Calibrated Test-Time Guidance for Bayesian Inference
#Test-time guidance #Bayesian posterior #Diffusion models #Miscalibrated inference #Black hole image reconstruction #Reward maximization #Probability sampling
📌 Key Takeaways
- Researchers identified flaws in existing test-time guidance methods for diffusion models
- Current approaches fail to recover proper Bayesian posterior distributions
- New alternative estimators enable calibrated sampling from Bayesian posterior
- The approach outperformed previous methods on Bayesian inference tasks and matched state-of-the-art in black hole image reconstruction
📖 Full Retelling
A team of researchers led by Daniel Geyfman and Felix Draxler published a groundbreaking paper on February 25, 2026, addressing critical flaws in test-time guidance mechanisms for diffusion models, revealing that existing approaches fail to recover proper Bayesian posterior distributions due to their focus on reward maximization rather than accurate probability sampling. The paper, 'Calibrated Test-Time Guidance for Bayesian Inference,' examines how current methods for steering pretrained diffusion models toward desired outcomes specified by reward functions are fundamentally flawed, particularly when accurate probabilistic representations are required. The researchers identified that these common approaches, while effective in achieving high-reward outputs, do not properly sample from the true Bayesian posterior distribution, leading to what they term 'miscalibrated inference' - a significant limitation in scientific and medical applications where understanding uncertainty is crucial. To address this issue, the team developed alternative estimators that enable calibrated sampling from the Bayesian posterior, which significantly outperformed previous methods on a range of Bayesian inference tasks and demonstrated state-of-the-art performance in black hole image reconstruction, a domain where accurate probabilistic modeling is essential.
🏷️ Themes
Machine Learning, Bayesian Inference, Diffusion Models
📚 Related People & Topics
Diffusion model
Technique for the generative modeling of a continuous probability distribution
In machine learning, diffusion models, also known as diffusion-based generative models or score-based generative models, are a class of latent variable generative models. A diffusion model consists of two major components: the forward diffusion process, and the reverse sampling process. The goal of ...
Entity Intersection Graph
Connections for Diffusion model:
🌐
Computer vision
3 shared
🌐
Vehicular automation
1 shared
🌐
Semantic change
1 shared
🌐
Recommender system
1 shared
🌐
Information retrieval
1 shared
Original Source
--> Computer Science > Machine Learning arXiv:2602.22428 [Submitted on 25 Feb 2026] Title: Calibrated Test-Time Guidance for Bayesian Inference Authors: Daniel Geyfman , Felix Draxler , Jan Groeneveld , Hyunsoo Lee , Theofanis Karaletsos , Stephan Mandt View a PDF of the paper titled Calibrated Test-Time Guidance for Bayesian Inference, by Daniel Geyfman and Felix Draxler and Jan Groeneveld and Hyunsoo Lee and Theofanis Karaletsos and Stephan Mandt View PDF HTML Abstract: Test-time guidance is a widely used mechanism for steering pretrained diffusion models toward outcomes specified by a reward function. Existing approaches, however, focus on maximizing reward rather than sampling from the true Bayesian posterior, leading to miscalibrated inference. In this work, we show that common test-time guidance methods do not recover the correct posterior distribution and identify the structural approximations responsible for this failure. We then propose consistent alternative estimators that enable calibrated sampling from the Bayesian posterior. We significantly outperform previous methods on a set of Bayesian inference tasks, and match state-of-the-art in black hole image reconstruction. Comments: Preprint. Under review Subjects: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI) Cite as: arXiv:2602.22428 [cs.LG] (or arXiv:2602.22428v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2602.22428 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Felix Draxler [ view email ] [v1] Wed, 25 Feb 2026 21:38:47 UTC (823 KB) Full-text links: Access Paper: View a PDF of the paper titled Calibrated Test-Time Guidance for Bayesian Inference, by Daniel Geyfman and Felix Draxler and Jan Groeneveld and Hyunsoo Lee and Theofanis Karaletsos and Stephan Mandt View PDF HTML TeX Source view license Current browse context: cs.LG < prev | next > new | recent | 2026-02 Change to browse by: cs cs.AI References & Citations NASA ADS Go...
Read full article at source