The reviewed record of science sign in
Pith

arxiv: 2607.05252 · v1 · pith:ZRLYPFMI · submitted 2026-07-06 · cs.LG

FUSE: FK-Steered Multi-Modal Flow Matching for Efficient Simulation-Based Posterior Estimation

Weichen Qin , Yufan Xie , Peihao Wang , Chia-Jui Chou , Minghui Du , Peng Xu , Ziren Luo , Yi Yang
show 3 more authors
Jingyi Yu Bo Liang Jiakai Zhang
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-07 21:22 UTCglm-5.2pith:ZRLYPFMIrecord.jsonopen to challenge →

classification cs.LG
keywords estimationfuseefficientinferencesimulation-basedcomplexexistingfk-steered
0
0 comments X

The pith

FUSE adapts a multimodal diffusion transformer for flow-matching-based posterior estimation and adds inference-time likelihood-guided particle resampling, achieving state-of-the-art posterior fidelity on SBI benchmarks and real-world exoplanet orbital estimation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Scientists often need to infer physical parameters (like a planet's orbit) from observations, but exact methods like MCMC are too slow for real-time use. Neural generative models can approximate the answer quickly, but they often produce posteriors that are too broad or miss complex correlations between parameters. FUSE tackles this in two ways. First, it uses a dual-track transformer architecture (adapted from image-generation models) that keeps parameter and observation information in separate tracks while letting them interact, rather than naively concatenating them. Second, during inference it runs multiple sampling trajectories in parallel and periodically resamples them based on how well their intermediate results match the simulator's likelihood, steering computation toward high-probability regions. On standard SBI benchmarks, the architecture alone outperforms prior methods when enough simulated data is available. On a real exoplanet orbital estimation task (β Pictoris b), the full system with FK-steering recovers sharp posterior features that baselines miss, completing in about three minutes versus 8.5 hours for MCMC.

Core claim

FUSE outperforms state-of-the-art neural SBI baselines (NPE, FMPE, Simformer) on the SBIBM benchmark across seven metrics at 10^5 simulations, and on the real-world β Pictoris b exoplanet orbital estimation task, FUSE with FK-steering recovers complex parameter degeneracies that baselines fail to capture, achieving a Mode L2 Distance of 3.85 versus 75.74 (NPE) and 132.41 (FMPE), while completing inference in ~3 minutes versus ~8.5 hours for PTMCMC.

Load-bearing premise

The FK-steering mechanism depends on the denoised proxy θ̂_t = θ_t − t·v_ϕ(θ_t, t, x) being a faithful approximation of E[θ_0 | θ_t, x] (Eq. 11, §4.3). This proxy is exact only when v_ϕ equals the optimal conditional velocity (Appendix I.2, Eq. 37). When the learned velocity field is imperfect—which is the regime where FK-steering is most needed—the proxy can produce misleading likelihood scores, potentially steering particles toward wrong regions. The paper acknowledges this is a 'tractable likelihood-guided correction rather than an exact posterior sampler' but does not empirically characterize how proxy error degrades FK performance, nor does it evaluate FK-steering across all 10 benchmark tasks (only SLCP and the exoplanet task).

Figures

Figures reproduced from arXiv: 2607.05252 by Bo Liang, Chia-Jui Chou, Jiakai Zhang, Jingyi Yu, Minghui Du, Peihao Wang, Peng Xu, Weichen Qin, Yi Yang, Yufan Xie, Ziren Luo.

