Representation learning / SIGMA
Learning representations beyond the surface
Teaching a sequence model that different strings can describe the same underlying molecule.
Improved mean performance across six molecular property-prediction benchmarks.
The problem
The same molecule can be written as different valid strings. A sequence model that focuses too heavily on the particular spelling can treat equivalent inputs inconsistently.
My contribution
I developed SIGMA, a token-level training objective that aligns different strings representing the same molecule. The method changes training without adding inference cost.
- Equivalent molecular strings
- Token-level semantic alignment
- Generation & prediction
Results
- Reduced generation-distribution error, measured by FCD, by 9.9% and 19.4% versus the strongest baselines across two QM9 encodings.
- Improved mean prediction performance on all six molecular benchmarks.
- Reduced sensitivity to equivalent input strings by up to 44.9%.
The experiments are in molecular modeling. The broader research interest is how to learn meaning that remains stable when the representation changes.
Paper
SIGMA: Semantic Identifier Grouping for Molecular Autoregression X. Wang, F. Dou, J. Bi, M. Song. Submitted to AAAI 2027.