← Back to the feed
AI & softwareResearch findings

Diagnosing Object Coverage Loss in Preference-Optimized Models and Restoring It via Base-Model Guidance

Original published 2026.11.18 03:24 KSTAdded to Blackdeck 2026.11.18 22:44 KST

What has changed?

By early stopping and appending high-confidence base-only categories, object coverage loss in preference-optimized models is restored without training

Key points

  1. Preference optimization reduces object coverage in vision-language models
  2. Early stopping and base-only category appending restore missing object categories without training
  3. Base-model-guided restoration achieves up to 86% of the coverage gain from optimized checkpoints

Connect it to your work

Preference optimization reduces object coverage; base-model-guided restoration recovers missing categories effectively

What remains unproven

This research note is based on an abstract; check the original for test conditions and application scope.

Originals used for this note

Links open the original publisher in a new tab.

  1. arxivSay Less, Hallucinate Less? Diagnosing and Restoring Object Coverage in Preference-Optimized Vision-Language Models