ReasonForge

Deterministic math tools for small language models.

ReasonForge gives small LLMs (8B–32B) access to a verified SymPy computation backend via tool calling. Instead of relying on the model to compute, all math is delegated to deterministic tools — the model only reasons about what to compute and how to present results.

Architecture

User Question → LLM (Qwen3) → Tool Calls → SymPy Backend → Verified Results → LLM → Final Answer

Multi-Turn Agentic Loop:

1. Reason: The model uses <think> tags to analyze the problem and decide on a strategy.
2. Execute: The model delegates computation to a deterministic tool (SymPy or Python sandbox).
3. Iterate: The model observes the verified tool output and either concludes the answer or calls another tool until solved (up to MAX_ROUNDS).

Tools

Tool	Operations	Backend
math_tool	compute, solve, simplify, factor, expand, gcd, lcm, prime_factors, divisors, mod_inverse, nsolve, crt + SymPy builtins (totient, fibonacci, isprime...)	SymPy
calculus_tool	differentiate, integrate, limit, series, summation, partial_fraction, trigsimp, ode_solve, laplace	SymPy
matrix_tool	determinant, inverse, eigenvalues, eigenvectors, rank, rref, transpose, multiply, add, trace, nullspace, columnspace, charpoly, norm, adjugate, solve (Ax=b)	SymPy
statistics_tool	describe, mean, median, mode, std, variance, correlation, regression, percentile, zscore, skewness, kurtosis, geometric_mean, harmonic_mean	Python stdlib
code_tool	run, check, ast_inspect — sandboxed Python code execution, syntax checking, and structure analysis	subprocess

Project Structure

MCP/
├── core.py                    # Shared LLM request logic, expert definitions, tool schemas
├── experts/
│   ├── math/
│   │   ├── server.py          # MCP server entry point (math tools)
│   │   └── tools/
│   │       ├── preprocess.py  # Expression parser (^ → **, implicit multiplication)
│   │       ├── algebra.py     # algebra + number theory
│   │       ├── calculus.py    # derivatives, integrals, ODEs
│   │       ├── matrix.py      # linear algebra
│   │       └── statistics.py  # descriptive & inferential stats
│   └── code/
│       ├── server.py          # MCP server entry point (code execution)
│       └── tools/
│           └── code.py        # Sandboxed Python runner & syntax checker
├── tests/
│   ├── sanity.py              # Tool unit tests (16 checks)
│   ├── math_benchmark.py      # A/B math benchmark (MATH-500 dataset)
│   ├── code_benchmark.py      # A/B code benchmark (HumanEval)
│   └── results/               # Local benchmark outputs 
├── ui/
│   ├── app.py                 # Gradio chat interface with intermediate thinking steps
│   └── style.css              # Custom UI styles (dark mode, thinking blocks)
├── ReasonForge_Colab.ipynb    # One-click Colab deployment notebook
├── pyproject.toml
├── requirements.txt
├── run_tests.bat              # Local tests launcher (Windows)
└── run_ui.bat                 # Local UI launcher (Windows)

Quick Start (Local)

# Requires: Ollama running with a supported model (qwen3:8b, qwen3:32b, etc.)
uv sync
uv run python -m ui.app
# Open at http://localhost:7861

Colab Deployment (GPU)

Open ReasonForge_Colab.ipynb in Google Colab Pro with an A100 GPU. It clones this repo, installs Ollama + qwen3:32b, and launches the UI with a public Gradio link.

Benchmarking

# Math benchmark — MATH-500 (requires Ollama running)
uv run python -m tests.math_benchmark --model llama3.2:3b --n 10
uv run python -m tests.math_benchmark --model qwen3:32b --n 50 --think

# Code benchmark — HumanEval (requires Ollama running)
uv run python -m tests.code_benchmark --model qwen3:8b --n 20
uv run python -m tests.code_benchmark --model qwen3:32b --n 164 --think

Running Sanity Tests

uv run python -m tests.sanity

Benchmark Results

MATH-500 (qwen3:8b, 50 problems)

Metric	Baseline	ReasonForge
Correct	43/50	45/50
Uniform Accuracy	86.0%	90.0% (▲ +4.0%)
Weighted Score	144/176	154/176
Weighted Accuracy	81.8%	87.5% (▲ +5.7%)

• Delegation: 40.0% (20/50) of tasks used tools
• Avg Rounds: 1.5
• Avg Time: Baseline 46.3s vs ReasonForge 31.0s (Δ -15.2s)

By Difficulty

Level 1      5/5   100%  ████████████████████
Level 2      7/7   100%  ████████████████████
Level 3      8/9   89%   █████████████████
Level 4     14/15  93%   ██████████████████
Level 5     11/14  79%   ███████████████  (+14%)

By Category

Algebra                   10/12  83%   ████████████████
Counting & Probability     4/4   100%  ████████████████████
Geometry                   4/4   100%  ████████████████████
Intermediate Algebra      11/13  85%   ████████████████  (+8%)
Number Theory              2/2   100%  ████████████████████
Prealgebra                 7/7   100%  ████████████████████
Precalculus                7/8   88%   █████████████████  (+12%)

HumanEval (Code: qwen3:8b, 160 problems)

Metric	Baseline	ReasonForge
Pass@1	4/160	102/160
Accuracy	2.5%	63.7% (▲ +61.2%)

• Delegation: 31.2% (50/160) of tasks used tools
• Avg Rounds: 1.5
• Avg Time: Baseline 23.9s vs ReasonForge 24.8s (Δ +0.9s)
• Wins vs Losses: ReasonForge successfully solved 100 problems that the Baseline failed on, while only losing 2.

Key Takeaways

Testing the 8-billion parameter qwen3 model reveals exactly why deterministic tool-delegation is crucial for smaller models:

1. Math (MATH-500): While both models achieved incredibly high baseline accuracy, giving the model access to the SymPy backend massively reduced latency (cutting the average computation time from 46.3s down to 31.0s), all while squeezing out an extra ~5% in weighted grading accuracy.
2. Code (HumanEval): Without sandboxed execution tools, the 8B model almost entirely collapsed on HumanEval, only passing a dismal 4/160 (2.5%) of the problems. However, the simple addition of the ReasonForge Python runtime tools allowed the exact same model to safely hypothesize, test, and iteratably structure its code, propelling its accuracy to 102/160 (63.7%)—a gigantic +61.2% improvement with zero fine-tuning required.

Tech Stack

• LLM Backend: Ollama (local) or any OpenAI-compatible API
• Math Engine: SymPy — symbolic computation
• Math Grading: math-verify — deterministic LaTeX parser (Linux/Colab)
• Code Grading: Self-contained HumanEval harness (inspired by openai/human-eval)
• UI: Gradio — chat interface with LaTeX rendering
• Protocol: MCP (Model Context Protocol) compatible