Research map / Large language models / Language, safety and interpretability
Controllable Molecule Generation via Sparse Representation Editing: An Interpretability-Driven Perspective
Zhuoran Li, Xu Sun, Chang Wen Chen, Wanyu Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| edit SAE concept activations (SpaRE) instead of baselines for site-specific generation | MolXPT-based LLM · ChEBI-20 site-specific | control success rate (%): 62.58 (RetMol) -> 98.92 | Improved |
| use SAE expansion factor 40 vs 10, 20, 80 or 100 (ablation) | SpaRE local control · ChEBI-20 | control success rate (%): 98.92 at 40; 93.52, 91.43, 96.47, 85.74 at 10, 20, 80, 100 | Mixed |
| place the SAE edit at layer 22 instead of earlier layers (ablation) | SpaRE local control · ChEBI-20 | control success rate (%): 0.92 (layer 0) -> 98.92 (layer 22) | Improved |
| apply global control at layer 10 rather than layers 4 or 16 (ablation) | SpaRE global solubility control · solubility LogP | controllability | Improved |
| add Monte Carlo Tree Search on top of SpaRE edits | SpaRE · GSK3b/JNK3 dual inhibitor design | success (%): 98.1 vs 95.2 best baseline (Atomas) | Improved |
| naive string substitution instead of structure-aware editing (ablation) | local site editing · ChEBI-20 | chemical plausibility | Hurt |
| use a small subset of SAE features (nine) for linear probing instead of the full feature set (ablation) | solubility probe · solubility classification | accuracy (%): 88% vs 93% upper bound | No clear effect |
Closest papers by meaning
- Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.90)
- Projecting Molecules into Synthesizable Chemical Spaces (ICML 2024 · similarity 0.88)
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.88)
- Representing Molecules as Random Walks Over Interpretable Grammars (ICML 2024 · similarity 0.88)
- Molecule Design by Latent Prompt Transformer (NeurIPS 2024 · similarity 0.88)
- Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
