Newest papers

CheckDistill: Adaptive Checklist Distillation for Efficient Distributional Reward Modeling in Image Editing

Yaqi Cai, Zuan Gao, Mengting Chen, Wei Li, Bo Rui Zheng · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26740 · cited by 0 works (OpenAlex) · open-access PDF

French

Reinforcement de l'apprentissage de MMBen pour l'édition d'images guidée par l'instruction nécessite des modèles de récompense qui fournissent des commentaires fiables, exacts et efficaces. La prédiction à échelle directe comprime les préférences d'édition multidimensionnelle, tandis que les évaluateurs génératifs peuvent exposer leur jugement à travers un raisonnement explicite. Cependant, les critères de notation larges ne spécifient pas tous les modèles d'édition dépendant de l'instance sur les contraintes d'édition et de conservation, et la génération de trajectoires de raisonnement longues est coûteuse pour l'optimisation en ligne. Nous introduisons CheckDistill, un cadre de modélisation de récompense

Machine translation (Meta M2M100); the English original follows.

Abstract

Reinforcement learning for instruction-guided image editing requires reward models that provide reliable, accurate, and efficient feedback. Direct scalar prediction compresses multidimensional editing preferences, whereas generative evaluators can expose their judgment through explicit reasoning. However, broad scoring criteria do not specify all instance-dependent editing and preservation constraints, and generating long reasoning trajectories is costly for online optimization. We introduce CheckDistill, a pointwise reward modeling framework that internalizes instruction-adaptive checklists. We decompose each editing condition into atomic, independently verifiable requirements and construct four-dimensional scoring labels together with structured verification evidence. After supervised score initialization, a privileged teacher conditions on checklists, item-level judgments, and dimension-level rationales to supervise the student's own scoring trajectories. A reliability-weighted on-policy self-distillation objective adjusts supervision at score tokens according to agreement between the teacher's expected scores and reference labels, while retaining candidate probability weights in a truncated token support. The deployed student directly predicts four score distributions without checklist inputs or explicit reasoning and aggregates their expectations into a scalar reward. CheckDistill achieves aggregate accuracies of 72.30%, 56.67% and 82.90% on MMRB2, MER-Bench and EditReward-Bench, respectively, exceeding the compared open-source baselines on all three metrics.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Verifiable secret sharing, Machine learning, Artificial intelligence, Image editing, Security token

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange