Research map / Probabilistic and scientific modeling / Variational and generative sampling
FALCON: Few-step Accurate Likelihoods for Continuous Flows
Danyal Rehman, Tara Akhound-Sadegh, Artem Gazizov, Yoshua Bengio, Alexander Tong
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 |
|---|---|---|---|
| FALCON (MeanFlow-based few-step CNF) instead of ECNF++ on larger peptides | FALCON · AL3/AL4/AL6 | ESS, E-W2, T-W2: AL6 E-W2 10.668 -> 0.892 | Improved |
| FALCON instead of ECNF++ on alanine dipeptide | FALCON · ALDP | ESS, T-W2: ESS 0.275 -> 0.067 | Hurt |
| FALCON (4-16 steps) instead of SBG discrete flow | FALCON · AL3/AL4/AL6 | E-W2: AL4 1.068 -> 0.686 | Improved |
| use high-NFE adaptive Dopri5 solver instead of few-step sampling (ablation) | FALCON · ALDP | ESS: 0.067 -> 0.264 | Improved |
| use high-NFE Dopri5 instead of few-step sampling (ablation) | FALCON · AL4 | E-W2: 0.686 -> 0.665 | No clear effect |
| use high-NFE Dopri5 instead of few-step sampling (ablation) | FALCON · AL6 | E-W2: 0.892 -> 1.013 | Hurt |
| FALCON instead of ECNF++ for training+inference time | FALCON · hexa-alanine | cumulative time: 137.4 -> 25.76 | Improved |
| increase invertibility regularization strength (ablation) | FALCON · ALDP | ESS, T-W2 | Mixed |
| use EDM inference schedule instead of other schedules (ablation) | FALCON · ALDP 8 steps | sample quality | Improved |
| replace MeanFlow loss with additivity loss (FALCON-A) (ablation) | FALCON · AL3/AL4/AL6 | E-W2: AL4 0.686 -> 2.929 | Hurt |
Closest papers by meaning
- Scalable Equilibrium Sampling with Sequential Boltzmann Generators (ICML 2025 · similarity 0.91)
- BoltzNCE: Learning likelihoods for Boltzmann Generation with Stochastic Interpolants and Noise Contrastive Estimation (NeurIPS 2025 · similarity 0.91)
- Temperature-Annealed Boltzmann Generators (ICML 2025 · similarity 0.89)
- Path Gradients after Flow Matching (NeurIPS 2025 · similarity 0.89)
- Conditional Normalizing Flows for Active Learning of Coarse-Grained Molecular Representations (ICML 2024 · similarity 0.87)
- Transferable Boltzmann Generators (NeurIPS 2024 · similarity 0.87)
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.