Figure 1
Figure 1. Figure 1: Pipeline of FUSE. (a) FUSE’s Architecture.The architecture employs an independent embedding interface to map heteroge￾neous inputs into a shared space, followed by an MM-DiT-based fusion module for multimodal integration. Finally, the posterior predictor estimates the denoising velocity based on refined, parameter-wise token representations. (b) FK-Steered Inference. We steer the FUSE sampling process via … view at source ↗
Figure 2
Figure 2. Figure 2: Sample efficiency comparison on 10 SBIBM tasks. We evaluate posterior fidelity using observation-wise ℓ-C2ST across simulation budgets from 103 to 105 , where 0.5 indicates an ideal match to the reference posterior. Curves report means over the official SBIBM observations, and shaded regions indicate the corresponding standard deviations. FUSE is evaluated without FK-steering in this benchmark to isolate t… view at source ↗
Figure 3
Figure 3. Figure 3: FK-steering improves sample quality. We compare our method against several baselines with the posterior samples and marginal distributions on the SLCP task. The addition of FK-steering further concentrates the samples into the high-density regions of the posterior. Baselines. We compare our approach against three state-of-the-art amortized inference baselines: (1) NPE (NSF) (Papa￾makarios & Murray, 2016; D… view at source ↗
Figure 4
Figure 4. Figure 4: Comprehensive evaluation of orbital parameter estimation for β Pictoris b. (a) Quantitative comparison: A heatmap of the Normalized Sinkhorn Divergence across all 8 orbital parameters. FUSE achieves the lowest error compared to baselines and the ablation version (w/o FK). Note the significant failure of NPE/FMPE in constraining orientation parameters (i, Ω), which our method resolves effectively. (b) Poste… view at source ↗
Figure 5
Figure 5. Figure 5: Posterior distribution comparison between FUSE and Naive Best-of-N selection. The corner plot displays the marginalized 1D and 2D posteriors for the β Pictoris b orbital parameters. The Naive Best-of-N strategy (orange) performs selection only at the final generation step and does not recover several sharp degeneracies and correlations present in the reference. In contrast, FUSE (blue) uses sequential FK-s… view at source ↗
read the original abstract

Simulation-Based Inference (SBI) is critical for scientific discovery, with generative models offering a promising path toward efficient inference. However, existing methods struggle with effective multimodal modeling. They often rely on brute-force fusion strategies that ignore the structural disparities between parameters and observations, thus limiting estimation fidelity. In this work, we introduce FUSE (Feynman-Kac steered mUlti-modal flow matching for efficient Simulation-based posterior Estimation). Unlike prior work, FUSE employs a dual-track architecture that preserves the distinct features of multimodal inputs while facilitating dynamic interaction. Additionally, we propose an FK-steered sampling strategy that leverages intermediate observation likelihoods to guide the generative trajectories, effectively improving the sample quality during inference. Our approach outperforms state-of-the-art baselines on standard SBI benchmarks, producing posteriors that closely match ground-truth MCMC. Furthermore, in a real-world exoplanet orbital estimation task, FUSE successfully resolves complex parameter degeneracies that challenge existing methods, highlighting its potential to accelerate complex scientific discoveries in astrophysics and beyond.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical entities, particles, forces, or dimensions are introduced. The method is a computational framework combining existing mathematical tools.

free parameters (6)
  • λ (FK scaling factor) = not explicitly stated
    Global scaling factor for FK potentials in Eq. 12; controls the strength of likelihood guidance. Value not reported in the main text.
  • α (noise scale) = 0.3
    Controls stochastic perturbation in the SDE sampler (Eq. 10). Set to 0.3 for both SLCP and exoplanet tasks (Appendix B, E).
  • B (beam width / particle count) = 8
    Number of parallel particles for FK resampling. Set to 8 for both SLCP and exoplanet tasks.
  • K (tokens per parameter) = 2
    Number of tokens per scalar parameter. Selected via ablation (Table 2) as optimal.
  • D (hidden dimension) = 128 (benchmark), 256 (exoplanet)
    Hidden width of transformer. Selected via ablation (Table 2).
  • Resampling schedule = every 5 steps, step 20-200
    Defines when FK resampling occurs during inference. Set identically for SLCP and exoplanet.
axioms (4)
  • domain assumption The learned velocity field v_ϕ approximates the optimal conditional velocity well enough that the denoised proxy θ̂_t = θ_t − t·v_ϕ(θ_t, t, x) is a meaningful estimate of E[θ_0 | θ_t, x]
    Invoked in §4.3 (Eq. 11) and justified in Appendix I.2 (Eq. 37-39). This is the load-bearing assumption for FK-steering correctness.
  • domain assumption The simulator likelihood p(x|θ) and prior p(θ) are available and tractable to evaluate at inference time
    Required for computing FK potentials (Eq. 11). Stated in §3.1 and §4.3. This holds for the SLCP and exoplanet tasks but not for all SBI problems.
  • standard math The rectified flow interpolation θ_t = (1−t)θ_0 + tϵ with constant velocity target ϵ−θ_0 is an appropriate transport for posterior estimation
    Standard conditional flow matching assumption (Lipman et al. 2023), used in Eq. 8.
  • domain assumption Multinomial resampling with stochastic rejuvenation sufficiently mitigates particle depletion for finite B
    Invoked in §4.3 and Appendix I.3. The paper acknowledges this is approximate and does not provide formal guarantees on particle diversity.

