ICLR 2026 Orals

FALCON: Few-step Accurate Likelihoods for Continuous Flows

Danyal Rehman, Tara Akhound-Sadegh, Artem Gazizov, Yoshua Bengio, Alexander Tong

Diffusion & Flow Matching Sat, Apr 25 · 4:03 PM–4:13 PM · 201 C Avg rating: 7.00 (4–10)
Author-provided TL;DR

Few-step Flow Matching with Accurate Likelihoods for Scalable Boltzmann Generators

Abstract

Scalable sampling of molecular states in thermodynamic equilibrium is a long-standing challenge in statistical physics. Boltzmann Generators tackle this problem by pairing a generative model, capable of exact likelihood computation, with importance sampling to obtain consistent samples under the target distribution. Current Boltzmann Generators primarily use continuous normalizing flows (CNFs) trained with flow matching for efficient training of powerful models. However, likelihood calculation for these models is extremely costly, requiring thousands of function evaluations per sample, severely limiting their adoption. In this work, we propose Few-Step Accurate Likelihoods for Continuous Flows (FALCON), a method which allows for few-step sampling with a likelihood accurate enough for importance sampling applications by introducing a hybrid training objective that encourages invertibility. We show FALCON outperforms state-of-the-art normalizing flow models for molecular Boltzmann sampling and is two orders of magnitude faster than the equivalently performing CNF model. FALCON code is available at: https://github.com/danyalrehman/FALCON.

One-sentence summary·Auto-generated by claude-haiku-4-5-20251001(?)

Triple-BERT addresses order dispatching via centralized SARL with action decomposition and BERT-based attention.

Contributions·Auto-generated by claude-haiku-4-5-20251001(?)
  • First centralized Single Agent Reinforcement Learning method for large-scale ride-hailing order dispatching
  • Action decomposition strategy breaking down joint action probability into individual driver probabilities
  • BERT-based network with parameter reuse capturing complex driver-order relationships with reduced parameters
Methods used·Auto-generated by claude-haiku-4-5-20251001(?)
  • Reinforcement learning
  • BERT attention mechanism
  • Action decomposition
  • Parameter reuse
Datasets used·Auto-generated by claude-haiku-4-5-20251001(?)
  • Manhattan ride-hailing real-world dataset
Limitations (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)
  • More sensitive to single points of failure compared to traditional MARL methods, decisions depend on comprehensive information from all drivers and orders
    from the paper
Future work (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)
  • Explore importance sampling within off-policy policy gradient-based actor optimization
    from the paper
  • Investigate offline training to replace pre-training phase
    from the paper
  • Identify more efficient SARL frameworks or enhancements to existing method
    from the paper

Author keywords

  • Generative Models
  • Flow Matching
  • Boltzmann Generators
  • AI for Science

Related orals

Generative Human Geometry Distribution

Introduces distribution-over-distribution model combining geometry distributions with two-stage flow matching for human 3D generation.

Avg rating: 5.50 (2–8) · Xiangjun Tang et al.
Something off? Let us know →