Starkly Speaking: Calibrating Generative Models to Distributional Constraints
2 views
Skip to first unread message
Hannes Stärk
11:58 AM (9 hours ago) 11:58 AM
to stark...@googlegroups.com
Hi together!
Join us now on zoom for this session :)
Speaker:
Henry Smith and Nathaniel Diamant who are PhD students at Stanford working with Brian Trippe
Paper:
Calibrating Generative Models to Distributional Constraints https://arxiv.org/abs/2510.10020 (Henry D. Smith, Nathaniel L. Diamant, Brian L. Trippe)
Generative models frequently suffer miscalibration, wherein statistics of the samples such as class probabilities deviate from desired values. In this work, we frame calibration as a constrained optimization problem and seek the closest model in Kullback-Leibler divergence satisfying calibration constraints. To address the intractability of imposing these constraints exactly, we introduce two surrogate objectives for fine-tuning: (1) the relax loss, which replaces the constraint with a miscalibration penalty, and (2) the reward loss, which converts calibration into a reward fine-tuning problem. We demonstrate that these approaches substantially reduce calibration error across hundreds of simultaneous constraints and models with up to nine billion parameters, spanning applications in protein design, image generation, and language modeling.
The work was completed under the supervision of Brian Trippe (Stanford Statistics).
Meeting Details:
Every Monday at 9:00 PT / 12:00 ET / 18:00 CE(S)T
https://zoom.us/j/5775722530?pwd=ZzlGTXlDNThhUDZOdU4vN2JRMm5pQT09
Henry Smith and Nathaniel Diamant who are PhD students at Stanford working with Brian Trippe
Paper:
Calibrating Generative Models to Distributional Constraints https://arxiv.org/abs/2510.10020 (Henry D. Smith, Nathaniel L. Diamant, Brian L. Trippe)
Generative models frequently suffer miscalibration, wherein statistics of the samples such as class probabilities deviate from desired values. In this work, we frame calibration as a constrained optimization problem and seek the closest model in Kullback-Leibler divergence satisfying calibration constraints. To address the intractability of imposing these constraints exactly, we introduce two surrogate objectives for fine-tuning: (1) the relax loss, which replaces the constraint with a miscalibration penalty, and (2) the reward loss, which converts calibration into a reward fine-tuning problem. We demonstrate that these approaches substantially reduce calibration error across hundreds of simultaneous constraints and models with up to nine billion parameters, spanning applications in protein design, image generation, and language modeling.
The work was completed under the supervision of Brian Trippe (Stanford Statistics).
Meeting Details:
Every Monday at 9:00 PT / 12:00 ET / 18:00 CE(S)T
https://zoom.us/j/5775722530?pwd=ZzlGTXlDNThhUDZOdU4vN2JRMm5pQT09
Meeting Details:
Every Monday at 9:00 PT / 12:00 ET / 18:00 CE(S)T
https://zoom.us/j/5775722530?pwd=ZzlGTXlDNThhUDZOdU4vN2JRMm5pQT09
Slack Workspace for discussion and paper voting:
https://join.slack.com/t/logag/shared_invite/zt-2zuxi7gd1-rLUgxg6gnCkhO7WlRsyElg
All information: Schedule of upcoming papers, recordings, mailing list:
https://portal.valencelabs.com/starklyspeaking
Hannes Stärk
Website: https://hannes-stark.com
PhD student at MIT
Reply all
Reply to author
Forward
0 new messages