pith-pipeline@v1.1.0-glm · 28016 in / 3591 out tokens · 195424 ms · 2026-07-07T21:22:08.723675+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · 3 internal anchors

  1. [1]

    Blunt, S., Wang, J

    URL https://openaccess.thecvf.co m/content/CVPR2023/html/Bao_All_Are_ Worth_Words_A_ViT_Backbone_for_Diffu sion_Models_CVPR_2023_paper.html. Blunt, S., Wang, J. J., Angelo, I., Ngo, H., Cody, D., De Rosa, R. J., Graham, J. R., Hirsch, L., Nagpal, V ., Nielsen, E. L., Pearce, L., Rice, M., and Tejada, R. or- bitize!: A comprehensive orbit-fitting software ...

    work page doi:10.3847/1538-3 2020

  2. [2]

    In: 2023 IEEE/CVF Conference on Com-puter Vision and Pattern Recognition (CVPR)

    doi: 10.1109/CVPR52729.2023.01764. URL https://openaccess.thecvf.com/conten t/CVPR2023/html/Brooks_InstructPix2P ix_Learning_To_Follow_Image_Editing_ Instructions_CVPR_2023_paper.html. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y ., Wang, Z., Kwok, J., Luo, P., Lu, H., and Li, Z. PixArt-α: Fast training of diffusion transformer for photorealistic tex...

    work page doi:10.1109/cvpr52729.2023.01764 2023

  3. [3]

    URL https://ojs.aaai.org/index.php/AAA I/article/view/30018

    doi: 10.1609/aaai.v38i18.30018. URL https://ojs.aaai.org/index.php/AAA I/article/view/30018. Cuturi, M. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems, volume 26, 2013. URL https: //papers.nips.cc/paper_files/paper/2 013/hash/af21d0c97db2e27e13572cbf59e b343d-Abstract.html. Dax, M., Gre...

    work page doi:10.1609/aaai.v38i18.30018 2013

  4. [4]

    URL https: //doi.org/10.1007/978-1-4684-9393-1

    doi: 10.1007/978-1-4684-9393-1. URL https: //doi.org/10.1007/978-1-4684-9393-1. Dhariwal, P. and Nichol, A. Diffusion Models Beat GANs on Image Synthesis. InAdvances in Neural Information Processing Systems, volume 34, pp. 8780–8794. Curran Associates, Inc., 2021. URL https://proceeding s.neurips.cc/paper/2021/hash/49ad23d 1ec9fa4bd8d77d02681df5cfa-Abstra...

    work page doi:10.1007/978-1-4684-9393-1 2021

  5. [5]

    Classifier-Free Diffusion Guidance

    doi: 10.1086/670067. URL https://doi.or g/10.1086/670067. Geffner, T., Papamakarios, G., and Mnih, A. Composi- tional score modeling for simulation-based inference. InProceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Ma- chine Learning Research. PMLR, 2023. URL https: //proceedings.mlr.press/v202/geffner 23a...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1086/670067 2023

