prteek/newton-net

Newton net

A tiny language model plus a fixed, differentiable Newtonian planner (the equations of motion and F = m·a) in one computation graph. It answers short driving questions ("doing 70 mph in the rain, how far to stop?") with exact maths and a trace of every step, and refuses anything it can't answer. An alternative to LLM tool calling: the physics runs inside the model's forward pass, so the reader learns to read questions from whether the physics gets the answer right.

Status: work in progress, no trained weights yet. This repo holds the pipeline as a minimal first version, each stage just enough for the demo:

file	stage
extract.py	1. Extractor: each number + unit → dimension and SI value (regex over a table of units); the text with values replaced by markers
reader.py	2. Reader + router: what is asked and how the values fit together; keyword rules standing in for the TinyLLM
assembler.py	3. Assembler: the extracted values and the reader's choices → one value per quantity
planner.py	4. Planner: chains the equations of motion to the target; differentiable on torch tensors

from model.assembler import assemble
from model.extract import extract, mask
from model.planner import solve
from model.reader import read

question = "A car at 30 m/s brakes at 0.7g. How long until it stops?"
quantities = extract(question)
choice = read(mask(question, quantities))
steps, values = solve(assemble(quantities, choice), choice["target"])
print(steps)  # ['t = (v - u)/a = 4.37']

• Training data: prteek/newton-net-questions
• Demo (keyword rules in place of the reader until it is trained): prteek/newton-net
• Code and design notes: github.com/prteek/newton-net, see decisions.md