Welcome to the LLM Reasoning Task Collection repository! This project is an open collaboration to create a comprehensive master list of reasoning tasks that can teach, elicit, or show reasoning samples to large language models (LLMs) for training purposes.

• Open Reasoning Tasks: LLM Reasoning Tasks Collection
  • Contents
  • Introduction
  • Contributing
  • License
  • Citation

The goal of this repository is to gather a diverse set of reasoning tasks designed to improve the reasoning capabilities of LLMs. Contributors are encouraged to submit tasks, provide examples, and optionally include diagrams or workflows to illustrate how the tasks function.

You can access the main tasks list table by clicking here (or open tasks.md file in the top level directory)

You can access the full table of reasoning tasks from our quarto based website by clicking here.

Coming Soon

Coming Soon

Coming Soon

This repository now includes the OpenCog Multi-Agent Orchestration Workbench, a sophisticated framework for autonomous multi-agent orchestration of LLM reasoning tasks.

The workbench now loads all 95 task definitions and exposes a ranked, prerequisite-aware curriculum with persisted mastery evidence. Bounded structured truth tables, syllogisms, Hanoi, and stack ordering have executable backends; other families explicitly report unsupported capabilities. Native OpenCog subsystems and advanced reasoning remain integration work, not existing support.

• Agent-Zero Master Builder: Coordinates cognitive architectures to develop sophisticated strategies for all reasoning challenges
• Pattern Language Library: A "living library" inspired by Christopher Alexander's "A Pattern Language" but specifically designed for AGI cognitive architectures
• Multi-Agent Orchestration: Sophisticated coordination of specialized agents with different roles (analyzer, strategist, executor, validator, learner)
• Strategy Development: Adaptive strategy creation and refinement for reasoning challenges

from opencog import OpenCogWorkbench

# Initialize the workbench with reasoning tasks
workbench = OpenCogWorkbench(tasks_dir="tasks")

# Solve a task autonomously with Agent-Zero
result = workbench.solve_task(
    "towers-of-hanoi", orchestration_mode="autonomous", instance={"disks": 3}
)
print(result)

# Get recommendations for a task
recommendations = workbench.get_task_recommendations("deductive-logic-puzzles")
print(recommendations)

# Only prerequisite- and backend-ready families are recommended.
print(workbench.recommend_next_tasks())

python opencog_demo.py

See OPENCOG_README.md for complete documentation of the OpenCog Multi-Agent Orchestration Workbench.

1. Agent-Zero - Master builder that analyzes tasks, builds cognitive architectures, and coordinates agents
2. Pattern Library - 10 foundational cognitive patterns for reasoning and problem-solving
3. Multi-Agent System - Orchestration framework with message passing and task distribution
4. Strategy Repository - 7 core strategies that can be composed and adapted

We welcome contributions from everyone! To contribute, please see our Contribution Guide.

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

@misc{nousresearch2024,
  title = {Open Reasoning Tasks: LLM Reasoning Tasks Collection},
  author = {Nous Research},
  url = {https://github.com/NousResearch/Open-Reasoning-Tasks},
  year = {2024},
}