REVERSE-Qwen2.5-VL-3B is a novel open-source vision-language model (VLM) that performs both next-token predictioin and self-verification / self-correction during the generation process. Built on top of Qwen2.5-VL-3B-Instruct, it is fine-tuned using the 100k-subset of REVERSE Visual Instruct 1.3M dataset and equipped with a retrospective resampling mechanism that allows it to detect and correct hallucinations during generation. The model is trained in early May, 2025.
Performance
REVERSE achieves state-of-the-art hallucination reduction across diverse captioning and open-ended visual question answering benchmarks. To ensure the apple-to-apple comparison, we fine-tune the released Qwen2.5-VL-3B model using both the LLaVA-FT setup and our REVERSE recipe, applying both on the same 100k subset. This allows us to directly compare the impact of our method against the LLaVA-FT baseline under consistent conditions as the Qwen2.5-VL's instruction tuning data is not publicly available.
Benchmark
Metric
Qwen2.5-VL-FT
REVERSE (τ=0.01)
CHAIR-MSCOCO
CHAIRi (↓)
12.2
10.5
CHAIRs (↓)
45.8
39.4
AMBER-G
CHAIR (↓)
7.7
7.5
Coverage (↑)
51.7
51.5
MMHal-Bench
Score (↑)
2.89
3.15
Hallucination Rate (↓)
0.43
0.29
HaloQuest
Avg. Accuracy (↑)
33.5
45.1
False Premise Acc. (↑)
25.4
42.9
Visual Challenging Acc. (↑)
51.6
41.8
Insufficient Context Acc. (↑)
26.4
55.5
It also performs competitively on discriminative tasks compared with the base VLM.
Benchmark
Metric
Qwen2.5-VL-FT
REVERSE (τ=0.5)
AMBER-D
F1 Score (↑)
85.0
85.7
POPE
F1 Score (↑)
87.1
86.5
MME-Hall
Score (↑)
550.4
589.5
Usage
Please refer to the installation guide on GitHub to get started:
👉 Installation Guide
