Implementation:Iamhankai Forest of Thought CoT Task Run

Knowledge Sources	Forest-of-Thought Chain-of-Thought Prompting
Domains	Reasoning, NLP
Last Updated	2026-02-14 03:00 GMT

Overview

Concrete tool for executing single-pass Chain-of-Thought reasoning provided by the Forest-of-Thought repository.

Description

The CoT_Task class extends SearchTask to implement Chain-of-Thought reasoning. It generates a single reasoning path via one LLM call, extracts a summary answer using dataset-specific extraction logic, and optionally applies self-criticism scoring. This is the simplest of the three base reasoning modes in FoT.

Usage

Instantiated by Monte_Carlo_Forest.cot_run() or directly by get_cot_answer() in the main orchestrator. Used when base_mode is set to cot in experiment configuration.

Code Reference

Source Location

• Repository: Forest-of-Thought
• File: methods/cot/task.py
• Lines: L8-180

Signature

class CoT_Task(SearchTask):
    def __init__(
        self, data, propose_method='glm', value_method='glm',
        temperature=0.7, max_tokens=2048, seed=170,
        max_length=2048, truncation=True, do_sample=True,
        max_new_tokens=1024, evaluate='', summary=False,
        lang='en', answer=None, verify_method='string',
        do_self_critic=False
    ):
        """
        Args:
            data: Input problem text.
            evaluate (str): Dataset type (math/scibench/scieval).
            summary (bool): Generate explicit summary.
            lang (str): Language ('en' or 'zh').
            answer: Ground truth for verification.
            do_self_critic (bool): Enable self-criticism scoring.
        """

    def run(self) -> dict:
        """
        Execute single-pass CoT reasoning.

        Returns:
            dict: Keys: content, solution, summary, accurate,
                  real_answer, self_critic (if enabled).
        """

Import

from methods.cot.task import CoT_Task

I/O Contract

Inputs

Name	Type	Required	Description
data	str	Yes	Problem text to solve
evaluate	str	No	Dataset type for answer extraction (default: )
lang	str	No	Language: en or zh (default: en)
answer	str	No	Ground truth for verification
do_self_critic	bool	No	Enable self-criticism confidence scoring (default: False)

Outputs

Name	Type	Description
result	dict	Keys: content (question), solution (full reasoning), summary (final answer), accurate (bool), real_answer (ground truth)

Usage Examples

from methods.cot.task import CoT_Task

task = CoT_Task(
    data="A train travels 60 km/h for 2 hours. How far does it go?",
    evaluate='math',
    lang='en',
    answer="120"
)

result = task.run()
print(f"Solution: {result['solution']}")
print(f"Answer: {result['summary']}")
print(f"Correct: {result['accurate']}")

Related Pages

Implements Principle

• Principle:Iamhankai_Forest_of_Thought_Chain_of_Thought_Reasoning

Requires Environment

• Environment:Iamhankai_Forest_of_Thought_Python_CUDA_Runtime