analiza aiqbench
This is an arXiv preprint on inference-time sampling for MeanFlow, a few-step generative model, evaluated on CIFAR-10, ImageNet-256, and AFHQ benchmarks. Its core contribution is decoupling the prediction horizon the model is queried with from the step size of the state update, and showing that the effect depends on both query position and the incoming state distribution. The authors are careful to say the gains are not universal, since matched-budget comparisons favor a simpler query-start shift in some tests and the proposed allocation adds nothing on two ImageNet models. For practitioners working on fast diffusion-style generation, the takeaway is a tunable, training-free inference knob, but its benefits should be verified per checkpoint rather than assumed.
Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlow
arXiv:2610.20071v1 Announce Type: new Abstract: MeanFlow predicts interval-average velocities, but conventional few-step sampling ties the interval queried from the model to the duration of the state update. We separate these two quantities at inference, fixing the model, update grid, and number of queries while varying the prediction horizon. Our central finding is that the horizon response depends on both query position and the states entering that query. Across seven independently trained CIFAR-10 checkpoints at classifier-free guidance (CFG) scale 2, a longer horizon worsens FID at early steps but improves it at the midpoint. To probe whether query time alone determines this response, we fix the midpoint query and current/subsequent guidance while changing only earlier CFG. In held-out factorial outcomes on five previously studied checkpoints, the mean midpoint horizon benefit increases from 0.109 to 1.308 FID, indicating sensitivity to the incoming-state distribution. A frozen four-query rule reduces mean CIFAR-10 FID by 2.023 without retraining or extra queries, including gains on five checkpoints trained after rule selection. A milder rule also improves mean FID by 1.041 on a stronger frozen grid, with fixed-rule FID gains on the evaluated public ImageNet-256 and AFHQ checkpoints. However, matched-budget calibration favors query-start shifting in historical CIFAR tests, and the tested Horizon allocation adds no FID benefit beyond Grid+Start on two public ImageNet models. Thus, prediction horizon is a distinct, position- and state-sensitive inference control, but not a universally preferred one.
Abstract:MeanFlow predicts interval-average velocities, but conventional few-step sampling ties the interval queried from the model to the duration of the state update. We separate these two quantities at inference, fixing the model, update grid, and number of queries while varying the prediction horizon. Our central finding is that the horizon response depends on both query position and the states entering that query. Across seven independently trained CIFAR-10 checkpoints at classifier-free guidance (CFG) scale 2, a longer horizon worsens FID at early steps but improves it at the midpoint. To probe whether query time alone determines this response, we fix the midpoint query and current/subsequent guidance while changing only earlier CFG. In held-out factorial outcomes on five previously studied checkpoints, the mean midpoint horizon benefit increases from 0.109 to 1.308 FID, indicating sensitivity to the incoming-state distribution. A frozen four-query rule reduces mean CIFAR-10 FID by 2.023 without retraining or extra queries, including gains on five checkpoints trained after rule selection. A milder rule also improves mean FID by 1.041 on a stronger frozen grid, with fixed-rule FID gains on the evaluated public ImageNet-256 and AFHQ checkpoints. However, matched-budget calibration favors query-start shifting in historical CIFAR tests, and the tested Horizon allocation adds no FID benefit beyond Grid+Start on two public ImageNet models. Thus, prediction horizon is a distinct, position- and state-sensitive inference control, but not a universally preferred one.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2610.20071 [cs.LG]
(or arXiv:2610.20071v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.20071 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 326 cuvinte
Subiecte conexe
Why do models task game?
AI Alignment Forum · 6 aug. 2026
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
arXiv · 25 sept. 2026
HIL-UMI: Bringing Human-in-the-Loop Post-Training of Vision-Language-Action Models to Universal Manipulation Interface
arXiv · 18 sept. 2026
TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
arXiv · 25 sept. 2026