AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Strengthening Multimodal Large Language Model with Bootstrapped Preference Optimization

Renjie Pi, Tianyang Han, Wei Xiong, Jipeng ZHANG, Runtao Liu, Rui Pan, Tong Zhang

ECCV 2024 · Vision-language and multimodal › Multimodal LLMs · approach: Bias-aware multi-image robustness · 44 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply BPO preference optimization to LLaVA1.5-7BLLaVA1.5-7B · MM-VetTotal score: 31.7 -> 36.8Improved
apply BPO to LLaVA1.5-13BLLaVA1.5-13B · MM-VetTotal score: 36.8 -> 41.4Improved
apply BPO to reduce hallucinationLLaVA1.5-7B · Object-HalBenchResponse-level hallucination rate (%): 45.9 -> 31.9Improved
apply BPO to reduce object-level hallucinationLLaVA1.5-13B · Object-HalBenchObject-level hallucination rate (%): 21.8 -> 12.9Improved
use preference learning with negatives instead of SFT on positive responses only (ablation)LLaVA1.5-7B · MM-VetTotal score: 33.3 -> 36.8Improved
use image-weakened negatives instead of self-generated negatives without weakening (ablation)LLaVA1.5-7B BPO · MM-VetTotal score: 32.4 -> 34.3Improved
increase preference dataset size (ablation)BPO · MM-Vet (LLaVA-7B, LLaVA-13B)Total scoreImproved
BPO versus SFT at all dataset sizes (ablation)LLaVA-7B and 13B · MM-VetTotal scoreImproved
self-generated negatives (no weakening) on math capability (ablation)LLaVA1.5-7B · MM-Vet MathMath score: 7.70 -> 1.90Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy