System 2 Reasoning AI

Experiments on deliberate, interpretable reasoning across neuro-symbolic models, symbolic regression, inference-time scaling, LLM agents, RL post-training in LLMs, and graph-based retrieval.

---

🧩 Overview

This repository gathers six advanced projects exploring how modern AI systems can perform System 2–style reasoning — deliberate, structured, and interpretable decision‑making that combines neural and symbolic approaches. Each notebook represents a distinct yet complementary direction toward explainable, compositional, and multi‑step reasoning.

The work spans topics from neuro‑symbolic program induction and symbolic regression to inference‑time reasoning optimization, reinforcement learning post‑training for LLMs, vision‑based LLM agents, and GraphRAG‑style retrieval.

---

📚 Contents

#	Project	Path	Description
1	Neuro‑Symbolic Reasoning	neuro_symbolic_reasoning/neuro_symbolic_reasoning.ipynb	CLEVR question answering via program induction, symbolic execution, and seq2seq models (LSTM + Transformer).
2	Symbolic Regression	symbolic_regression/symbolic_regression.ipynb	Equation discovery with Equation Learner (EQL) layers and a Transformer Seq2Seq model (token vocabulary → SymPy expressions).
3	Inference‑Time Scaling	inference_time_scaling/inference_time_scaling.ipynb	Compares Chain‑of‑Thought, Best‑of‑N, Beam Search, Self‑Refine, Tree‑of‑Thoughts, A*, and MCTS on math reasoning.
4	RL Post‑Training for LLMs	rl_post_training/rl_post_training.ipynb	Two‑stage fine‑tuning (SFT → GRPO/TRL) with custom rewards for structure (<think>…</think> & <answer>…</answer>) and correctness.
5	Vision LLM Agent	vision_llm_agent/vision_llm_agent.ipynb	Multi‑agent vision reasoning combining OpenCV heuristics with a VLM (Qwen‑VL). Includes ablations and a small 100‑image dataset.
6	GraphRAG Pipeline	graph_rag/graph_rag.ipynb	Graph‑based retrieval and community reasoning with entity extraction, Leiden community detection, and community‑scoped answering.

---

🎯 Core Themes

• System 2 Reasoning: deliberate, multi‑step problem solving
• Hybrid Neuro‑Symbolic Learning: combining neural inference with symbolic structure
• Reasoning‑Time Optimization: inference‑time scaling, beam search, and ToT/MCTS exploration
• Reinforcement Learning Post‑Training: reward shaping for structured reasoning in LLMs
• LLM Agents & Vision Integration: combining perception with reasoning chains
• Graph‑Based Retrieval: leveraging community structure for contextual memory

---

⚙️ Repository Structure

System2-Reasoning-AI
│   .gitignore
│   LICENSE
│   README.md
│
├── graph_rag
│   └── graph_rag.ipynb
│
├── inference_time_scaling
│   └── inference_time_scaling.ipynb
│
├── neuro_symbolic_reasoning
│   │   neuro_symbolic_reasoning.ipynb
│   │   prompt_example.txt
│   │
│   ├── dataH5Files
│   │   └── (dataset files)
│   │
│   └── utils
│       ├── clevr_executor.py
│       ├── logger.py
│       ├── preprocess.py
│       ├── preprocess_questions.py
│       ├── programs.py
│       ├── utils.py
│       └── __init__.py
│
├── rl_post_training
│   └── rl_post_training.ipynb
│
├── symbolic_regression
│   ├── dataset.csv
│   └── symbolic_regression.ipynb
│
└── vision_llm_agent
    │   vision_llm_agent.ipynb
    │
    └── agent_data
        │   data.csv
        │
        └── images
            ├── 1018.png
            ├── 10461.png
            ├── 10546.png
            ├── 10916.png
            ├── 11286.png
            ├── ...
            └── 9588.png

Each folder contains an independent Jupyter notebook and (where applicable) the supporting data included in your upload.

---

Environment Setup

These notebooks target Python 3.10+. Install only what you need for the notebook you plan to run.

pip install torch transformers accelerate datasets tqdm numpy matplotlib pandas sympy scikit-learn             trl peft vllm opencv-python pillow             langchain langchain-community langchain-graphrag networkx cdlib pypdf

Notes

• Some parts (e.g., RL post‑training and GraphRAG community detection) benefit from a GPU runtime.
• If using verifier or external API calls in the inference‑time scaling notebook, configure your API keys via environment variables instead of hard‑coding.

---

📊 Selected Observations

• Vision LLM Agent: In the provided ablations, the best deep‑agent configuration (Agent 1 + Agent 3) outperformed zero‑shot and classic pipelines on the 100‑image set.
• Inference‑Time Scaling: Search‑based and verification‑based strategies (e.g., Best‑of‑N, ToT/MCTS) showed higher accuracy than plain CoT at additional compute cost.
• RL Post‑Training for LLMs: GRPO with structured rewards improved format compliance and answer correctness relative to SFT‑only baselines.
• GraphRAG: Community summaries derived from the entity‑relation graph improved long‑document question answering quality relative to naive chunk retrieval.

(Exact metrics depend on runtime settings and hardware; see notebook outputs for details.)

---

Motivation

System 1 reasoning in LLMs is fast but often shallow. This project explores System 2 reasoning — deliberate, symbolic, and interpretable — by experimenting with architectures and training strategies that encourage models to reason, plan, and reflect.

---

Citation

If you use this repository in your research, please cite it as follows:

@misc{System2-Reasoning-AI,
  author       = {[Radin Cheraghi/SUT]},
  title        = {Experiments on System 2 reasoning},
  year         = {2025},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/radinch/System2-Reasoning-AI.git}}
}