Through the PRISM: Preference Representation in Intermediate States of Video Diffusion Models

A team of researchers has introduced PRISM, a video evaluation model that discriminates preferences directly from the noisy intermediate states of a diffusion process, bypassing the need for clean, pixel-based reward models. [1] The model, formally named Preference Representation in Intermediate States of Diffusion Models, was detailed in a paper submitted to arXiv on June 18, 2026. [1] The work challenges the conventional paradigm of evaluating generated video with clean, pixel-based reward models, which the authors argue disconnects evaluation from the noisy diffusion process and incurs massive VAE decoding costs. [2] PRISM addresses this by employing a lightweight Query-based Aggregation head attached to a frozen video diffusion backbone to decode preference signals directly from noisy latents. [2] The approach achieves state-of-the-art preference accuracy and demonstrates strong noise-robustness, a property that enables early-stage Best-of-N sampling. [2] This capability allows the system to filter suboptimal candidates at the very beginning of the denoising process, drastically reducing computation while boosting final video quality. [2] The researchers also revealed a strong positive correlation between a backbone's generative performance and its inherent evaluative power, a finding they suggest could enable self-improving video backbones. [2] The paper was posted on arXiv, an open-access repository for electronic preprints in fields such as computer science, physics, and mathematics that has hosted over two million articles since its founding in 1991. [6] The abstract page for the PRISM paper features several community-developed tools under the arXivLabs framework, a program launched in 2020 to allow collaborators to build experimental features on the platform. [4] These tools, which appear as tabs at the bottom of the abstract page, include the Bibliographic Explorer for navigating citation trees and the CORE Recommender for discovering related open-access papers. [5] arXivLabs is currently pausing new proposals while the development team focuses on modernizing the repository's infrastructure and moving its systems to the cloud. [3]