 12 cookies in a dozen and she makes 8 dozen cookies for a total of 12*8 = 96 cookies
3. She splits the 96 cookies equally amongst 16 people so they each eat 96/16 = 6 cookies
4. So, the final answer is 6 cookies per person.

{% endcode %}

Verifying a step

Given a list of steps, let’s first think about how we can verify the last step, assuming all previous ones are correct.

This is effectively the same as the global verifier above, except that we need to render the steps before we make the prompt. We’ll also already factor out the step-verification into a function check_step so that we can reuse it later.

{% code title="verify/last.py" overflow="wrap" %}

from fvalues import F

from ice.recipe import recipe
from ice.recipes.primer.verify.utils import *


def make_verification_prompt(question: str, steps: list[str]) -> str:
    return F(
        f"""Consider this question: "{question}"

Here are the first few steps of an answer:

{render_steps(steps)}

Q: Is step {len(steps)} correct, assuming that the previous steps are correct? Say "A: Yes" or "A: No".
A:"""
    )


async def check_step(question: str, steps: list[str]) -> float:
    """
    Return the probability that the step is correct
    """
    prompt = make_verification_prompt(question=question, steps=steps)
    answer_probs, _ = await recipe.agent().classify(
        prompt=prompt, choices=(" Yes", " No")
    )
    return answer_probs.get(" Yes", 0.0)


async def verify_answer(
    question: str = DEFAULT_QUESTION, steps: list[str] = DEFAULT_STEPS
):
    return await check_step(question=question, steps=steps)


recipe.main(verify_answer)

{% endcode %}

If we run this with the default question and steps:

python verify_last.py

We get:

0.8373182599538002

Note that (as we’d expect) this probability of the last step being correct is significantly higher than the probability the model assigned to the entire answer being correct.

Verifying all steps

To verify all steps, we simply replace verify_answer with an (async) map over the prefix of each step:

{% code title="verify/steps.py" %}

from ice.recipe import recipe
from ice.recipes.primer.verify.last import check_step
from ice.recipes.primer.verify.utils import *
from ice.utils import map_async


async def verify_answer(
    question: str = DEFAULT_QUESTION, steps: list[str] = DEFAULT_STEPS
):
    """
    For each prefix of 1..n steps, check if the nth step is correct.
    """
    step_indices = list(range(1, len(steps) + 1))
    step_probs = await map_async(
        step_indices,
        lambda index: check_step(question=question, steps=steps[:index]),
    )
    return list(zip(step_probs, steps))


recipe.main(verify_answer)

{% endcode %}

Instead of just returning the probabilities, we return pairs of probabilities and steps to make the result easier to read. It looks like this:

{% code overflow="wrap" %}

[
    (
        0.7626456256640988,
        'Beth bakes 4x 2 dozen batches of cookies for a total of 4*2 = 8 dozen cookies'
    ),
    (
        0.5670302526651101,
        'There are 12 cookies in a dozen and she makes 8 dozen cookies for a total of 12*8 = 96 cookies'
    ),
    (
        0.5074000096282046,
        'She splits the 96 cookies equally amongst 16 people so they each eat 96/16 = 6 cookies'
    ),
    (
        0.827695609429836,
        'So, the final answer is 6 cookies per person.'
    )
]

{% endcode %}

The more difficult the math, the lower the probability the model assigns to the step being correct.

Execution trace (view online)

Exercises

1. How could you use the probabilities we get for each step? One idea is to use a model to resample steps that are wrong. Can you use this to answer questions more correctly?
2. If we multiply the probabilities above to get the probability that the argument overall is correct, we get $$0.76 \cdot 0.57 \cdot 0.51 \cdot 0.83 = 0.18$$. In general, the more steps, the lower we should expect the product probability to be. If we can’t get high probability by just checking the answer, and we can’t get it by checking many steps, how can we ever confidently conclude that an answer is correct? What does your answer to this question mean for how to implement and check reasoning using language models?

Get feedback on exercise solutions

If you want feedback on your exercise solutions, submit them through this form. We—the team at Ought—are happy to give our quick take on whether you missed any interesting ideas.

References

1. Cobbe, Karl, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training Verifiers to Solve Math Word Problems. October 27, 2021.