Research map / Vision-language and multimodal / Human and embodied modeling
Can Transformers Smell Like Humans?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that MoLFormer, a transformer pre-trained on chemical SMILES without perceptual labels, yields odorant representations that align with human olfactory perception across expert labels, rating prediction, and similarity judgments.
Problem
Whether representations from transformers pre-trained on chemical structures align with human olfactory perception, given the scarcity of large labeled olfaction datasets.
Approach
Use frozen MoLFormer SMILES embeddings as odorant representations and measure alignment with human perception via linear probes (expert labels, continuous ratings) and direct cosine-similarity RSA against human similarity judgments, comparing with supervised Open-POM and physicochemical-descriptor DAM baselines.
Key results
- MoLFormer representational similarity with human similarity ratings reaches r=0.64 on Snitz and r=0.66 on Ravia (p<0.0001), on par with supervised Open-POM and well above DAM.
- On continuous rating prediction, MoLFormer achieves CC 0.20 on Keller and 0.25 on Sagar, slightly below Open-POM (0.22, 0.29) and below DAM on Sagar (0.35).
- MoLFormer predicts 13 of 15 physicochemical descriptors as well as or better than Open-POM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoLFormer representations instead of DAM representations | continuous rating prediction · Keller | Pearson correlation (CC): 0.17\pm{0.00} -> 0.20\pm{0.00} | Helped |
| Use MoLFormer representations instead of Open-POM representations | continuous rating prediction · Keller | Pearson correlation (CC): 0.22\pm{0.01} -> 0.20\pm{0.00} | Hurt |
| Use MoLFormer representations instead of DAM representations | continuous rating prediction · Sagar | Pearson correlation (CC): 0.35\pm{0.01} -> 0.25\pm{0.01} | Hurt |
| Use MoLFormer representations instead of Open-POM representations | continuous rating prediction · Sagar | Pearson correlation (CC): 0.29\pm{0.01} -> 0.25\pm{0.01} | Hurt |
| Use MoLFormer representations instead of DAM representations | continuous rating prediction · Keller | NRMSE: 0.15\pm{0.00} -> 0.15 \pm{0.00} | No clear effect |
| Use MoLFormer representations instead of DAM representations | continuous rating prediction · Sagar | NRMSE: 0.17\pm{0.00} -> 0.19 \pm{0.00} | Hurt |
| Use MoLFormer representations instead of DAM representations | odorant similarity alignment · Snitz | Pearson correlation | Helped |
| Use MoLFormer representations instead of DAM representations | odorant similarity alignment · Ravia | Pearson correlation | Helped |
| Use deeper MoLFormer layers for odorant representations (ablation) | MoLFormer · Snitz and Ravia | representational alignment | Helped |
| Use deeper MoLFormer layers for physicochemical descriptor prediction (ablation) | MoLFormer · physicochemical descriptors | correlation | Hurt |
| Fine-tune MoLFormer on GS-LF before extracting representations (ablation) | MoLFormer · Keller and Sagar | continuous rating prediction | Hurt |
| Fine-tune MoLFormer on GS-LF before extracting representations (ablation) | MoLFormer · Snitz and Ravia | representational similarity alignment | No clear effect |
Surprising findings
- MoLFormer, never trained on perceptual labels, matches supervised Open-POM on similarity alignment and expert-label-based classification is only somewhat below it.
- Alignment with physicochemical descriptors decreases across MoLFormer layers while alignment with perception increases with depth.
Limitations
- Mixture intensity and concentration of components are not modeled in MC-odorant encoding.
- Available perceptual datasets are small and lack variation, especially for continuous rating regression.
- Only averaged ratings are evaluated; per-subject alignment is not assessed.
Experimental setup
Datasets: GS-LF Sagar Keller Ravia Snitz
Benchmarks: GS-LF Sagar Keller Ravia Snitz
Metrics: micro-averaged ROC-AUC NRMSE Pearson correlation
Baselines: Open-POM DAM
Builds on: MoLFormer Principal Odor Map Distance Angle Model
Closest papers by meaning
- Evaluating alignment between humans and neural network representations in image-based learning tasks (NeurIPS 2024 · similarity 0.81)
- Inference with correlated priors using sisters cells (NeurIPS 2025 · similarity 0.80)
- Model–Behavior Alignment under Flexible Evaluation: When the Best-Fitting Model Isn’t the Right One (NeurIPS 2025 · similarity 0.80)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · similarity 0.80)
- Not all solutions are created equal: An analytical dissociation of functional and representational similarity in deep linear neural networks (ICML 2025 · similarity 0.80)
- From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers (NeurIPS 2025 · similarity 0.80)
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.