  6. [6]

    Rectified Flow: A Marginal Preserving Approach to Optimal Transport

    URL https://openreview.net/forum ?id=PqvMRDCJT9t. Liu, J., Liu, G., Liang, J., Li, Y ., Liu, J., Wang, X., Wan, P., Zhang, D., and Ouyang, W. Flow-GRPO: Training flow matching models via online RL. InAdvances in Neural Information Processing Systems, volume 38, 2025. URL https://proceedings.neurips.cc/paper _files/paper/2025/hash/3a10c46572628 d58cb44fb70...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1073/pnas.0306899100 2025

  7. [7]

    Radford, A., Kim, J

    URL https://openaccess.thecvf.co m/content/ICCV2023/html/Peebles_Scal able_Diffusion_Models_with_Transform ers_ICCV_2023_paper.html. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In...

    work page doi:10.1007/978-3-319-24574-4 2021

  8. [8]

    Ruth, M.Exoplanet Orbital Characterization Using Simulation-Based Inference

    URL https://openaccess.thecvf.co m/content/CVPR2023/html/Ruiz_DreamBo oth_Fine_Tuning_Text-to-Image_Diffusi on_Models_for_Subject-Driven_Generati on_CVPR_2023_paper.html. Ruth, M.Exoplanet Orbital Characterization Using Simulation-Based Inference. PhD thesis, University of Li`ege, Li`ege, Belgium, June 2024. URL https://ma theo.uliege.be/handle/2268.2/203...

    work page 2024

  9. [9]

    Song, J., Meng, C., and Ermon, S

    URL https://proceedings.mlr.pres s/v267/singhal25b.html. Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021a. URL https://openrevi ew.net/forum?id=St1giarCHLP. Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling ...

    work page doi:10.1103/shym-w46f 2023

  10. [10]

    Joint Multimodal Learning with Deep Generative Models

    arXiv:1611.01891 [stat]. Tavar´e, S., Balding, D. J., Griffiths, R. C., and Donnelly, P. Inferring Coalescence Times From DNA Sequence Data.Genetics, 145(2):505–518, February 1997. ISSN 1943-2631. doi: 10.1093/genetics/145.2.505. URL 13 FUSE: FK-Steered Multi-Modal Flow Matching https://academic.oup.com/genetics/ar ticle/145/2/505/6018089. Tejero-Cantero,...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1093/genetics/145.2.505 1997

  11. [11]

    URL https: //doi.org/10.21105/joss.02505

    doi: 10.21105/joss.02505. URL https: //doi.org/10.21105/joss.02505. V ousden, W. D., Farr, W. M., and Mandel, I. Dynamic temperature selection for parallel tempering in Markov chain Monte Carlo simulations.Monthly Notices of the Royal Astronomical Society, 455(2):1919–1937, January

    work page doi:10.21105/joss.02505 1919

  12. [12]

    Wang, Q., Kulkarni, S

    doi: 10.1093/mnras/stv2422. Wang, Q., Kulkarni, S. R., and Verd ´u, S. Divergence es- timation for multidimensional densities via k-nearest- neighbor distances.IEEE Transactions on Information Theory, 55(5):2392–2405, 2009. doi: 10.1109/TIT.2009 .2016060. Wildberger, J., Dax, M., Buchholz, S., Green, S. R., Macke, J. H., and Sch ¨olkopf, B. Flow matching ...

    work page doi:10.1093/mnras/stv2422 2009

  13. [13]

    Naive Best-of-N

    To ensure numerical stability across different tasks, we normalize C by the maximum squared distance found within the reference samples. The distance is then obtained by solving an entropy-regularized optimal transport problem: Sε = X i,j PijCij,(33) where P is the optimal coupling matrix and ε is the regularization parameter. In our implementation, we so...

    work page arXiv 2021