Implementation:Iamhankai Forest of Thought Monte Carlo Forest
| Knowledge Sources | |
|---|---|
| Domains | Reasoning, Search_Algorithms, Ensemble_Methods |
| Last Updated | 2026-02-14 03:00 GMT |
Overview
Concrete tool for orchestrating multi-tree reasoning forests provided by the Forest-of-Thought repository.
Description
The Monte_Carlo_Forest class is the top-level orchestrator for FoT evaluation. It manages a configurable number of reasoning trees, dispatches to the appropriate base search algorithm (MCTS via mctsr_run, ToT via tot_run, or CoT via cot_run), tracks per-tree answers and activation signals, implements early stopping via majority consensus, and delegates final answer selection to get_fot_final_answer.
Usage
Instantiate with parsed arguments, then call run() for each dataset example in the evaluation loop. The run() method returns per-example results including the forest's final answer, per-tree answers, and correctness flags.
Code Reference
Source Location
- Repository: Forest-of-Thought
- File: run_with_mcf_stop_noearly.py
- Lines: L314-599
Signature
class Monte_Carlo_Forest:
def __init__(self, args) -> None:
"""
Args:
args (argparse.Namespace): Configuration with:
- max_iter (int): MCTS iterations per tree
- dataset (str): Dataset identifier
- model_path (str): LLM checkpoint path
- tree_nums (int): Number of trees in forest
- output_dir (str): Output directory
- stop (str): Stopping strategy (cgdm/random/majority/score)
- base_mode (str): Base algorithm (mcts/cot/tot)
"""
def run(self, example) -> list:
"""
Execute forest search on a single dataset example.
Args:
example: Dataset example with query and ground truth.
Returns:
list[dict]: Per-example results with keys:
- query (str): Original question
- ground_truth (str): Expected answer
- fot_ans (str): Forest's final answer
- trees_ans (list): Per-tree answers
- is_correct (bool): Correctness flag
- correct_num (int): Running accuracy counter
"""
def get_fot_final_answer(
self, query, activated_answers_list,
total_answers_list, activated_answer_scores=[],
t=-1, fot=False
) -> str:
"""
Select final answer using configured stopping strategy.
Args:
query (str): Original question
activated_answers_list (list): Extracted answers from trees
total_answers_list (list): All raw tree outputs
activated_answer_scores (list): Reward scores
t (int): Tree index
fot (bool): Forest-of-trees context flag
Returns:
str: Selected best answer
"""
Import
# Monte_Carlo_Forest is defined inline in the main script
# Not importable as a separate module
from run_with_mcf_stop_noearly import Monte_Carlo_Forest
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| args | argparse.Namespace | Yes | Full experiment configuration from parse_args() |
| example | datasets.Dataset row | Yes | Single dataset example with query and ground truth fields |
Outputs
| Name | Type | Description |
|---|---|---|
| results | list[dict] | Per-example result dictionaries with query, ground_truth, fot_ans, trees_ans, is_correct, correct_num |
| JSON files | Files | Per-example JSON records written to output_dir/{dataset}/jsons/{hash}.json |
Usage Examples
Basic Forest Evaluation
from models.load_local_model import Pipeline
from utils.utils import mcts_load_data
# 1. Setup
args = parse_args()
client = Pipeline(model_id=args.model_path, model_type=args.model_type)
dataset = mcts_load_data(args)
# 2. Create forest
mcf = Monte_Carlo_Forest(args)
# 3. Evaluate each example
for i, example in enumerate(dataset):
results = mcf.run(example)
for r in results:
print(f"Q: {r['query'][:50]}... A: {r['fot_ans']} Correct: {r['is_correct']}")
Related Pages
Implements Principle
Requires Environment
Uses Heuristic
Page Connections
Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment