Vidai: Neural Mathematical Parsing Demo
Work in Progress: This demo showcases a novel neuro-symbolic architecture. Simple expressions work well; complex expressions need more training data.
The Approach
Vidai (Tamil for "answer") uses transformers for what they're good at: pattern recognition and sequence transformation. Instead of asking a model to "learn" that 7 × 8 = 56 from text, Vidai learns to recognize the tree structure inherent in mathematical expressions.
- ContextEncoder: Recognizes expression structure with tree-depth embeddings
- SymbolicParserDecoder: Outputs prefix notation that explicitly encodes the tree
- SymPy: Handles exact computation deterministically
x^2 + 3*y becomes + ** x 2 * 3 y (a tree with + at root).
What Works
| Expression Type | Status |
|---|---|
Simple arithmetic: 3 + 5 * 2 |
✅ Works |
Parenthesized: (x^2) + (3*y) |
✅ Works |
Single functions: sin(x), sqrt(16) |
✅ Works |
| Complex without parens | ⚠️ Needs more training data |
| Deep nesting (3+ levels) | ⚠️ Needs more training data |
Links
- Read the Full Story - Blog post on the journey
- Model Card
- GitHub Repository
- Technical Report
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
🙋
Ask for provider support